Prevalence, causes and impact of TP53-loss phenocopying events in human tumors

Fito-Lopez, Bruno; Salvadores, Marina; Alvarez, Miguel-Martin; Supek, Fran

doi:10.1186/s12915-023-01595-1

Research article
Open access
Published: 24 April 2023

Prevalence, causes and impact of TP53-loss phenocopying events in human tumors

Bruno Fito-Lopez¹,
Marina Salvadores¹,
Miguel-Martin Alvarez¹ &
…
Fran Supek ORCID: orcid.org/0000-0002-7811-6711^1,2

BMC Biology volume 21, Article number: 92 (2023) Cite this article

2253 Accesses
3 Citations
6 Altmetric
Metrics details

Abstract

Background

TP53 is a master tumor suppressor gene, mutated in approximately half of all human cancers. Given the many regulatory roles of the corresponding p53 protein, it is possible to infer loss of p53 activity – which may occur due to alterations in trans – from gene expression patterns. Several such alterations that phenocopy p53 loss are known, however additional ones may exist, but their identity and prevalence among human tumors are not well characterized.

Results

We perform a large-scale statistical analysis on transcriptomes of ~ 7,000 tumors and ~ 1,000 cell lines, estimating that 12% and 8% of tumors and cancer cell lines, respectively, phenocopy TP53 loss: they are likely deficient in the activity of the p53 pathway, while not bearing obvious TP53 inactivating mutations. While some of these cases are explained by amplifications in the known phenocopying genes MDM2, MDM4 and PPM1D, many are not. An association analysis of cancer genomic scores jointly with CRISPR/RNAi genetic screening data identified an additional common TP53-loss phenocopying gene, USP28. Deletions in USP28 are associated with a TP53 functional impairment in 2.9–7.6% of breast, bladder, lung, liver and stomach tumors, and have comparable effect size to MDM4 amplifications. Additionally, in the known copy number alteration (CNA) segment harboring MDM2, we identify an additional co-amplified gene (CNOT2) that may cooperatively boost the TP53 functional inactivation effect of MDM2. An analysis of cancer cell line drug screens using phenocopy scores suggests that TP53 (in)activity commonly modulates associations between anticancer drug effects and various genetic markers, such as PIK3CA and PTEN mutations, and should thus be considered as a drug activity modifying factor in precision medicine. As a resource, we provide the drug-genetic marker associations that differ depending on TP53 functional status.

Conclusions

Human tumors that do not bear obvious TP53 genetic alterations but that phenocopy p53 activity loss are common, and the USP28 gene deletions are one likely cause.

Background

Mutations in the TP53 tumor suppressor gene are a very common feature across almost all types of human cancer. These mutations abrogate or reduce TP53 activity via various mechanisms: dominant-negative acting missense mutations, loss-of-function missense, nonsense, frameshift indel, splice site, or synonymous mutations, or copy number losses that frequently delete one TP53 allele while the other allele is inactivated by a mutation. That such TP53 genetic alterations occur at high frequency in many cancer types implies that they have very strong selective advantages for the expanding cancer cell clones [1, 2]; indeed this is borne out in experimental data on cell lines and animal models of cancer [3, 4].

The large selective advantage of TP53 losses are consistent with its roles in arresting the cell cycle or triggering apoptosis upon threats to genome integrity. TP53-null cells better tolerate genomic instability, which can result from endogenous causes, most prominently oncogene-overexpressing and thus replication-stressed cancerous genetic backgrounds. Consistently, TP53-mutant tumors have higher frequencies of segmental copy number alterations (CNA), whole-genome duplication, and overall mutation rates [5, 6]. Moreover, TP53-null cells better tolerate DNA damaging conditions that would normally trigger cell cycle checkpoints, such as those resulting from DNA-acting drugs or radiation [7, 8]. Consistently, TP53 mutation-bearing tumors tend to be more resistant to various cancer chemotherapies [4, 9,10,11] and radiotherapy [10,11,12], and more aggressive TP53 R273 and R248 mutants are associated with accelerated cancer progression in colorectal tumors [13].

The frequency of TP53 mutations – highest of all cancer genes, standing at 37% in The Cancer Genome Atlas (TCGA) cohort – indicates that most cancers benefit from the loss of TP53. However, there are nonetheless many tumors which do not bear a mutation in TP53. A part of those is explained by genetic events that phenocopy TP53 loss i.e. that have similar downstream phenotypic consequences as TP53 gene loss itself. There are three established examples of TP53 loss phenocopying events occurring in tumors. Most prominently, this is the amplification of the MDM2 and MDM4 oncogenes and overexpression of the corresponding proteins. These negatively regulate TP53 protein levels by promoting its proteasomal degradation, and otherwise inhibit TP53 activity by binding to its transactivation domain [14,15,16]. The third implicated gene is PPM1D, whose amplification overexpresses a serine/threonine phosphatase acting upon various targets including TP53, reducing its activity. (We note that PPM1D can also be affected by point mutations that result in gain-of-function [17,18,19]).

Given the strong selective advantages of the TP53 activity loss in cancer evolution, we hypothesized that TP53 loss phenocopying in human cancers extends beyond these known examples of MDM2, MDM4 and PPM1D alterations. If indeed other common mechanisms of TP53 phenocopying exist, this would be relevant to predicting tumor cell response to various drugs, and to predicting tumor aggressiveness, thus having implications to personalized medicine. Because TP53 loss has clear consequences on the mRNA expression levels of various downstream targets [4, 20], the TP53-null-like phenotype can be inferred from large scale transcriptomic data [20,21,22,23]. Here, we apply a statistical framework to jointly analyse 966 cancer cell line and ~7000 tumor genomes and transcriptomes, aiming to identify additional TP53 phenocopying genetic events and impact on drug sensitivity. We find that TP53 loss phenocopies are remarkably common across tumors and cancer cell lines, and we identify USP28 gene deletions as one cause of TP53 loss phenocopying, and reveal many links between drugs and their targets that are modulated by TP53 activity.

Results

Inferring the functional TP53 status of tumors from transcriptomes

We developed a machine learning method to detect TP53 loss phenocopies in tumors and cell lines, integrating RNA-seq data with TP53 mutation data in a logistic regression, regularized with an Elastic Net penalty (very similar cross-validation accuracy was obtained with Ridge or Lasso penalties; see Methods). Regression models were trained using cross-validation on mRNA levels of 7131 tumor samples from the TCGA project, across 20 different cancer types, controlling for cancer type. In addition to using this global analysis of mRNA expression levels to infer the functional TP53 status state of each tumor, we also identified the genes whose expression patterns are associated with TP53 status. Tumors with TP53 putatively causal mutations were included as positive examples (TP53 status was categorized according to Genomics of Drug Sensitivity in Cancer Project (GDSC) methodology; see Methods). Previously known phenocopying events (MDM2, MDM4 and PPM1D amplifications), as well as samples with TP53 deletions were excluded from the training set (these known phenocopying events will be used to calibrate decision thresholds; see below). Our classifier learned a combination of relevant gene weights that differentiate samples with an aberrant TP53 activity. Tumor samples that are not TP53 mutated (by GDSC criteria), but are classified as mutated by the machine learning model are considered to be TP53 loss phenocopies.

Our classifier showed a high performance with an area under the receiver operating characteristic (AUROC) curve of 96% in cross-validation on TCGA tumors (out-of-sample accuracy), and 95% on the testing set (consisting of 10% of the samples held out from training set, Fig. 1a). Thus, we were able to often correctly detect TP53 status in tumor samples the classifier was not exposed to, with an area under precision-recall curve = 0.9654. The TP53 loss phenocopy scores for each TCGA tumor sample are provided in Additional file 3: Data S1.

To account for cancer type-specific effects, our classifier included the cancer type as a variable. Leave-one-cancer-type-out (LOCTO) cross-validation showed that classifier performance was robust to removal of individual cancer types, with LOCTO AUCs ranging from 0.951 (LGG removal) to 0.964 (SKCM removal) compared to an are under the curve (AUC) of 0.961 using the full dataset (Additional file 1: Fig. S1a). TP53 phenocopy scores from the LOCTO models were highly correlated with those from the full model (e.g. R = 0.89, p = 4e-163, for BRCA removal, Additional file 1: Fig. S1b), indicating that the classifier would generalize well to unseen cancer types. Additionally, we tested for correlations between tumor purity estimates and TP53 phenocopy scores across all cancer types and found no significant relationship overall (pan-cancer Pearson R = -0.03, p = 0.11, 1/10 cancer types was found correlated, Additional file 1: Fig. S1c).

Out of the ~ 12,000 genes available to the classifier, 217 genes were deemed relevant for TP53 status classification (non-zero coefficients; gene score provided in Additional file 3: Data S2). These represent a sparse (but not necessarily exhaustive) set of genes that are, considered together, highly informative for predicting TP53 functional status. Expectedly, many of the classifier's most relevant genes are known to be related to TP53 functionality. For instance, apoptosis-enhancing nuclease (AEN) was the gene with the highest absolute importance score. This exonuclease is a direct TP53 target whose expression is regulated by the phosphorylation of TP53 and its tumor suppressor role has been reported [24]. Tumors with a high expression of AEN are expected to be p53 functional, and indeed highly expressed AEN was associated with TP53 WT status in our classifier’s coefficients. On the other extreme, COP1, a ubiquitin ligase that acts as an important p53 negative regulator, was the highest negative importance score with TP53 mutated status in the classifier [25]. We further performed a GO enrichment analysis, revealing that top functional enriched sets were related to apoptotic signals, supporting the biological rationale underlying this set (Additional file 1: Fig. S1d). Most enriched pathways were: intrinsic apoptotic signalling pathway in response to DNA damage by p53 class mediator (8.1-fold enrichment, false discovery rate (FDR) = 4.2%), pyrimidine deoxyribonucleoside monophosphate biosynthetic process (47.4 fold enrichment, FDR = 1.9%) and response to UV-B (17.2 fold enrichment, FDR = 3.7%) (ShinyGO, see Methods).

Validation and calibration of the TP53 phenocopy score model

Our classifier extends recent gene expression-based models for TP53 functionality [20,21,22,23] by being able to generalize across both tumor and cancer cell lines (important for identifying drug sensitivity associations, see below), and moreover it can provide calibrated FDR estimates for TP53 functional status of each tumor or cell line. In particular, to assess the reliability of the individual predictions from the model, FDR for each tumor was computed by considering the cross-validation precision-recall curves (Fig. 1b). The previously known phenocopying events (MDM2, MDM4 and PPM1D amplifications) and TP53 deep deletions, which were held out from the training set, were largely scored as TP53 mutated. Tumors harbouring a known phenocopying amplification were assigned higher scores than the rest of TP53 wild-type tumors (mean score = 0.56 and 0.27 respectively, p = 1e-65 by t-test). Cells harbouring a TP53 deep deletion also had higher scores (mean TP53 deleted = 0.47, mean TP53 not deleted = 0.27, p = 1e-08). We tested multiple threshold values for defining TP53 loss phenocopies, which expectedly yielded varying percentages of phenocopies across cancer types (Additional file 1: Fig. S1e). Our choice of threshold to detect TP53 phenocopied tumors was set based on these known phenocopying events, conservatively, corresponding to score > 0.6, Methods; Fig. 1b).

This resulted in an empirical FDR estimated at 15% (i.e. precision of 85%), based on the known TP53 mutations. Importantly this 15% is a conservative upper-bound estimate of FDR, since it is based on the assumption that there do not exist any unknown TP53 phenocopying events: it classifies all high-scoring TP53 wild-type tumors as false positives. Conversely, using the known phenocopying events we estimate a lower-bound recall (sensitivity) of this classifier at 63% (Fig. 1b). Again, this estimate is conservatively biased, since it is not a priori known whether every copy number gain in MDM2/MDM4/PPM1D causes a phenocopy; some low-level gains may not have effects and thus would appear as false-negatives.

To additionally validate the classifier, we inspected the relationship between known phenocopy genes’ allele copy-number (see Methods), and the TP53 phenocopy score. There were significant positive correlations between three known phenocopying genes copy-number, and the TP53 phenocopy score in TP53 wild-type tumors (Fig. 1c).

To independently validate our TP53 phenocopy classifier, we applied it to RNA-seq data from 555 tumors across 7 cancer types from the Pan-Cancer Analysis of Whole Genomes (PCAWG) study. Using the same procedure as with the TCGA dataset (FDR = 15%, recall = 80%), the FDR on this validation cohort was 8.8% (recall = 77%, Additional file 1: Fig. S1f). TP53 phenocopy scores were significantly higher for TP53-altered samples compared to TP53 wild-type samples, both in a pan-cancer analysis (0.87 vs. 0.48, p = 2e-60, t-test) and in individual cancer types (Additional file 1: Fig. S2a,b). Similarly, known phenocopying genes scored a high phenocopy score (Additional file 1: Fig. S2c). These results were consistent with those from the TCGA dataset (0.86 versus 0.32 for TP53 altered versus wild-type, p≈0, t-test). Together, these data independently validate our TP53 phenocopy classifier and findings. The TP53 function loss phenocopy scores for each PCAWG tumor sample are provided in Additional file 3: Data S3.

