Profile of PTEN mutational hotspots

We used an extended cohort combining colorectal tumor specimens from FMI, cBioportal, AACR-GENIE, and other sources of publicly available data (PAD) (Fig. 1a, Supplementary Fig. S1a, b), to augment and contrast with specimens available from FMI5. In the PAD dataset, 86.9% of tumors for which MS status was known or imputed were MSS/tumor mutation burden (TMB) low (designated MT-L), 13.0% MSI/TMB high (designated MT-H), and 0.1% MSS with a high TMB (MSS-htmb, often associated with POLE mutations) versus 94.6% MT-L, 4.7% MT-H, and 0.7% MSS-htmb in the FMI dataset (Fig. 1b; see discussion of rationale for use of MT-L, MT-H, and MSS-htmb terminology in methods and in5). The frequencies of PTEN alterations we observed in the combined PAD and FMI datasets are MT-L, 3.9%; MT-H, 18.1%; MSS-htmb, 44.0%.

Fig. 1: Description of the datasets used in the study, and identification of PTEN alterations and alteration hotsponts.

a Publicly Available Dataset (PAD) for CRC used in this study, with information on numbers of CRCs acquired from AACR-GENIE, cBioportal, ICGC, and publications10,58; data were manually reviewed to avoid duplicated entry of tumors into the datasets, resulting in 18,679 non-redundant records. The FMI CRC dataset is shown for the comparison (not to scale). b Fractions of MT-L, MT-H, and MSS-htmb tumors in the PAD and FMI datasets. c Structural domains in PTEN include a phosphatidylinositol 4,5-bisphosphate (PIP2)-binding domain which regulates catalytic activity (PBD; residues 6–15, purple); a phosphatase domain active against phospholipids and phosphoproteins (residues 14–185, yellow; blue triangles mark active site residues); a C2 domain, which regulates PTEN localization (residues 190–350; light blue); and a C-terminal tail that influences protein stability and target specificity (residues 352–403; green). Location and relative abundance of PTEN hotspots in the PAD and FMI datasets. The height of each lollipop indicates the relative frequency of the corresponding mutation in the dataset, with the frequency of the most abundant observed hotspot in FMI set (R130) set as 1. For the top 6 hotspots, the absolute number of mutations in FMI/PAD datasets are shown in parentheses. Number of PTEN mutations analyzed: FMI - 2120; PAD- 1398. d PTEN 3D hotspots based on AlphaFold prediction. High-confidence (detected in all five top-ranking AlphaFold models with p-value < 0.005) novel 3D hotspots impacting the C2 domain and/or including residues at the phosphatase-C2 domain interface are shown in magenta (centered around D252) and light-orange (centered around F273). Other 3D hotspots (including those partially overlapping with the previously reported5) detected in at least one out of five best AlphaFold models with p-value < 0.005 are indicated in gray. These are predominantly located within the phosphatase domain. Phosphatase domain (left) is color-coded in yellow, and C2 domain (right) in blue. Specific affected residues are in Supplementary Table S3. e Location of PTEN residues at the dimer interface between two PTEN monomers, involving residues in the phosphatase (yellow) and C2 (blue) domains. See Supplementary Table S4 for specific affected residues.

Analysis of the pattern of recurrently mutated amino acids in PTEN (e.g. hotspots) indicated similar profiles in the FMI and PAD datasets (Fig. 1c, Supplementary Table S2). To determine if specific sites of mutation are clustered on the 3D folded structure of the protein, we modeled PTEN with Alphafold2 (AF2) and ESMfold (Supplementary Fig. S1c), which allowed us to extend the coverage beyond that present in the experimental models. Using the 5 best models (Supplementary Fig. S1d) we identified statistically significant 3D hotspots (Supplementary Table S3), defined as overabundance of mutations occurring in a group of amino acids no more than 5\(\mathring\) apart in the PTEN monomer. This identified two new 3D hotspots partially encompassing the C2 domain of PTEN (Fig. 1d), in addition to 3D hotspots residing within the phosphatase domain. In addition, since PTEN is known to function as