The prevalence of phenocopying events was substantial: overall 12% of all TCGA tumor samples were redefined into a TP53 mutated-like category (Fig. 1d) by our criteria. Different cancer types display different phenocopy frequencies (Fig. 1d), overall frequency ranging from 19% for breast cancer (BRCA cancer type) to 3% for B-cell lymphoma (DLBC cancer type); phenocopy frequencies are shown in Additional file 1: Fig. S1d. For instance, most breast cancer TP53-phenocopied tumors derive from previously known events i.e. the MDM4/MDM2/PPM1D amplifications are the most common event, while the remaining 27% of the TP53 loss phenocopies (5% of all breast cancer samples) are not associated with a known phenocopying event (proportion shown for every cancer type Fig. 1d). We note that it is still possible that individual examples of tumors may be erroneously classified as TP53-deficient at the applied score threshold. Overall, 51% of TP53-loss phenocopied tumor samples across all cancer types were not linked with one of the three known causal genes nor a CNA deletion in TP53 itself, suggesting that additional TP53 phenocopying mechanisms are commonly occurring in human tumors.

TP53 loss-of-function (LoF) mutations can result in either complete or partial LoF. Our TP53 phenocopy score distinguished between TP53 alterations that likely disrupt p53 protein activity to varying extents (Additional file 1: Fig. S2e). The VARITY and EVE variant pathogenicity scores (see Methods) predict the likelihood that a given missense variant is disease-causing. TP53 phenocopy scores were modestly but significantly correlated with these variant scores (VARITY: R = 0.16, p = 7e-11; EVE: R = 0.09, p = 1e-4), indicating that our classifier also can capture differential functional impact of TP53 missense variants. Nonsense mutations and frameshifting indels often have higher TP53 phenocopy scores than missense variants (major cancer types shown in Additional file 1: Fig. S2e).

USP28 deletion phenocopies a TP53 mutated state in tumors

Prompted by the abundance of tumor samples that are functionally TP53 null but lacking an obvious TP53 loss or a known phenocopying event, we sought to identify other phenocopying genes across all cancer types. To this end, we designed a custom association-testing methodology that combines six different statistical tests across four different genomic data types (see Methods, Additional file 1: Fig. S3).

In brief, our methodology is based on the rationale that genes that cause a TP53 loss phenocopy via altered dosage at DNA and mRNA levels should exhibit a distinct copy number variant (“CNV” tests) and also gene expression (“GE” tests) pattern. Each of these two genomic data types is considered in two tests, one comparing TP53 phenocopying against TP53 wild-type tumors, and another comparing TP53 phenocopying against TP53-mutant tumors, for a total of four tests. As two additional tests, we considered external data from genetic screens across large panels of cancer cell lines [26, 27]. In particular, we tested for significant codependency scores, explaining how a knockout (“CRISPR”) or knock-down (“RNAi”) of a candidate phenocopying gene affects fitness across a panel of cell lines, when compared with the fitness profile of a TP53 knockout/knock-down across the same panel [28, 29]. As an example supporting the use of our methodology that combines cancer genomic analysis and genetic screening data analysis, the CRISPR knockout of the known TP53 negative regulator MDM2 decreases cell line fitness, in a manner anticorrelated to a TP53 knockout across cell lines. (Additional file 1: Fig. S4a).

In summary, we tested differences of tumor genomics CNV and GE patterns (two tests each as above), additionally considering “CRISPR” and “RNAi” test scores from genetic screens, for each gene, performing tests stratified by cancer type. Our final score combines the 6 tests together, providing a ranking of potential TP53 phenocopying genes.

As anticipated, top 3 genes by the overall prioritization score are MDM2, MDM4 and PPM1D genes (Fig. 2a). A breakdown of our scores by different cancer types is provided in Additional file 1: Fig. S3. Following those known TP53 phenocopying genes, the gene USP28 was the 4th ranked gene (genes neighboring MDM2/MDM4/PPM1D excluded) in terms of overall statistical significance (p = 5.9e-07, combined across all six tests), and in particular scored highly on CRISPR codependency (pan-cancer score for USP28 = 0.54, compared with -0.72 for MDM2 and -0.53 for MDM4, breakdown by cancer type in Additional file 1: Fig. S4a). An overview of the top 20 genes according to our prioritization scores is provided in Additional file 1: Fig. S4c.We note that, in contrast to MDM2 and MDM4, it is the deletions not amplifications of USP28 that were associated with TP53 loss phenocopying; this is reflected in the mirrored direction of the CRISPR codependency score. USP28 encodes a deubiquitinase enzyme with substantial evidence from previous biochemistry and cell model studies that link it to p53 activity and apoptotic responses.

While it was proposed that USP28 was linked to DNA damage apoptotic responses through the ATM-CHK2-p53 pathway, recent evidence however suggests that a distinct pathway involving TP53BP1 and USP28 induces p53 and cell cycle arrest after mitotic delays [30,31,32,33,34,35]. The observation that USP28 deletions are identified as a TP53 phenocopying event in our analysis with higher priority than ATM or CHK2 alterations, suggests that loss of TP53BP1-USP28-p53 pathway drives selected events to a higher degree than loss of the ATM-CHK2-p53 pathway in many tumors. This could help tumors adapt to replication stress that can slow mitotic progression, and to tolerate supernumerary centrosomes.

Overall, diverse experimental evidence from genetic screens strongly supports our identification of USP28 deletions as p53-loss phenocopying events, and our cancer genomic and transcriptomic analysis suggests a widespread distribution of causal USP28 deletions across human tumors.

We also assessed how many previously unexplained TP53 phenocopying tumors (“Unknown cause” group in Fig. 1d) could be attributed to USP28 deletions. USP28 deletions accounted for a subset of TP53 phenocopying tumors (12% had exclusively a USP28 deletion, 24% had a USP28 deletion and a TP53 shallow deletion) by transcriptome score that did not bear TP53 mutations nor deep deletion and that were not associated with known phenocopy mechanisms of MDM2/MDM4/PPM1D amplification (Additional file 1: Fig. S4d).

Additional hits from this association study might provide promising genes for follow-up. For instance, MSI2 was the 5th most highly prioritized gene, predicted to phenocopy TP53 loss by amplification. MSI2 encodes a transcriptional regulator that has been recently identified as an oncogene in hematologic and solid cancers [36,37,38]; we note that the MSI2 locus is in the broader neighborhood of PPM1D (17q22 and 17q23.2, respectively; 3.3 Mb genomic distance). Similar results to CRISPR analyses were observed using RNAi screening codependency scores, further supporting the association of the USP28 losses with TP53 phenocopying, as well as the MSI2 gains (Additional file 1: Fig. S4b). Other apoptosis-related genes such as CHEK2 and ATM [39, 40] were also in the prioritized genes in our analysis, albeit at more modest statistical significance. Of note, the TPR gene also had a highly significant codependency score but was driven by a single cancer type (kidney) and thus with less clear relevance to diverse tumor types. Top prioritized genes and significance calls from the individual association tests are provided at Additional file 3: Data S4.

Phenocopy scores prioritize causal genes within CNA segments

Amplifications of certain chromosomal segments (or whole arms) in case of MDM2, MDM4 and PPM1D commonly underlie TP53 phenocopies. Such CNA genetic events however often affect multiple adjacent genes, where an open question in cancer genomics is which gene or genes in the affected segment are causal [41]. We hypothesized that the known TP53 phenocopying gene CNA segments might in some cases harbor more than one causal gene. Our genomic combination testing approach can prioritize genes with enriched gene expression and/or CNA in the TP53 loss phenocopying group. Considered together with the prioritization by CRISPR and RNAi codependency, this method provided a plausible ranking of possible TP53 loss phenocopying genes. Applied globally, this identified USP28 as a novel phenocopying gene (see above). To more formally study if the USP28-adjacent genes could contribute to this, we considered that the same method could be applied on a local scale: by examining profiles of CNVs along chromosomes, our genomic prioritization scores would be able to single out the causal gene(s) in the segment of recurrent CNA.

As a control for this approach, we sought to confirm previously known phenocopies. Indeed, MDM4 amplification was found to be enriched in the TP53-phenocopying group of tumor samples, but not in the rest of tumor groups – the TP53 mutant and the non-phenocopying TP53 wild-type (Fig. 2b). The local profile of this enrichment for the chromosome 1q segment 32.1 peaks at the MDM4 gene and falls off towards its flanking genes (Fig. 2b). Our CRISPR and RNAi data analysis, consistently, indicate MDM4 as the gene with the strongest effect in the region (Fig. 2b). As expected, similar CNA and CRISPR/RNAi profiles were observed at PPM1D (Additional file 1: Fig. S4e). The MDM2 CNA enrichment score segment peak was narrower, suggesting a more focal gene amplification process (Additional file 1: Fig. S4e).

Next, we examined the local shape of the USP28-adjacent CNA profiles. USP28 deletions were found to be enriched in the TP53 phenocopying group when compared to the rest of tumor groups (2.3-fold in TP53 w.t, 2.8-fold in TP53 mutant). USP28 enrichment was comparable to MDM4 region enrichments of 2.5–3.7-fold (TP53 wt., TP53 mutant) (Fig. 2b,c). TP53 phenocopying tumor samples appear to have more deletions in the USP28 region than TP53 wild-type (non-phenocopying) and TP53 mutant samples. The local profile of enrichments presents a plateau-like pattern rather than a sharp peak, and USP28 is within the top-ranked genes in the plateau, however some neighbouring genes do appear similarly so. Therefore, we further used the CRISPR and RNAi codependency scores to prioritize the causal genes in this deletion segment; these scores clearly distinguishes USP28 from its immediate neighbours (Fig. 2c), suggesting that USP28 is indeed the causal gene in the chromosomal segment.

Furthermore, this ‘local scan’ can be applied chromosome-wide, where we noted another small region on chromosome 11q.12.1-q1.13.1 modestly enriched with amplifications in TP53-phenocopying tumors (Additional file 1: Fig. S5a), thus raising our interest. However, neither the genes in this region nor other chromosome 11 regions showed a positive CRISPR codependency score of even half of the USP28 score (Fig. 2c). Of note, the USP28 codependency score exceeds, in absolute magnitude, the score of the known MDM4 phenocopy (Fig. 2b,c).

In the broader neighborhood of USP28, the gene ATM seems to also be frequently deleted in the TP53-phenocopying tumor group, meaning ATM is also a candidate for the causal gene in this deletion segment at chr11 q22.3-q23.2. However, the statistical support from genomic enrichment scores (using our methodology for meta-analysis across 6 statistical tests) for ATM were less strong than for USP28 (p = 1e-5 versus p = 6e-7, respectively). Consistently, comparing the RNAi and CRISPR TP53-codependency scores of ATM versus USP28 shows a more robust effect of the USP28 knockout (USP28 RNAi codependency score p = 4.9e-112 versus ATM p = 3e-80, in a pan-cancer analysis; Additional file 1: Fig. S5b). To further rule out that ATM has the causal role in this deleted segment, we considered the cases of tumors where ATM is disrupted by a point mutation; unlike CNA in the ATM gene, these cases are not commonly genetically linked with disruptions in USP28. The ATM mutated but USP28 wild-type tumors had considerably weaker TP53 phenocopy transcriptomic scores (median = 0.36) than the USP28 deleted but ATM non-mutated tumors (median = 0.84; p = 0.0013 by Mann–Whitney test; Fig. 2d). The cases where both USP28 and ATM were disrupted, by deletion or mutation, had very similar phenocopy scores (median = 0.73) as the USP28 deleted but ATM non-mutated cases. This analysis of ATM mutations supports that USP28 deletion, rather than ATM disruption, is the causal change in the deleted segment at chr11 q22-q23.

To validate the USP28 finding, we analyzed an independent CRISPR data set, consisting of 3 genome-wide screens performed on TP53 wild-type and TP53 -/- isogenic pairs of cell lines: one on the A549 cell line pair and two on the RPE1 cell line pairs (see details in Methods). In the TP53 wild-type background, the TP53 k.o. increases cell fitness, as expected for a high-effect tumor suppressor gene (Fig. 2e). Thus, if the USP28 loss was to phenocopy TP53 loss, the USP28 k.o. by CRISPR should also increase fitness. Indeed, compared to the 10 neighboring control genes residing within 1 Mb of USP28, the USP28 k.o. has a stronger fitness effect (beta score from MAGeCK tool, see Methods) for 10 out of 10 neighboring genes in 2 out of 3 screens, and for 8 out of 10 neighboring genes in the remaining screen (Fig. 2e). For ATM, this effect is less pronounced (Additional file 1: Fig. S5c). In 3 out of 3 cell screening experiments, USP28 fitness effect was stronger than ATM effect (1.4-fold, 2.4-fold and 2.6-fold increased beta score). To further support this finding, we asked if the fitness gain resulting from USP28 loss is because of downstream effects on TP53 activity. We thus considered the isogenic cells where TP53-/- was ablated, in which indeed the fitness gain from USP28 k.o. was attenuated or disappeared (Fig. 2e) compared to TP53 wild-type cells. In 2 out of 3 cell line screens, the fitness attenuation effect of TP53-/- background cells was stronger in USP28 than in the neighboring ATM gene, supporting the causal role of USP28 in that segment (Additional file 3: Data S5). Of note, in this analysis the effect sizes of USP28 k.o. were less than the full effect of TP53 k.o., however they were still substantial: in 2 out of 3 screens, the fitness gain of USP28 disruption was comparable in size to the fitness loss of MDM4 disruption (Fig. 2e).

Overall, these analyses highlight USP28 as the likely causal gene for TP53 loss phenocopying via deletions in the chr11 q22-q23 segment.

Cancer type spectrum of TP53 phenocopying events

As stated above, not every cancer type is affected by the same types of phenocopies. For instance, MDM2 amplification phenocopy occurs often in BRCA, CESC, BLCA, LUAD and STAD but less so in HNSC, OV, MESO and LIHC (Fig. 1d). To further elucidate the tissue-specificity of USP28 phenocopying events, we considered the prioritization scores separately for different cancer types (Additional file 1: Fig. S3). We observed that BRCA, BLCA and LUAD were the cancer types which showed the strongest signal in our prioritization score for USP28 phenocopies, with a suggestive signal in STAD.

To elucidate the cancer type spectrum of the USP28 phenocopies, we considered the USP28 amplifications as a negative control (in this gene, deletions are expected to phenocopy). In particular, we determined in which tumor types USP28 deletions had a higher TP53 phenocopy score than USP28 copy number-amplified samples. As expected, statistical significance (when comparing the TP53 phenocopy score) of USP28 copy number-neutral tumor samples versus those bearing deletions was stronger than when comparing copy number-neutral to amplifications. This supports the impact of USP28 deletions on TP53 loss phenocopy score. The strongest effect was found in BLCA, STAD, BRCA, LIHC and LUAD (Fig. 2e). In further support of this tissue spectrum, when CRISPR TP53 codependency scores were checked, highest USP28 scores were found in cancer cell lines originating from BLCA, STAD, BRCA, LIHC and LUAD (Fig. 2e). The leading codependency score was found in BLCA (effect size = 0.73, p = 2.2e-08) and BLCA also had the most significant value when comparing deletions to neutral copy number TP53 phenocopy score (p = 4.2e-06, Additional file 1: Fig. S5d). LUAD had the second most significant codependency p-value (p = 3.8e-6), and is also highly ranked in comparison of phenocopy score between deletion versus neutral USP28 CNV tumors (Additional file 1: Fig. S5d). We found a positive association between USP28 CRISPR codependency score and the effect of USP28 deletions in TP53 phenocopying score across cancer types (Additional file 1: Fig. S5d). Of note, the “oncogene-tumor suppressor” dichotomy of USP28 was reported [42], which might imply that USP28 amplification could also result in a TP53 phenocopy in certain contexts. However, our analysis did not support this in the majority of cancer types: out of 14 cancer types, only 3 of them had a stronger TP53 phenocopy score in USP28-amplified samples than in USP28-deleted samples (Fig. 2e); this was the case for none of the cancer types with common USP28 phenocopying (BLCA, STAD, LIHC, BRCA and LUAD).

Taken together, these results suggest that USP28 deletion is a novel TP53 phenocopying alteration that commonly affects major cancer types such as breast cancer (6.2% of total breast tumors, not counting known phenocopying events and TP53 deletions) and also bladder, lung, liver and stomach cancer (7.6%, 7.0%, 3.8% and 2.9% cases, respectively).

Multiple neighboring genes in CNA segment can contribute to a TP53 deficient state

Some of the top hits found in our combined testing approach were located near to known TP53 loss phenocopying genes such as MDM2. We thus hypothesized that there exist cases of ‘collaboration’ of two or more neighboring genes, affected by a single copy-number alteration event, which jointly bear upon the TP53 loss phenotype. This would represent a special case of epistasis between two genes, caused by a single alteration that spans and affects both genes. Our data suggests that the CNOT2 gene, residing in the MDM2 segment in the chromosome 12q15, is a likely example of this.

In particular, in our data, MDM2 was frequently co-amplified with CNOT2, in 72% of the cases of MDM2 amplification (Additional file 1: Fig. S6a, check by cancer type at Additional file 1: Fig. S6b). Data from CRISPR and RNAi screening experiments can help resolve associations from genomic analysis, where effects of neighboring genes are in genetic linkage (here implying being jointly affected by CNA). No other gene in that neighborhood that was amplified together with MDM2 had as high CRISPR codependency scores as CNOT2 (effect size = -0.24, p = 4.1e-14, Fig. 3a,b); next best gene in the 20 Mb neighborhood is CDK4 with effect size = -0.16, p = 3e-7. However, CDK4 is co-amplified with MDM2 in only 20% of the cases (Fig. 3a). CNOT2-only amplifications (i.e. without concurrent MDM2 CNA) do not significantly associate with TP53 phenocopy score (Pearson's TP53 phenocopy score vs CNOT2 CNV p = 0.45, effect size = -0.83, Additional file 1: Fig. S6c). More interestingly, MDM2 CNV was not found to be associated with our TP53 phenocopy score when MDM2-only amplified without CNOT2 (Pearson´s TP53 phenocopy score vs MDM2 CNV p = 0.57, effect size = 0.09, Additional file 1: Fig. S6c). On the other hand, MDM2-CNOT2 co-amplifications were significantly associated with a TP53 deficiency transcriptomic score in tumors (Pearson's correlation TP53 phenocopy score vs MDM2 CNV p = 2e-05, effect size = 0.41, Additional file 1: Fig. S6c,e).

This genomic evidence is supported by recent experimental studies, indicating a role for CNOT2 in p53-dependent apoptosis, and suggesting therapeutic potential of CNOT2 suppression (see Additional file 2: Text S1 for a summary and references). As supporting evidence, we considered fitness effects of CNOT2 k.o. by CRISPR in various subsets of cell lines. The MDM2-gain but CNOT2-neutral genetic backgrounds had more modest fitness effects of CNOT2 k.o. (median = -0.37) than the CNOT2-gain but MDM2-neutral genetic backgrounds (median = -0.62; p = 0.072 by Mann–Whitney test, Additional file 1: Fig. S6d). Consistently, the CNOT2 k.o. by CRISPR had stronger fitness effects (median = -0.55) in the TP53 wild-type backgrounds than in TP53-mutant background cell lines (median = -0.45, p = 0.0091 by Mann–Whitney test). In other words, fitness effects of CNOT2 disruption by CRISPR are conditional upon MDM2 alterations and TP53 alterations, implicating CNOT2 in a three-way genetic interaction with the two other genes.

We hypothesized that this role of CNOT2 in boosting the TP53-phenocopying effect of MDM2 amplification may be variable across tissues. Our data suggests that in some cancer types TP53 functional loss seems to rely on amplifications of both genes together, rather than solely MDM2, but not in all (Additional file 2: Text S2). This suggests a model where the MDM2-CNOT2 coamplification enhances the TP53 loss effect via a genetic interaction, and amplification of MDM2 alone but not CNOT2 alone able to generate a TP53 functional loss phenotype. Gene expression profiles match this observation from CNA: having a MDM2 and CNOT2 co-overexpression (over the 97th percentile; n = 40) implies a high mean TP53 phenocopy score (above the 84th percentile, mean phenocopy score MDM2_CNOT2 = 0.65, Fig. 3c, Additional file 1: Fig. S6f), however less so for a MDM2-only overexpression (76th percentile; mean MDM2 only score = 0.46, Fig. 3c, Additional file 1: Fig. S6f), and, expectedly, even less so for a CNOT2-only overexpression (73th percentile; mean CNOT2 only score= 0.41).

This principle might extend beyond the MDM2-CNOT2 pair. For instance, the MSI2 gene, another highly prioritized hit in our combined test (Additional file 1: Fig. S6g,h,i), resides near the known phenocopying gene PPM1D and has the potential to boost the effects of the linked amplification of the PPM1D gene to cause a TP53 deficient state. Considered jointly, these findings suggest the possibility of TP53-loss like phenotype being a result of multiple phenocopying events of neighboring genes resulting from a single segmental CNA.

Detecting TP53 loss phenocopies in cancer cell line panels

It is well known that TP53 mutations associate with overall poorer drug response in tumors [43,44,45], consistent with a lower ability of TP53 deficient cells to trigger cell cycle arrest and/or apoptosis [46,47,48,49,50]. We hypothesized that, in addition to conferring a generalized drug resistance, the TP53 function loss may also modulate the association between specific drugs and alterations in their target genes. In other words, we asked whether in TP53 wild-type cancer cells, for instance, amplification in a particular oncogene predicts sensitivity to a particular drug, while in TP53 mutant cells the same amplification does not associate with sensitivity. Cancer cell line drug screening panels [51, 52] provide a convenient system to test this hypothesis, because many drugs were tested systematically across both TP53 wild-type and mutant cells originating from multiple cancer types. Considering TP53 function loss via phenocopy score should afford additional statistical power and clarify the associations discovered; otherwise, some effectively TP53 null cells would be erroneously considered wild-type during statistical testing, making it more difficult to identify associations.

First, we aimed to generalize our tumor TP53 phenotype classifier to cancer cell lines. Because cell lines exhibit strong global (i.e. affecting many genes) shifts in gene expression patterns, compared to their tumor tissue of origin, we applied an adjustment methodology as in our recent work [53] using the COMBAT tool [54]. Upon adjusting gene expression from cell lines in the CCLE and GDSC panels to make it comparable with TCGA tumors (see Methods), we applied the TP53 classifier and obtained ranked scores. As a positive control, the classifier assigned a significantly higher TP53 phenocopy score to TP53 mutated cell lines (mean TP53_wt = 0.43, TP53_mut = 0.83, p = 1.1e-49 by t-test), therefore cell line data served as an independent validation set for the classifier. Of the 610 cell lines labeled as TP53 mutant based on genomic sequence (see Methods), 87% were classified as having a TP53-loss phenotype (Fig. 4a), suggesting a reasonable ability of the classifier trained on TCGA tumor transcriptomes to generalize to cell line data.

Similarly as in tumors, a notable fraction of cell lines were apparent false positives, labelled as TP53 wild-type by the genomes, but classified as TP53 deficient using the phenocopy score. We subdivided these apparent false positives into a high-confidence phenocopying set and the rest; the TP53 phenocopy score of the TP53 deep-deleted tumor samples was used as the threshold (see Methods). The high-confidence set was composed of 76 cell lines (FDR = 18%, see Methods, Fig. 4 B). Only 79% of the cell lines labelled as TP53 wild-type genetically were also classified as TP53 wild-type by the phenocopy score, suggesting that TP53-loss phenocopying events are common among cancer cell lines. For comparison, this percentage was 77% in cancer samples.

Some of the apparent false positive cell lines had a MDM2, MDM4 or PPM1D amplification or a USP28 deletion (43 out of 109, 39% of the high-confidence set). Cell lines harboring one of these CNA in known phenocopying genes had higher scores than the rest of TP53 wild-type cell lines (mean score = 0.58 and 0.37, respectively; t-test p = 5.4e-5, Additional file 1: Fig. S7a). Cells harboring a TP53 deep deletion (90th percentile of CNA scores) also had higher phenocopy scores than samples without deletion (mean score = 0.78 and 0.33, respectively, t-test p = 5.4e-8). 28% of the cell lines in the high-confidence harbor a TP53 deep deletion (22 out of 76, 90th percentile of TP53 deletion CNA). These data support that the apparent false positives are often bona fide TP53 phenocopying events in cancer cell lines. All TP53 phenocopy scores and cell line functional TP53 status information are provided in Additional file 3: Data S6.

Effects of TP53 on drug resistance are clarified by TP53 phenocopy scores

Next, we considered the GDSC drug response distributions for various drugs, in light of the TP53 functional status as determined by the TP53 mutations, and additionally as determined by our TP53 phenocopy scores. To identify drugs to which response is affected by TP53 functional status, we predicted drug response (log IC50) values of 449 GDSC drugs individually, using TP53 status as an independent variable (see Methods).

For most of the tested drugs (105 out of 188 drugs that were significantly associated at < 25% FDR, pan-cancer), the associations with TP53 had a lower FDR when testing using TP53 phenocopy score, over when using the TP53 CFE labels (denoting TP53 mutations which alter gene function) (Fig. 4c, effect size at Additional file 1: Fig. S7b). For the drugs that affected pathways related to TP53, this effect of improved significance by using the phenocopy score was prominent (hits with FDR by TP53 phenocopy score < FDR by TP53 CFE labels: DNA replication, 12/12 drugs; genome integrity, 8/10; p53 pathway, 3/5; apoptosis regulation, 4/6; cell cycle, 4/7; Additional file 1: Fig. S7c). As a negative control, randomized TP53 labels were not significantly associated with any drug. As a positive control, the drugs known to be affected by TP53 status such as nutlin-3a (effect size = 1.48 vs 1.01, p = 6.7e-68 vs 1.2e-44) or bleomycin (effect size = 0.25 vs 0.16, p = 0.009 vs 0.07), exhibited a stronger association with the TP53 score than with TP53 CFE mutation labels (Fig. 4c).

Next, we examined the IC50 drug sensitivity values towards all drugs together, considering the different groups of cell lines defined by our TP53 functional status classifier (Additional file 1: Fig. S7d). Here, the mean IC50 values of high-confidence TP53 phenocopying cell lines is more similar to the TP53 mutated cell-lines than to the TP53 wild-type cell lines. In drugs known to be affected by TP53 status, such as bleomycin (Fig. 4d), IC50 values were not notably different between TP53 mutant and the TP53 phenocopying high-confidence cell lines. All drug associations effect size and p-value are plotted in Additional file 1: Fig. S8a,b. Cancer type-specific associations are shown at Additional file 1: Fig. S9.

Taken together, the above analyses support the utility of the phenocopy score in identifying TP53-associated drug sensitivity or resistance, and also support that our tumor-derived classifier is able to generalize to cancer cell line transcriptomes to detect a phenotype of functional TP53 loss.

Associations between drug sensitivity and genetic markers are modified by TP53 status

A central goal in personalized cancer medicine is to discover actionable mutations, which are used as genetic markers to decide which therapy to apply. Based on the role of TP53 mutations in dysregulating various processes relevant to tumorigenesis, we hypothesized that various druggable cancer vulnerabilities may be conditional upon TP53 functional status. To investigate, a regression was fit to predict activity (log IC50) for each drug, from cancer type and each cancer gene mutation status (via the CFE classification, see Methods), and additionally introducing TP53 status (either via TP53 mutation (“CFE”), or via TP53 phenocopy score) as an interaction term in the regression. The TP53 phenocopy score was binarized (> 0.6 used as threshold) to be able to compare fairly with the TP53 mutation status. Comparing TP53 phenocopy association FDRs against TP53 mutation association FDRs suggested that the application of phenocopy score allowed to more confidently identify the drug-gene associations where TP53 status modulates the effect size (see the comparison of FDR values in Fig. 5a, broken down by pathway that targets the drug). Out of the identified three-way associations (gene * drug * TP53 status), 34% were identified only by using the TP53 phenocopy score, but not by the TP53 mutation status (Fig. 5a). For comparison, only 15% three-way associations were uniquely identified by TP53 mutation status but not by phenocopy score. We provide a tally of all gene-drug associations that were conditional upon TP53 in Additional file 1: Fig. S10a, and a per-gene comparison of associations identified with TP53 phenocopy score labels, versus those identified by TP53 mutational status, in Additional file 1: Fig. S10b.

Next, we aimed to select the more robust associations. To this end, we applied the “two-way” testing approach to identify replicated drug-marker links [55] seen across two or more drugs that share the same target protein or pathway. These were tested separately for individual cancer types, comparing TP53-deficient versus wild-type cell lines. Here, this “two-way” randomization test [55] was further modified to be able to detect interactions with a third factor, the TP53 functional status. As an additional criterion ensuring confidence of associations, only the hits that recur in more than one cancer type were taken into consideration (as a trade-off, this will cause tissue-specific associations to be missed). Stratifying by TP53 functional status, we identified a number of drug-gene CFE associations that would not be significant if ignoring the TP53 status during testing (60% of total, < 25% FDR, Additional file 1: Fig. S10c). This corresponds to a total of 2303 associations of a drug to specific gene mutational status by cancer type (total number of tests ignoring TP53 n = 486,417 versus n = 402,945 controlling for TP53 status, Additional file 1: Fig. S10c,d). Of these, 133 associations were found in both approaches, but had a lower FDR when considering TP53 stratification (mean FDR = 15% versus 19% if not stratifying, p = 5e-08); all associations from the “two-way” replication test are listed in Additional file 4: Data S7.

Sensitizing effects of driver mutations to HDAC and ATR inhibitors are modulated by TP53

Studies suggested a role of the drug AR-42 (a HDAC1 inhibitor) in prolonging p53 life and triggering apoptosis [56, 57]. We hypothesized that, if p53 activity is impaired, this effect of HDAC inhibitors may be altered. Interestingly, our testing reveals that mutations in the PIK3CA oncogene are associated with sensitivity to HDAC1 inhibition in a manner conditional upon TP53 mutation. In other words, when TP53 is functional, the resistance to HDAC1 inhibitor AR-42 due to PIK3CA mutation is higher than when TP53 is mutant or otherwise inactivated as indicated by phenocopy score (TP53 wild-type *A PIK3CA_mut regression coefficient test p = 0.005, Cohen’s d = 1.3, TP53 mutant PIK3CA regression coefficient test p = 0.08, Cohen d = -0.38, Fig. 5b). We would not identify this association with AR-42 while ignoring TP53 status (test on regression coefficient only using PIK3CA mutation status p = 0.67, Cohen d = -0.08). In particular, in LUAD the difference in AR-42 sensitivity (median of normalized log IC50 across cell lines) between PIK3CA mutant and PIK3CA wild-type is hardly evident: 0.26 versus 0.24 respectively, while in TP53-functional LUAD this difference is -0.43 (PIK3CA wild-type) versus 0.35 (PIK3CA wild-mutant). This response is observed across three different HDAC inhibitors and in three different cancer types. For instance, AR-42 and belinostat were found significantly associated with PIK3CA mutation in HNSC + LUSC (here considered jointly because of known molecular similarities of the cancer types), in BRCA, and in LUAD cancer types (Fig. 5b). Similarly, the AR-42 association with PIK3CA mutation was supported in the HDAC1-targeting drug CAY10603 (Additional file 1: Fig. S11b). Furthermore, in an independent drug screening dataset, PRISM [51], we were able to recover these associations (Additional file 1: Fig. S11b). This example illustrates how being aware of TP53 functional inactivation status allows to detect drug-gene associations that may be specific to the TP53 wild-type or to the TP53 deficient backgrounds but not both.

We also noted that the HDAC1i resistance/PIK3CA mutation association (conditional upon TP53 functional status) was only recovered when controlling for TP53 phenocopy score, but not when using simply the TP53 mutation status (per CFE method, see Methods) as an interaction term (belinostat IC50-PIK3CA mutation Mann–Whitney test, in the TP53 mutation wild-type background p = 0.13, while in TP53 w.t. phenocopy labels background p = 0.01, Fig. 5b). This example illustrates how the use of TP53 phenocopy scores provides additional power to identify drug-gene associations, as already indicated by the comparison of FDR scores for many associations above (Fig. 5a).

Recent reports have pointed out the potential therapeutic benefit of ATR inhibitors such as VE-821 or VE-822 in PTEN-deficient breast, glioma and melanoma cells [58, 59]. ATR is a crucial kinase regulating DNA repair and safeguarding genome integrity. ATR inhibition in PTEN-deficient cells was associated with accumulation of DSBs, cell cycle arrest and induction of apoptosis [58, 59], thus based on these phenotypes we hypothesized that the TP53 functional status may modulate this effect. Inspecting our data supports that the ATR inhibitors VE-821, VE-822, and AZD6738 were associated with a lower fitness in PTEN-mutant cells of the SKCM, OV, BRCA and DLBC cancer types (Fig. 5c, Additional file 1: Fig. S11d). This effect was however revealed only when TP53 status was taken into consideration, since p53-deficient cells had an increased survival that obscured this association of PTEN and ATRi (Fig. 5c, Additional file 1: Fig. S11d). While significance of the TP53 interaction term was not reached in this particular example, probably because the number of cell lines with a PTEN mutation (but TP53 wild-type) was low, nonetheless the association of ATRi IC50 values was found to be more robustly significant in a TP53 wild-type context than in a TP53 deficient context. This means there was a more prominent difference in cell fitness upon ATRi treatment comparing PTEN-mutated to PTEN wild-type cells in a TP53-proficient background (TP53 wild-type IC50-PTEN Cohen’s d = -0.41 vs TP53 deficient AZD6738 IC50-PTEN Cohen's d = -0.05).

Overall, above we highlighted two examples where TP53 functional status modulates the association between HDAC1 inhibitors and PIK3CA driver mutations, and ATR inhibitors and PTEN driver mutations. There were however many other significant associations involving TP53 status, cancer driver gene mutations (or CNA) and activity of drugs (listed in Additional file 4: Data S7), for example the association between PIK3R1 mutations and sensitivity to MET inhibitors (Additional file 1: Fig. S11c).

To estimate the importance in considering TP53 status in discovering drug associations more generally, we considered overlap in associations recovered when TP53 status was accounted for versus when TP53 status was ignored. Only 14% of significant associations of a given molecular target to driver gene alteration status were shared between two analyses (Additional file 1: Fig. S10c), indicating that considering TP53 status strongly alters the drug-gene links that can be recovered from statistical testing of drug screens on cancer cell lines. The TP53 status-aware testing recovered a higher number of associations (n = 12,150 versus 7853, both at < 25% FDR). We also noted this effect depended on the particular gene: drug responses in genes such as KRAS or TP53BP1 are well explained by gene mutational status alone, not benefitting from TP53 stratification (Additional file 1: Fig. S11a). Nevertheless, for most genes, their drug associations were more confidently retrieved when TP53 status was accounted for (e.g. BRAF, HRAS, ATM, APC; n = 18 genes total). Overall, the above data suggests that TP53 functionality should be considered when matching drugs to cancer patients based on the various driver gene mutations in their tumor, and that this TP53 functional status should preferentially be estimated via a phenocopying score rather than TP53 gene mutations.

Discussion

Disabling the master tumor suppressor gene TP53 provides cancer cells with important advantages such as avoiding cell cycle arrest or apoptosis upon replication stress or DNA damage. Because TP53 acts as a transcription factor controlling expression of hundreds of genes, a functional read-out of TP53 activity can be obtained using gene expression data, both at the level of RNA or at the protein level [20,21,22,23]. These gene expression-based scores of TP53 function have potential clinical relevance in predicting cancer aggressiveness/patient survival and therapy response [22, 23, 60, 61]. In this study, we developed a robust global transcriptome score of TP53 deficiencies, and applied it to ~ 7,000 tumors and ~ 1,000 cancer cell lines, to answer three questions.

Firstly, we asked how common are the TP53-mutation phenocopying events across human cancers. We estimated a 12% frequency of TP53 loss phenocopies, compared to a 58% prevalence of TP53 mutant tumors. In some cancer types such as BRCA and BLCA, the TP53 phenocopies may constitute a high fraction of 19% and 16% tumor samples, respectively, suggesting that the TP53 status of tumors should preferentially be measured via a functional readout (here, transcriptome-wide signature) rather than considering only the TP53 gene mutations. Supporting this notion, a recent study using a four-gene expression signature of TP53 activity demonstrated that this significantly predicts patient survival across 11 cancer types, and that in the majority of those it performs better than considering TP53 mutations [22].

Secondly, given the high prevalence of TP53 phenocopies, we asked if there exist additional genetic events that are associated with these phenocopies. We developed a method that considers CNA profiles jointly with gene expression in tumors, further integrating experimental data from CRISPR and RNAi screens, which confidently identified the USP28 gene deletion as a common TP53-loss phenocopying event. This is relevant for at least five cancer types: BLCA, STAD, BRCA, LIHC and LUAD, and affects 2.9%-7.6% tumor samples therein. The same statistical methodology also highlighted additional genes neighbouring the known phenocopies MDM2 and PPM1D – the CNOT2 and MSI2 genes, respectively – which are often co-amplified with the ‘primary’ gene in the CNA gain segment and may boost the resulting TP53-loss phenotype. Our analysis provides an example illustrating a more general principle of how molecular phenotypes (here, a transcriptional signature and fitness effects from a CRISPR screen) can be used to identify causal genes in a CNA segment. Our consideration of CRISPR k.o. screening data allowed us to circumvent the confounding effects of genetic linkage within the recurrent, phenocopying CNA segments, because the CRISPR reagent effects are not strongly coupled to the localization of the genes. Related genomics methodologies could be applied in future work to interrogate various recurrent CNA events observed in tumors, for which the causal gene(s) are often not known with confidence.

Thirdly, we asked if a better measurement of the TP53 functional inactivation status may be helpful for predicting cancer cell response to antitumor drugs based on genetic markers. Given that TP53 deficiencies have myriad downstream consequences on the cell, including e.g. suppression of cell cycle checkpoints, or inactivation of various DNA repair pathways [4] it is conceivable that the TP53 background may affect the activity of various drugs to kill cancer cells, including drugs targeted towards a particular driver mutation outside of TP53 itself. We searched large data sets for three-way statistical interactions involving TP53 status, each drug activity, and each mutated cancer driver gene. This suggested for instance that TP53 status modulates the selective resistance towards HDAC1-inhibitor drugs in PIK3CA-mutant cells. The associations we identified were filtered to retain those supported in multiple compounds targeting the same protein or pathway; enforcing agreement across multiple measurements may allay concerns of reproducibility in cell line screening databases [62,63,64]. Recent work by us and others [55, 65] has used statistical methods to integrate over diverse screening datasets, considering drug and CRISPR genetic screens jointly, thus improving reliability of drug-target association discovery. Our robustly supported set of drug-target gene links that may be modulated by TP53 status (Additional file 4: Data S7) provides a comprehensive resource for follow-up validation of the TP53 functional status in modulating gene-drug associations.

The statistical model that we employed to identify TP53 loss phenocopying events draws on the expression levels of 217 genes, and is largely portable across various human tissues. Given that the model’s predictive accuracy is high (demonstrated using cross-validation and using an independent data set of cancer cell line transcriptomes), the errors it makes may be of interest. While the apparent false-positives are often TP53 loss phenocopies using still-unknown mechanisms, extending the USP28 example addressed in this study, it would also be interesting to look into the apparent false negatives in future work. These TP53-mutant tumors classified as TP53 wild-type-like by our transcriptome score were not considered here, because of their relatively modest number, making statistical analyses difficult. Going forward, analyses of genomes from larger cohorts of cancer patients may provide enough examples to reveal mechanisms of re-establishing activity of mutated TP53 in certain cancers. Conceivably, this may happen by normalizing expression of the TP53-downstream genes which have been dysregulated by the TP53 mutation; understanding these events may inspire new avenues for therapy of TP53 mutant tumors.

The general statistical approach presented here could be applied beyond TP53 also to other driver gene phenocopying events which may occur in tumors. For instance, RAS pathway activation transcriptomic scores were proposed [21], and similarly so the homologous recombination DNA repair scores were proposed based on mutational signatures [66, 67]. Conceivably, other cancer pathways may be similarly addressed as well, systematically analyzing their distribution across tumors to identify phenocopying events, as well as their implications to drug response prediction, as we have done here for TP53 phenocopies.

Conclusions

Our study provides insights into the prevalence and implications of TP53 deficiency in human cancers, which often happens by phenocopying, thus highlighting the need to consult a functional readout of TP53 activity in precision medicine efforts. Integration of a transcriptomic signature of TP53 inactivity with experimental data from CRISPR and RNAi screens identified USP28 gene deletions as a common TP53-loss phenocopying event in tumors, and in addition we identified auxiliary genes neighbouring MDM2 and PPM1D that may enhance their phenocopying effect. Furthermore, our study suggests that a measurement of TP53 functional status can improve the prediction of cancer response to antitumor drugs, such as HDAC and ATR inhibitors.

Methods

Data collection and preparation

Gene expression and Copy Number Alteration (CNA) data

We downloaded gene expression data (transcripts per million, TPM) from GDC Data Portal [68] for human tumor samples (TCGA) and from GDSC [51] and CCLE [69] for cell line samples (CL).

We filtered out genes with missing values in more than 100 samples and selected the overlapping genes between cell lines and tumors. Cancer types with less than 10 samples were filtered out. Low expressed genes were removed (TPM < 1 in 90% of the samples) and applied a square-root transformation to TPM. Cancer types. Tumors with less than 10 samples were filtered out. In total, we have 12,419 features for 966 CL samples and 9149 TCGA samples.

Tumor purity estimates were obtained from TCGA ABSOLUTE (using Tumor.purity function, R package TCGAbiolinks 2.25.3). Samples with top quality (ABSOLUTE score = 86) were removed for correlation testing.

We collected CNA from GDC Data Portal [68] for TCGA samples and from DepMap [63] for CL samples.

For external validation, the RNA-Seq data and CNV data for PCAWG [70] was downloaded using Xenahub [71]. Samples shared between TCGA and PCAWG datasets were removed from PCAWG. Only the genes present in the TCGA dataset were kept in PCAWG RNAseq data. RNAseq data was square root transformed. Cancer types that were considered equivalent between the two datasets were the following: ESAD/ESCA, CLLE/DLBC, MALY/DLBC, RECA/KIRC, PACA/PAAD, LIRI/LIHC. In total, we used 12,419 gene expression features for 555 PCAWG tumors across 7 cancer types.

Data alignment between tumors and cell lines

In order to later generalize the model to cell lines we proceed to align TCGA and CL data. For this, we applied ComBat, a batch adjustment method, to account for intrinsic differences between tumor signal and cell lines signal [54]. For the alignment of TCGA and CL data, we first applied quantile normalization (normalize.quantiles function, preprocessCore R 1.48.0 package) using tumor data as reference and then applied ComBat (ComBat function, R package sva 3.32.1). Each group (TCGA, GDSC or CLLC) was treated as a different batch. We proceed similarly using PCAWG gene expression data.

TP53 status label (according to GDSC)

TCGA Pan-Cancer Atlas somatic mutation data were extracted from the MC3 Public MAF (v0.2.8) data set [72]. We followed the Iorio et al. methodology [73] to determine bona fide TP53 mutations (0:wild type, 1: mutated). We identify recurrent variants that are likely to contribute to carcinogenesis. We considered mutated variants: non-synonymous missense mutations, indels (in frame insertions and deletions and out of frame insertions and deletions), nonsense mutations and specific splice-site mutations (such as “p.X125_splice”). Samples without any of these mutations annotated were considered TP53 wild type. In only 5% of the cases (179 out of 3416) our labels differed from the ones provided by Iorio et al. In total, we obtained TP53 labels for 7788 TCGA tumors. TP53 variant pathogenic scores were obtained from the EVE data portal [74] and from VARITY data portal [75]. For the validation dataset, TP53 mutations were downloaded from PCAWG XenaHub (simple somatic mutation (SNVs and indels)—coding driver mutations, [70]).

TP53 score classifiers in human tumors

We used the aligned human tumor data to train a supervised elastic [20,21,22,23] net penalized logistic regression (using cv.glmnet function with alpha = 0.5, R package glmnet 4.0–2) classifier with cyclical coordinate descendent optimization [76]. The choice of Elastic net penalization aims to deal with two concerns: the large number of variables can lead to high complexity (overfitting) and the feature multicollinearity. Elastic net regressions are seen as a good trade-off that benefit from the dimensionality reduction provided by Lasso penalization while keeping as many informative variables as possible (Ridge penalization). Of note, these three regularization methods yielded similar cross-validation accuracy: Elastic net (i.e. alpha = 0.5) AUC 0.960, Lasso (i.e. alpha = 1) AUC 0.965, and Ridge (i.e. alpha = 0) AUC 0.952, suggesting that the default alpha = 0.5 in Elastic net method is a reasonable choice. The model is trained using RNAseq data (X matrix) to infer TP53 status (Y matrix). As a reference (Y) during training we used TP53 mutation status labels.

For the training set, we excluded the tumor samples that have an amplification (not neutral, > 0, according to GISTIC CNA thresholded calls downloaded using FirebrowseR package, Analyses.CopyNumber.Genes.Thresholded function) in previously known TP53 phenocopying genes (MDM2, MDM4, PPM1D) or a deep deletion of TP53 (n = 2065), to prevent the model from relying too much on dosage effects of these genes, instead of the downstream response. Known phenocopies and TP53 gene expression were removed from the training dataset with the same purpose.

In addition, to control for cancer type specific signals we included cancer type as a dummy variable. To control for class imbalance, we included weights in the classifier.

The model learns a vector of gene-specific weights that better classifies TP53 status. The score from the models determines the probability of a given tumor of being TP53 deficient. Optimization of the penalized regression formula and further details of the classifier can be consulted at [76].

Leave-one-cancer type-out (LOCTO) classifier

We used PCAWG gene expression data aligned with TCGA data using ComBat to train an Elastic Net classifier with the same features and structure as above. The model is trained using RNAseq data (X matrix) to infer TP53 status (Y matrix). In this case the models were fit excluding from the training data all examples from one cancer type, iterating over cancer types. Metrics and curves from Additional file 1: Fig. S1a were calculated using the held out cancer type data.

Assessment of the classifier and calculation of FDR score

Using 90% of the training set and 5 balanced folds (balanced based on TP53 mutational state) we performed cross-validation. We measured the performance of the training set (folds used for training) and the testing set (10% held out). Areas under the Receiving Operating Curve (AUROC) and the Precision Recall curve (AUPRC) were calculated for both training (cross-validation) and testing sets.

FDR was calculated by sample using each sample probability score from the classifier as threshold for determining positive and negative samples FDR = false positive / (false positive + true positive). Samples harboring an amplification (GISTIC thresholded amplifications, FirebrowseR package, Analyses.CopyNumber.Genes.Thresholded function) of known phenocopying genes (MDM2, MDM4, PPM1D) or TP53 deletions (GISTIC thresholded deep deletions, FirebrowseR package, Analyses.CopyNumber.Genes.Thresholded function)) were considered as true positives when calculating FDR.

In Fig. 1b, density of known phenocopies was calculated using MDM4, MDM2, PPM1D (amplifications) and TP53 (deletions) CNA over/under the 95/0.05 th quantile. All TP53 Phenocopy scores (probabilities of being TP53 dysfunctional) are provided at Data S2.

For the validation data (PCAWG), the same procedure was followed. In this case, we used the downloaded consensus CNV data to determine samples that were amplified for MDM4/MDM2/PPM1D or deep-deleted for TP53.

Precision and recall in the TCGA and PCAWG dataset were obtained similarly to FDR. MDM2/PPM1D/MDM4 amplification and TP53 deep deletion were considered to be positive cases.

The classifier coefficients were analyzed using the GO enrichment tool ShinyGO [77]. The 12,419 genes from the gene expression matrix with a coefficient equal to zero were used as background. Full classifiers relevant coefficients are provided at Data S1.

The coefficients of the TP53 model should be interpreted with care, for several reasons: some of these genes may change in expression levels via indirect association meaning they may not be directly regulated by TP53; the gene set may omit genes that are bona fide TP53 targets if the information contained in them is redundant with other genes; and finally these genes may individually be only weakly associated with TP53 status, since the method optimizes the expression markers’ collective power. Visualization was performed using Revigo [78].

TP53 status detection in cell lines

Using the downloaded RNAseq from GDSC cell lines data we applied our trained tumor classifier to cell lines. As stated above, RNAseq data was square rooted, normalized and ComBat batch corrected. Cell line prediction performance was measured using as reference TP53 COSMIC labels [79] combined with Iorio et al. methodology [73] as we did in tumors. FDR was calculated again using samples harboring an amplification of known phenocopying genes (MDM2, MDM4, PPM1D) or TP53 deletions as true positives.

Using the classifier scores we separate the cell lines high-confidence set (FDR < = 18%) using as threshold reference GISTIC tresholded TP53 deep deletions (-2) (threshold = 0.93) (FirebrowseR package, Analyses.CopyNumber.Genes.Thresholded function). Therefore, we determine 3 sets derived from our Phenocopy score: high-confidence set (predicted TP53 phenocopies, classified as mutant but originally labeled as wild type), TP53 mutant (classified and labeled as mutant) and TP53 wild type (classified and labeled as wild type). All cell line predictions are provided at Data S3.

Due to a lack of positive controls, samples that were classified as wild type being originally labeled as TP53 mutant were not considered further. However, in the future, analyses with a higher number of cancer genomes may reveal mechanisms of re-establishing TP53 activity in some TP53 mutant cancers (e.g. by normalizing expression of the TP53-downstream genes which have been dysregulated by the TP53 mutation).

Gene codependency with TP53 knockout/knockdown in cell line screens

Following data of the 2021 Q4 release downloaded from the DepMap project website: CRISPR data from PROJECT Score [26] (“Achilles_gene_effect.csv”), combined RNAi from DEMETER2 scores repository [27] (“D2_combined_gene_dep_scores.csv”), and the cell line metadata (“sample_info.csv”). In this data, negative scores imply cell growth inhibition and/or death following gene knockout.

CRISPR data is normalized so non-essential genes scores are close to 0. We used Pearson's correlation to correlate the gene effect of CRISPR TP53 knockout in every cell line to other genes' effect. We tested 990 cell lines for 12,419 genes. This correlation score was calculated both pan-cancer and by cancer type.

Similarly as with CRISPR codependency data we correlated gene knockdown effect with TP53 knockdown (RNAi) effect using Pearson's correlation test. We tested 700 cell lines for our 12,419 genes. This score was calculated both for pan-cancer and by cancer type.

Calculation of the combined gene prioritization score

We sought to rank possible TP53 loss phenocopying genes testing different data: copy number variant data, gene expression data (RNAseq), RNAi codependency score and CRISPR codependency score. We used the downloaded tumor data (previously described) and our TP53 Phenocopy score to test for differences across our 3 main TP53 groups: TP53 wild type (labeled and classified as wild type), TP53 mutated (labeled and classified as mutated) and predicted TP53 phenocopied (labeled as wild type but classified as mutated). We guessed that phenocopying genes should have a differential expression in the phenocopies group when comparing to wild type and mutated TP53 groups individually. We tested 12,419 genes (by cancer type) in the following manner (via Student's t-test):

CNV_gene_TP53_wt versus CNV_gene_TP53_phenocopies (CNV0 test),
CNV_gene_TP53_mut versus CNV_gene_TP53_phenocopies (CNV1 test)
GE_gene_TP53_wt versus GE_gene_TP53_phenocopies (GE0 test)
GE_gene_TP53_mut versus GE_gene_TP53_phenocopies (GE1 test)
RNAi_score_gene versus RNAi_score_TP53 (RNAi codependency score, methodology described above)
CRISPR_score_gene versus CRISPR_score_TP53 (CRISPR codependency score, methodology described above).

Three thousand ten genes did not have gene expression data so GE1 and GE0 tests were omitted from the combination test for those genes. We combined the p-values values from all available tests by cancer type using Fisher's method for combining p-values. For each category (CNV and GE) we only use in the combination the worst p-value (i.e. maximum) between CNV0 and CNV1 test, and separately between GE1 and GE0 test, as a way of aggregating. Genes in which the test direction is not coherent in CNV, GE and codependency score were dropped. A gene with a negative codependency score, such as a negative regulator such as MDM2, is expected to cause a phenocopy of TP53 by amplification and overexpression (therefore a higher expression in the phenocopies group that TP53 wt or mut). P-values were FDR adjusted using Benjamini–Hochberg method (p.adjust function of the stats R package). We further merged each cancer type combined p-values into one single p- value using Fisher's approach, and FDR adjusted. That way we obtained the final Prioritization score for each gene in a cancer-combined way. Pearson R score of CRISPR/RNAi codependency was merged across cancer types by taking its mean value.

We set as reference points the known phenocopying genes’ (MDM2, MDM4, and PPM1D) FDRs and CRISPR codependency scores. To establish a stringent threshold for new candidate phenocopying genes, we required that a gene's prioritization score should have an FDR as significant as the best-ranked known phenocopying gene (considered by cancer type). Same requirement was applied for the CRISPR codependency score. For example, the known phenocopying genes with the best score by cancer type was MDM4 in the LUAD cancer type, with an combined FDR of 4e-05 and a CRISPR codependency score of -0.21. Finally, we additionally applied a pan-cancer CRISPR and RNAi codependency score threshold, requiring genes to have at least a modest correlation (absolute value > = 0.1) between the TP53 knockout/knockdown fitness a candidate gene knockout fitness in both CRISPR and RNAi screens (codependency score); this filter removes many of the passenger neighboring genes that are co-amplified with known phenocopying genes.

TP53 wild-type and TP53 -/- isogenic cell line genetic screens

Mean beta scores were calculated using MAGeCK-MLE [80] for TP53-isogenic pair cell lines A549 [81] and two RPE1 cell lines [82, 83]. Beta scores represent the effect that gene knock-out has on cell fitness.

We calculated the Z-scores (distance from the mean expressed as number of standard deviations) of either USP28 or ATM within the distribution of their respective neighbor genes, for each dataset and TP53 status "1Mbp neighbor genes" are genes present in Brunello [84] and Gecko v2 [85] libraries and located within a 1Mbp window surrounding either USP28 or ATM, obtained from genecards.weizmann.ac.il.

Drug response associations with TP53 status

We collected GDSC [73] drug data for a total of 1000 cell lines. We used IC50 as a measure of activity of a compound against a specific cell line. If drug data was available in both GDSC1 and GDSC2 versions, GDSC1 data was selected.

We also collected each drug putative target and target pathway information from the GDSC website (https://www.cancerrxgene.org/). We filtered out NA values and transformed IC50 scores to log scale. We downloaded GDSC mutational Cancer Functional Events (CFEs) [73] in order to: make comparisons between TP53 Phenocopy score and GDSC TP53 CFEs and to test other gene status-drug response associations, while controlling for TP53 status. Mutational CFEs consist of a GDSC curated set of cancer genes (CGs) for which the mutation pattern in whole-exome sequencing (WES) data is consistent with positive selection.

We first used drug response (IC50) values of 449 GDSC drugs to fit a pan-cancer regressions against TP53 status using cancer type as control variable. We fit three different regressions per drug response: against TP53 CFEs, against predicted TP53 Phenocopy thresholded scores and against TP53 random labels.

$$log\mathit(IC\mathit{50}\mathit)\mathit\sim TP\mathit{53}\mathit.status\mathit+cancer\mathit.type$$

For the TP53 status we used the groups obtained from our phenocopy score being the TP53 high-confidence phenocopying set (classified as mutant by the score, but labeled as wild-type genetically) and TP53 mutant set (classified as mutant, labeled as mutant) the TP53 deficient set (TP53.status = 1) and TP53 wild type (classified as wild-type, labeled as wild-type) as wild type set (TP53.status = 0). Due to uncertainty, we filtered out samples with a TP53 mutation classified as wild-type. Cancer types with less than 3 cases for any category were filtered out. We used the esc R package to calculate effect size (cohens_d function). P-values of associations were FDR corrected using the Benjamini-Hochberg (“fdr”) correction of the p.adjust function (stats package).

We separate the drugs into groups according to the pathway the gene they target belong to. By pathway, we calculated the slope resulting from the comparison of the FDR phenocopy score regression versus the FDR TP53 CFEs. For the visualization we plotted raw IC50 values of different drugs and all drugs together across the different cell line defined sets. For further analysis, we merged the cancer types that are thought to have some similarity: HNSC with LUSC (jointly known as HNSC_LUSC), GBM with LGG (LGG_GBM) and OV with UCEC (OV_UCEC).

Drug response associations of gene status controlling for TP53 status

We collected drug screening data from the PRISM project [52] and GDSC project [51]. NA values were filtered out and IC50 values were transformed to logarithmic scale. We downloaded mutation features (GDSC mutational CFEs, see above) from [73].

First, we fit a regression for each drug and gene CFE including TP53 loss Phenocopy score and the interaction term as it follows:

$$log\mathit(IC\mathit{50}\mathit)\mathit\sim genCFEs\mathit+TP\mathit{53}Phenocopy\mathit.status\mathit+genCFEs\mathit\ast TP\mathit{53}Phenocopy\mathit.status$$

For comparison, we performed the same analysis using TP53 random and TP53 CFEs instead of TP53 Phenocopy.status.

We tested every gene mutational CFEs out of the 300 genes provided by GDSC. We filtered out cases with lss than 3 samples in any category (mutated:1 or wildtype:0) for TP53 status and gen CFEs. Regressions were fitted by cancer type using glm package (glment 4.0–2 R package). We selected genCFEs p.value and FDR correct using the Benjamini-Hochberg (“fdr”) correction of the p.adjust function (stats R package). The coefficient of the genCFEs variable informs us about the fold change of the different variable states (mutant:1-wildtype:0) when TP53Phenocopy.status is set to its reference levels (wildtype:0). We compared these scores when using TP53 Phenocopy to TP53 CFEs by plotting FDR values and calculating slope (Fig. 5a, Additional file 1: Fig. S10a).

Two-way association tests for gene-drug associations modified by TP53

To further analyze TP53 interaction in a more stringent way we implemented a version of the “two-way association test” approach recently developed by us and reported in Levatic et al. [55]. In this methodology we enforced that, for a given drug, an association between a GDSC gene feature (e.g. gene mutation CFEs) and GDSC drug response is reproduced in other drugs with the same molecular target; here this is additionally controlled by TP53 status.

For this, we curated 996 sets of two drugs with the same target (ie: Dabrafenib and AZ628, target = BRAF). For the two drugs separately, we fitted a regression comparing the GDSC drug response against gene mutational status (CFE) controlling for TP53 status (as stated above) by cancer type. We tested the different labels in the regression: TP53 CFEs, TP53 random labels and TP53 phenocopy labels. We considered associations by cancer type. We calculated the two-way association score by averaging the estimates (effect size) obtained between drug 1 and drug 2. To calculate the p-value for each drug-drug combination, we shuffled the TP53 labels and compared the obtained random estimates with the actual estimate as described in our previous work [55].

For an association to be selected, we require that it is observed in more than one cancer type (merged cancer types excluded), FDR < 25% across all cancer types where the hit is observed and that the direction (value from gene CFEs variable estimate) is maintained across drugs. When selecting relevant hits we also required that each hit TP53 interaction term variable in regression is significant (FDR < 25%). This informs us of deviation from the behavior of the regression variables gen_status = 1 and gen_status = 0 when TP53 is controlled as interaction. We filtered out cases with less than 3 samples in any category (mutated:1 or wildtype:0) for TP53 status and gene CFEs in a cancer type-specific manner. Supported hits by this methodology are reported at Fig. 6 b,c, Additional file 1: Fig. S10c, Additional file 1: Fig. S11a and e and in Additional file 4: Data S7.

In addition, as a validation for some hits we performed a “two-way” using PRISM data. In this case we enforced that, for a given drug, an association between a gene feature (GDSC gen mutational CFEs) and GDSC drug response is reproduced in the same drug using the PRISM dataset. The rest of the methodology was applied in the same manner (see GDSC “two-way test” above).

As control, we followed the same procedure of the two-way testing method but fitting regressions of IC50 ~ gene CFEs (without interaction term). FDR corrected p-values of gen CFEs coefficient in regressions with and without interaction term were compared. We made different types of comparisons: by gene associations (Additional file 1: Fig. S10b), molecular target-gen CFEs associations (different 2-sets of drugs can target the same molecular feature) and all associations (Additional file 1: Fig. S10a).

Availability of data and materials

In this study published datasets were reanalyzed. TCGA RNAseq gene expression levels were downloaded from the TCGA repository at NCI Genomic Data Commons [GDC Data Portal https://portal.gdc.cancer.gov/ (2016)]. TCGA CNV data was downloaded using the Firebrowser R package [https://github.com/mariodeng/FirebrowseR] using the Analyses.CopyNumber.Genes.Thresholded function. Tumor purity data was downloaded using the Tumor.purity function from TCGAbiolinks 2.25.3 [https://github.com/BioinformaticsFMRP/TCGAbiolinks]. PCAWG RNAseq and CNV were downloaded using the Xena browser [donor centric miRNA expression and copy number,https://pcawg.xenahubs.net]. Cell line RNAseq data was downloaded from CLLE [21Q4 CCLE_expression.csv https://depmap.org/portal/download/all/ (2021)] and GDSC [Expression array, https://www.cancerrxgene.org/downloads/bulk_download (2014)]. Cell line CNV data was downloaded from Depmap [21Q4 CCLE_gene_cn.csv https://depmap.org/portal/download/all/ (2021)].Somatic mutation calls for TCGA were downloaded from the MC3 project [mc3.v0.2.8.PUBLIC.maf.gz https://gdc.cancer.gov/about-data/publications/mc3-2017 (2017)]. TP53 variant pathogenic scores were downloaded from EVE [TP53 EVE score https://evemodel.org/download/protein/P53_HUMAN (2021)] and Varity [gen name: TP53, http://varity.varianteffect.org (2021)]. PCAWG TP53 status was downloaded from Xena [simple somatic mutation (SNVs and indels) – coding driver mutations, https://pcawg.xenahubs.net]. Cell line gene mutation data (CFEs) and TP53 mutational CFEs were downloaded from GDSC [Mutational CFEs, https://www.cancerrxgene.org/gdsc1000/GDSC1000_WebResources/Clinically_Relevant_Features.html (2014)]. CRISPR PROJECT score data was downloaded from Depmap [2021Q4 Achillles_gene_effect.csv, https://doi.org/10.6084/m9.figshare.16924132.v1 (2021)]. DEMETER2 RNAi scores were downloaded from repository project [D2_combined_gene_dep_scores.csv, DEMETER2 data https://doi.org/10.6084/m9.figshare.6025238.v6 (2018)]. Drug data, drug putative target and target pathway information from the GDSC website [Drug Screening – IC50s, https://www.cancerrxgene.org/downloads/bulk_download (2014)] and from PRISM project [PRISM data https://doi.org/10.6084/m9.figshare.20564034.v1 (2022)].

Abbreviations

AUROC:: Area under the receiver operating characteristic
LOCTO:: Leave-one-cancer-type-out
AUC:: Area under the curve
AEN:: Apoptosis-enhancing nuclease
CFE:: Cancer Functional Events
CNA:: Copy number alteration
FDR:: False discovery rate
LoF:: Loss-of-function
ROC:: Receiver operating characteristic
TPM:: Transcripts per million

References

Rogers ZN, McFarland CD, Winters IP, Naranjo S, Chuang C-H, Petrov D, et al. A quantitative and multiplexed approach to uncover the fitness landscape of tumor suppression in vivo. Nat Methods. 2017;14:737–42.
Article CAS PubMed PubMed Central Google Scholar
Martincorena I, Raine KM, Gerstung M, Dawson KJ, Haase K, Van Loo P, et al. Universal Patterns of Selection in Cancer and Somatic Tissues. Cell. 2018;173:1823.
Article CAS PubMed PubMed Central Google Scholar
Olive KP, Tuveson DA, Ruhe ZC, Yin B, Willis NA, Bronson RT, et al. Mutant p53 gain of function in two mouse models of Li-Fraumeni syndrome. Cell. 2004;119:847–60.
Article CAS PubMed Google Scholar
Janic A, Valente LJ, Wakefield MJ, Di Stefano L, Milla L, Wilcox S, et al. DNA repair processes are critical mediators of p53-dependent tumor suppression. Nat Med. 2018;24:947–53.
Article CAS PubMed Google Scholar
Steele CD, Abbasi A, Islam SMA, Bowes AL, Khandekar A, Haase K, et al. Signatures of copy number alterations in human cancer. Nature. 2022;606:984–91.
Article CAS PubMed PubMed Central Google Scholar
Priestley P, Baber J, Lolkema MP, Steeghs N, de Bruijn E, Shale C, et al. Pan-cancer whole-genome analyses of metastatic solid tumours. Nature. 2019;575:210–6.
Article CAS PubMed PubMed Central Google Scholar
Fei P, El-Deiry WS. P53 and radiation responses. Oncogene. 2003;22:5774–83.
Article CAS PubMed Google Scholar
Hientz K, Mohr A, Bhakta-Guha D, Efferth T. The role of p53 in cancer drug resistance and targeted chemotherapy. Oncotarget. 2017;8:8921–46.
Article PubMed Google Scholar
Aas T, Børresen AL, Geisler S, Smith-Sørensen B, Johnsen H, Varhaug JE, et al. Specific P53 mutations are associated with de novo resistance to doxorubicin in breast cancer patients. Nat Med. 1996;2:811–4.
Article CAS PubMed Google Scholar
Rusch V, Klimstra D, Venkatraman E, Oliver J, Martini N, Gralla R, et al. Aberrant p53 expression predicts clinical resistance to cisplatin-based chemotherapy in locally advanced non-small cell lung cancer. Cancer Res. 1995;55:5038–42.
CAS PubMed Google Scholar
Lee JM, Bernstein A. p53 mutations increase resistance to ionizing radiation. Proc Natl Acad Sci U S A. 1993;90:5742–6.
Article CAS PubMed PubMed Central Google Scholar
Stengel A, Kern W, Haferlach T, Meggendorfer M, Fasan A, Haferlach C. The impact of TP53 mutations and TP53 deletions on survival varies between AML, ALL, MDS and CLL: an analysis of 3307 cases. Leukemia. 2017;31:705–11.
Article CAS PubMed Google Scholar
Hassin O, Nataraj NB, Shreberk-Shaked M, Aylon Y, Yaeger R, Fontemaggi G, et al. Different hotspot p53 mutants exert distinct phenotypes and predict outcome of colorectal cancer patients. Nat Commun. 2022;13:2800.
Article CAS PubMed PubMed Central Google Scholar
Barboza JA, Iwakuma T, Terzian T, El-Naggar AK, Lozano G. Mdm2 and Mdm4 loss regulates distinct p53 activities. Mol Cancer Res. 2008;6:947–54.
Article CAS PubMed PubMed Central Google Scholar
Li Q, Lozano G. Molecular pathways: targeting Mdm2 and Mdm4 in cancer therapy. Clin Cancer Res. 2013;19:34–41.
Article PubMed Google Scholar
Woodfield SE, Shi Y, Patel RH, Chen Z, Shah AP, Srivastava RK, et al. MDM4 inhibition: a novel therapeutic strategy to reactivate p53 in hepatoblastoma. Sci Rep. 2021;11:2967.
Article CAS PubMed PubMed Central Google Scholar
Kleiblova P, Shaltiel IA, Benada J, Ševčík J, Pecháčková S, Pohlreich P, et al. Gain-of-function mutations of PPM1D/Wip1 impair the p53-dependent G1 checkpoint. J Cell Biol. 2013;201:511–21.
Article CAS PubMed PubMed Central Google Scholar
Milosevic J, Fransson S, Gulyas M, Olsen TK, Gallo-Oller G, Treis D, et al. High Expression of PPM1D Induces Tumors Phenotypically Similar to TP53 Loss-of-Function Mutations in Mice. Cancers. 2021;13:5493.
Article CAS PubMed PubMed Central Google Scholar
Bulavin DV, Demidov ON, Saito S, Kauraniemi P, Phillips C, Amundson SA, et al. Amplification of PPM1D in human tumors abrogates p53 tumor-suppressor activity. Nat Genet. 2002;31:210–5.
Article CAS PubMed Google Scholar
Kenzelmann Broz D, Spano Mello S, Bieging KT, Jiang D, Dusek RL, Brady CA, et al. Global genomic profiling reveals an extensive p53-regulated autophagy program contributing to key p53 responses. Genes Dev. 2013;27:1016–31.
Article PubMed PubMed Central Google Scholar
Way GP, Sanchez-Vega F, La K, Armenia J, Chatila WK, Luna A, et al. Machine Learning Detects Pan-cancer Ras Pathway Activation in The Cancer Genome Atlas. Cell Rep. 2018;23:172-80.e3.
Article CAS PubMed PubMed Central Google Scholar
Miller LD, Smeds J, George J, Vega VB, Vergara L, Ploner A, et al. An expression signature for p53 status in human breast cancer predicts mutation status, transcriptional effects, and patient survival. Proc Natl Acad Sci U S A. 2005;102:13550–5.
Article CAS PubMed PubMed Central Google Scholar
Donehower LA, Soussi T, Korkut A, Liu Y, Schultz A, Cardenas M, et al. Integrated Analysis of TP53 Gene and Pathway Alterations in The Cancer Genome Atlas. Cell Rep. 2019;28:1370-84.e5.
Article CAS PubMed PubMed Central Google Scholar
Kawase T, Ichikawa H, Ohta T, Nozaki N, Tashiro F, Ohki R, et al. p53 target gene AEN is a nuclear exonuclease required for p53-dependent apoptosis. Oncogene. 2008;27:3797–810. GDSC data portal https://www.cancerrxgene.org/gdsc1000/GDSC1000_WebResources/Home.html 2014.
Article CAS PubMed Google Scholar
Dornan D, Wertz I, Shimizu H, Arnott D, Frantz GD, Dowd P, et al. The ubiquitin ligase COP1 is a critical negative regulator of p53. Nature. 2004;429:86–92.
Article CAS PubMed Google Scholar
Dwane L, Behan FM, Gonçalves E, Lightfoot H, Yang W, van der Meer D, et al. Project Score database: a resource for investigating cancer cell dependencies and prioritizing therapeutic targets. Nucleic Acids Res. 2021;49:D1365-72. https://doi.org/10.6084/m9.figshare.16924132.v1.
Article CAS PubMed Google Scholar
McFarland JM, Ho ZV, Kugener G, Dempster JM, Montgomery PG, Bryan JG, et al. Improved estimation of cancer dependencies from large-scale RNAi screens using model-based normalization and data integration. Nat Commun. 2018;9:4610. https://doi.org/10.6084/m9.figshare.6025238.v6.
Article PubMed PubMed Central Google Scholar
Boyle EA, Pritchard JK, Greenleaf WJ. High-resolution mapping of cancer cell networks using co-functional interactions. Mol Syst Biol. 2018;14:e8594.
Article PubMed PubMed Central Google Scholar
Giacomelli AO, Yang X, Lintner RE, McFarland JM, Duby M, Kim J, et al. Mutational processes shape the landscape of TP53 mutations in human cancer. Nat Genet. 2018;50:1381–7.
Article CAS PubMed PubMed Central Google Scholar
Zhang D, Zaugg K, Mak TW, Elledge SJ. A role for the deubiquitinating enzyme USP28 in control of the DNA-damage response. Cell. 2006;126:529–42.
Article CAS PubMed Google Scholar
Fong CS, Mazo G, Das T, Goodman J, Kim M, O’Rourke BP, et al. 53BP1 and USP28 mediate p53-dependent cell cycle arrest in response to centrosome loss and prolonged mitosis. Elife. 2016;5:e16270.
Article PubMed PubMed Central Google Scholar
Cuella-Martin R, Oliveira C, Lockstone HE, Snellenberg S, Grolmusova N, Chapman JR. 53BP1 Integrates DNA Repair and p53-Dependent Cell Fate Decisions via Distinct Mechanisms. Mol Cell. 2016;64:51–64.
Article CAS PubMed PubMed Central Google Scholar
Bernhard SV, Seget-Trzensiok K, Kuffer C, Krastev DB, Stautmeister L-M, Theis M, et al. Loss of USP28 and SPINT2 expression promotes cancer cell survival after whole genome doubling. Cell Oncol. 2022;45:103–19.
Article CAS Google Scholar
Lambrus BG, Daggubati V, Uetake Y, Scott PM, Clutario KM, Sluder G, et al. A USP28–53BP1–p53–p21 signaling axis arrests growth after centrosome loss or prolonged mitosis. J Cell Biol. 2016;214:143–53.
Article CAS PubMed PubMed Central Google Scholar
Phan TP, Maryniak AL, Boatwright CA, Lee J, Atkins A, Tijhuis A, et al. Centrosome defects cause microcephaly by activating the 53BP1-USP28-TP53 mitotic surveillance pathway. EMBO J. 2021;40:e106118.
Article CAS PubMed Google Scholar
Ito T, Kwon HY, Zimdahl B, Congdon KL, Blum J, Lento WE, et al. Regulation of myeloid leukaemia by the cell-fate determinant Musashi. Nature. 2010;466:765–8.
Article CAS PubMed PubMed Central Google Scholar
Sheng W, Dong M, Chen C, Wang Z, Li Y, Wang K, et al. Cooperation of Musashi-2, Numb, MDM2, and P53 in drug resistance and malignant biology of pancreatic cancer. FASEB J. 2017;31:2429–38.
Article CAS PubMed Google Scholar
Sheng W, Dong M, Chen C, Li Y, Liu Q, Dong Q. Musashi2 promotes the development and progression of pancreatic cancer by down-regulating Numb protein. Oncotarget. 2017;8:14359–73.
Article PubMed Google Scholar
Pusapati RV, Rounbehler RJ, Hong S, Powers JT, Yan M, Kiguchi K, et al. ATM promotes apoptosis and suppresses tumorigenesis in response to Myc. Proc Natl Acad Sci U S A. 2006;103:1446–51.
Article CAS PubMed PubMed Central Google Scholar
Hirao A, Cheung A, Duncan G, Girard P-M, Elia AJ, Wakeham A, et al. Chk2 is a tumor suppressor that regulates apoptosis in both an ataxia telangiectasia mutated (ATM)-dependent and an ATM-independent manner. Mol Cell Biol. 2002;22:6521–32.
Article CAS PubMed PubMed Central Google Scholar
Taylor AM, Shih J, Ha G, Gao GF, Zhang X, Berger AC, et al. Genomic and Functional Approaches to Understanding Cancer Aneuploidy. Cancer Cell. 2018;33:676-89.e3.
Article CAS PubMed PubMed Central Google Scholar
Prieto-Garcia C, Tomašković I, Shah VJ, Dikic I, Diefenbacher M. USP28: Oncogene or Tumor Suppressor? A Unifying Paradigm for Squamous Cell Carcinoma. Cells. 2021;10:2652.
Article CAS PubMed PubMed Central Google Scholar
Kuno I, Takayanagi D, Asami Y, Murakami N, Matsuda M, Shimada Y, et al. TP53 mutants and non-HPV16/18 genotypes are poor prognostic factors for concurrent chemoradiotherapy in locally advanced cervical cancer. Sci Rep. 2021;11:19261.
Article CAS PubMed PubMed Central Google Scholar
Cho S-Y, Park C, Na D, Han JY, Lee J, Park O-K, et al. High prevalence of TP53 mutations is associated with poor survival and an EMT signature in gliosarcoma patients. Exp Mol Med. 2017;49:e317.
Article CAS PubMed PubMed Central Google Scholar
Zainuddin N, Berglund M, Wanders A, Ren Z-P, Amini R-M, Lindell M, et al. TP53 mutations predict for poor survival in de novo diffuse large B-cell lymphoma of germinal center subtype. Leuk Res. 2009;33:60–6.
Article CAS PubMed Google Scholar
Cooks T, Pateras IS, Tarcic O, Solomon H, Schetter AJ, Wilder S, et al. Mutant p53 prolongs NF-κB activation and promotes chronic inflammation and inflammation-associated colorectal cancer. Cancer Cell. 2013;23:634–46.
Article CAS PubMed PubMed Central Google Scholar
Muller PAJ, Vousden KH. Mutant p53 in cancer: new functions and therapeutic opportunities. Cancer Cell. 2014;25:304–17.
Article CAS PubMed PubMed Central Google Scholar
Li T, Kon N, Jiang L, Tan M, Ludwig T, Zhao Y, et al. Tumor suppression in the absence of p53-mediated cell-cycle arrest, apoptosis, and senescence. Cell. 2012;149:1269–83.
Article CAS PubMed PubMed Central Google Scholar
Walerych D, Napoli M, Collavin L, Del Sal G. The rebel angel: mutant p53 as the driving oncogene in breast cancer. Carcinogenesis. 2012;33:2007–17.
Article CAS PubMed PubMed Central Google Scholar
Tan BS, Tiong KH, Choo HL, Chung FF-L, Hii L-W, Tan SH, et al. Mutant p53–R273H mediates cancer cell survival and anoikis resistance through AKT-dependent suppression of BCL2-modifying factor (BMF). Cell Death Dis. 2015;6:e1826.
Article CAS PubMed PubMed Central Google Scholar
Yang W, Soares J, Greninger P, Edelman EJ, Lightfoot H, Forbes S, et al. Genomics of Drug Sensitivity in Cancer (GDSC): a resource for therapeutic biomarker discovery in cancer cells. Nucleic Acids Res. 2013;41 Database issue:D955-61. https://doi.org/10.1093/nar/gks1111. (2012).
Article CAS Google Scholar
Corsello SM, Nagari RT, Spangler RD, Rossen J, Kocak M, Bryan JG, et al. Discovering the anti-cancer potential of non-oncology drugs by systematic viability profiling. Nat Cancer. 2020;1:235–48. https://doi.org/10.6084/m9.figshare.20564034.v1. (2022).
Article CAS PubMed PubMed Central Google Scholar
Salvadores M, Fuster-Tormo F, Supek F. Matching cell lines with cancer type and subtype of origin via mutational, epigenomic, and transcriptomic patterns. Sci Adv. 2020;6:eaba1862.
Article CAS PubMed PubMed Central Google Scholar
Zhang Y, Parmigiani G, Johnson WE. ComBat-seq: batch effect adjustment for RNA-seq count data. NAR Genom Bioinform. 2020;2:lqaa078.
Article PubMed PubMed Central Google Scholar
Levatić J, Salvadores M, Fuster-Tormo F, Supek F. Mutational signatures are markers of drug sensitivity of cancer cells. Nat Commun. 2022;13:2926.
Article PubMed PubMed Central Google Scholar
Zhou R, Wu J, Tang X, Wei X, Ju C, Zhang F, et al. Histone deacetylase inhibitor AR-42 inhibits breast cancer cell growth and demonstrates a synergistic effect in combination with 5-FU. Oncol Lett. 2018;16:1967–74.
PubMed PubMed Central Google Scholar
Chen Y-J, Wang W-H, Wu W-Y, Hsu C-C, Wei L-R, Wang S-F, et al. Novel histone deacetylase inhibitor AR-42 exhibits antitumor activity in pancreatic cancer cells by affecting multiple biochemical pathways. PLoS ONE. 2017;12:e0183368.
Article PubMed PubMed Central Google Scholar
Turchick A, Liu Y, Zhao W, Cohen I, Glazer PM. Synthetic lethality of a cell-penetrating anti-RAD51 antibody in PTEN-deficient melanoma and glioma cells. Oncotarget. 2019;10:1272–83.
Article PubMed PubMed Central Google Scholar
Al-Subhi N, Ali R, Abdel-Fatah T, Moseley PM, Chan SYT, Green AR, et al. Targeting ataxia telangiectasia-mutated- and Rad3-related kinase (ATR) in PTEN-deficient breast cancers for personalized therapy. Breast Cancer Res Treat. 2018;169:277–86.
Article CAS PubMed PubMed Central Google Scholar
De Ramón C, Rojas EA, Cardona-Benavides IJ, Mateos M-V, Corchete LA, Gutiérrez NC. Transcriptional signature of TP53 biallelic inactivation identifies a group of multiple myeloma patients without this genetic condition but with dismal outcome. Br J Haematol. 2022;199:344–54.
Article PubMed PubMed Central Google Scholar
Jeay S, Gaulis S, Ferretti S, Bitter H, Ito M, Valat T, et al. A distinct p53 target gene set predicts for response to the selective p53-HDM2 inhibitor NVP-CGM097. Elife. 2015;4:e06498.
Article PubMed PubMed Central Google Scholar
Haverty PM, Lin E, Tan J, Yu Y, Lam B, Lianoglou S, et al. Reproducible pharmacogenomic profiling of cancer cell line panels. Nature. 2016;533:333–7.
Article CAS PubMed Google Scholar
Cancer Cell Line Encyclopedia Consortium, Genomics of Drug Sensitivity in Cancer Consortium. Pharmacogenomic agreement between two cancer cell line data sets. Nature. 2015;528:84–7.
Article Google Scholar
Haibe-Kains B, El-Hachem N, Birkbak NJ, Jin AC, Beck AH, Aerts HJWL, et al. Inconsistency in large pharmacogenomic studies. Nature. 2013;504:389–93.
Article CAS PubMed PubMed Central Google Scholar
Gonçalves E, Segura-Cabrera A, Pacini C, Picco G, Behan FM, Jaaks P, et al. Drug mechanism-of-action discovery through the integration of pharmacological and CRISPR screens. Mol Syst Biol. 2020;16:e9405.
Article PubMed PubMed Central Google Scholar
Nguyen L, W M Martens J, Van Hoeck A, Cuppen E. Pan-cancer landscape of homologous recombination deficiency. Nat Commun. 2020;11:5584.
Article CAS PubMed PubMed Central Google Scholar
Gulhan DC, Lee JJ-K, Melloni GEM, Cortés-Ciriano I, Park PJ. Detecting the mutational signature of homologous recombination deficiency in clinical samples. Nat Genet. 2019;51:912–9.
Article CAS PubMed Google Scholar
Grossman RL, Heath AP, Ferretti V, Varmus HE, Lowy DR, Kibbe WA, et al. Toward a Shared Vision for Cancer Genomic Data. N Engl J Med. 2016;375:1109–12. GDC Data Portal https://identifiers.org/gdc.
Article PubMed PubMed Central Google Scholar
Barretina J, Caponigro G, Stransky N, Venkatesan K, Margolin AA, Kim S, et al. The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drug sensitivity. Nature. 2012;483:603–7. Depmap CCLE data https://figshare.com/articles/dataset/DepMap_21Q4_Public/16924132 2021.
Article CAS PubMed PubMed Central Google Scholar
The ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium, Aaltonen LA, Abascal F, Abeshouse A, Aburatani H, Adams DJ, et al. Pan-Cancer Analysis of Whole Genomes. Nature. 2020;578:82–93. https://doi.org/10.1038/s41586-020-1969-6.
Article CAS Google Scholar
Goldman MJ, Craft B, Hastie M, Repečka K, McDade F, Kamath A, et al. Visualizing and interpreting cancer genomics data via the Xena platform. Nat Biotechnol. 2020;38:675–8.
Article CAS PubMed PubMed Central Google Scholar
Ellrott K, Bailey MH, Saksena G, Covington KR, Kandoth C, Stewart C, et al. Scalable Open Science Approach for Mutation Calling of Tumor Exomes Using Multiple Genomic Pipelines. Cell Syst. 2018;6:271-81.e7. https://doi.org/10.1016/j.cels.2018.03.002.
Article CAS PubMed PubMed Central Google Scholar
Iorio F, Knijnenburg TA, Vis DJ, Bignell GR, Menden MP, Schubert M, et al. A Landscape of Pharmacogenomic Interactions in Cancer. Cell. 2016;166:740–54.
Article CAS PubMed PubMed Central Google Scholar
Frazer J, Notin P, Dias M, Gomez A, Min JK, Brock K, et al. Disease variant prediction with deep generative models of evolutionary data. Nature. 2021;599:91–5. EVE scores https://evemodel.org/download/bulk.
Article CAS PubMed Google Scholar
Wu Y, Liu H, Li R, Sun S, Weile J, Roth FP. Improved pathogenicity prediction for rare human missense variants. Am J Hum Genet. 2021;108:2389. VARIANT portal http://varity.varianteffect.org.
Article CAS PubMed PubMed Central Google Scholar
Friedman J, Hastie T, Tibshirani R. Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Softw. 2010;33:1–22.
Article PubMed PubMed Central Google Scholar
Ge SX, Jung D, Yao R. ShinyGO: a graphical gene-set enrichment tool for animals and plants. Bioinformatics. 2020;36:2628–9.
Article CAS PubMed Google Scholar
Supek F, Bošnjak M, Škunca N, Šmuc T. REVIGO summarizes and visualizes long lists of gene ontology terms. PLoS ONE. 2011;6:e21800.
Article CAS PubMed PubMed Central Google Scholar
Tate JG, Bamford S, Jubb HC, Sondka Z, Beare DM, Bindal N, et al. COSMIC: the Catalogue Of Somatic Mutations In Cancer. Nucleic Acids Res. 2019;47:D941–7.
Article CAS PubMed Google Scholar
Li W, Köster J, Xu H, Chen C-H, Xiao T, Liu JS, et al. Quality control, modeling, and visualization of CRISPR screens with MAGeCK-VISPR. Genome Biol. 2015;16:281.
Article CAS PubMed PubMed Central Google Scholar
Biayna J, Garcia-Cao I, Álvarez MM, Salvadores M, Espinosa-Carrasco J, McCullough M, et al. Loss of the abasic site sensor HMCES is synthetic lethal with the activity of the APOBEC3A cytosine deaminase in cancer cells. PloS Biol. 2021;19:e3001176.
Article CAS PubMed PubMed Central Google Scholar
Drainas AP, Lambuta RA, Ivanova I, Serçin Ö, Sarropoulos I, Smith ML, et al. Genome-wide Screens Implicate Loss of Cullin Ring Ligase 3 in Persistent Proliferation and Genome Instability in TP53-Deficient Cells. Cell Rep. 2020;31:107465.
Article CAS PubMed PubMed Central Google Scholar
Haapaniemi E, Botla S, Persson J, Schmierer B, Taipale J. CRISPR-Cas9 genome editing induces a p53-mediated DNA damage response. Nat Med. 2018;24:927–30.
Article CAS PubMed Google Scholar
Doench JG, Fusi N, Sullender M, Hegde M, Vaimberg EW, Donovan KF, et al. Optimized sgRNA design to maximize activity and minimize off-target effects of CRISPR-Cas9. Nat Biotechnol. 2016;34:184–91.
Article CAS PubMed PubMed Central Google Scholar
Sanjana NE, Shalem O, Zhang F. Improved vectors and genome-wide libraries for CRISPR screening. Nat Methods. 2014;11:783–4.
Article CAS PubMed PubMed Central Google Scholar

Download references

Acknowledgements

We thank Travis Stracker for discussions on roles of USP28.

Funding

Work in the F.S. lab is supported by an ERC StG “HYPER-INSIGHT” (757700), Horizon2020 project “DECIDER” (965193), Spanish government project “REPAIRSCAPE”, CaixaResearch project “POTENT-IMMUNO” (HR22-00402), an ICREA professorship to F.S., the SGR funding of the Catalan government, and the Severo Ochoa centers of excellence award of the Spanish government to the hosting institution. M.S. was funded by a FPU fellowship of the Spanish government, Ministry of Universities.

Author information

Authors and Affiliations

Institute for Research in Biomedicine (IRB Barcelona), The Barcelona Institute for Science and Technology (BIST), Barcelona, Spain
Bruno Fito-Lopez, Marina Salvadores, Miguel-Martin Alvarez & Fran Supek
Catalan Institution for Research and Advanced Studies (ICREA), Barcelona, Spain
Fran Supek

Authors

Bruno Fito-Lopez
View author publications
You can also search for this author in PubMed Google Scholar
Marina Salvadores
View author publications
You can also search for this author in PubMed Google Scholar
Miguel-Martin Alvarez
View author publications
You can also search for this author in PubMed Google Scholar
Fran Supek
View author publications
You can also search for this author in PubMed Google Scholar

Contributions

BFL, MS and FS conceived and designed the project. BFL and MS collected the data. BFL performed the analysis with the advice of MS and under the guidance of FS. BFL and FS drafted the manuscript with MS assistance. MMA performed isogenic cell line screens analysis. FS supervised the project. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Fran Supek.

Ethics declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1: Supplementary Fig. 1.

Mechanisms related to TP53 loss phenocopying in tumors. Supplementary Fig. 2. Validation of TP53 phenocopy scores. Supplementary Fig. 3. Prioritization score methodology. Supplementary Fig. 4. Further evidence supporting the role of individual TP53-loss phenocopying genes. Supplementary Fig. 5. Evidence supporting USP28 as a chromosome 11 phenocopying gene. Supplementary Fig. 6. Supporting evidence for effects of CNOT2 and MSI2 in the amplified segments bearing the MDM2 and PPM1D genes, respectively. Supplementary Fig. 7. Additional data on effects of TP53 loss and phenocopying thereof on cancer cell line drug sensitivity. Supplementary Fig. 8. Pathway-wise effects of TP53 mutations or phenocopies on drug sensitivity of cancer cell lines. Supplementary Fig. 9. Cancer type specific effect of TP53 phenocopies or mutations in drug testing. Supplementary Fig. 10. Associations between mutation status of various genes and the sensitivity to various drugs can be modulated by the TP53 functional status. Supplementary Fig. 11. Additional analyses of TP53 status modulating associations between driver mutations and drugs.

Additional file 2.

Additional file 3: Supplementary Data 1.

TCGA TP53 Phenocopy scores. Supplementary Data 2. Gene coefficients. Supplementary Data 3. PCAWG TP53 Phenocopy scores. Supplementary Data 4. Prioritization scores. Supplementary Data 5. USP28/ATM fitness effect. Supplementary Data 6. Cell lines TP53 Phenocopy score.

Additional file 4: Supplementary Data S7.

Two-way association tests for drug activity.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Cite this article

Fito-Lopez, B., Salvadores, M., Alvarez, MM. et al. Prevalence, causes and impact of TP53-loss phenocopying events in human tumors. BMC Biol 21, 92 (2023). https://doi.org/10.1186/s12915-023-01595-1

Download citation

Received: 23 November 2022
Accepted: 12 April 2023
Published: 24 April 2023
DOI: https://doi.org/10.1186/s12915-023-01595-1

Prevalence, causes and impact of TP53-loss phenocopying events in human tumors

Abstract

Background

Results

Conclusions

Background

Results

Inferring the functional TP53 status of tumors from transcriptomes

Validation and calibration of the TP53 phenocopy score model

USP28 deletion phenocopies a TP53 mutated state in tumors

Phenocopy scores prioritize causal genes within CNA segments

Cancer type spectrum of TP53 phenocopying events

Multiple neighboring genes in CNA segment can contribute to a TP53 deficient state

Detecting TP53 loss phenocopies in cancer cell line panels

Effects of TP53 on drug resistance are clarified by TP53 phenocopy scores

Associations between drug sensitivity and genetic markers are modified by TP53 status

Sensitizing effects of driver mutations to HDAC and ATR inhibitors are modulated by TP53

Discussion

Conclusions

Methods

Data collection and preparation

Gene expression and Copy Number Alteration (CNA) data

Data alignment between tumors and cell lines

TP53 status label (according to GDSC)

TP53 score classifiers in human tumors

Leave-one-cancer type-out (LOCTO) classifier

Assessment of the classifier and calculation of FDR score

TP53 status detection in cell lines

Gene codependency with TP53 knockout/knockdown in cell line screens

Calculation of the combined gene prioritization score

TP53 wild-type and TP53 -/- isogenic cell line genetic screens

Drug response associations with TP53 status

Drug response associations of gene status controlling for TP53 status

Two-way association tests for gene-drug associations modified by TP53

Availability of data and materials

Abbreviations

References

Acknowledgements

Funding

Author information

Authors and Affiliations

Contributions

Corresponding author

Ethics declarations

Ethics approval and consent to participate

Consent for publication

Competing interests

Additional information

Publisher’s Note

Supplementary Information

Additional file 1: Supplementary Fig. 1.

Additional file 2.

Additional file 3: Supplementary Data 1.

Additional file 4: Supplementary Data S7.

Rights and permissions

About this article

Cite this article

Share this article

Keywords

BMC Biology

Contact us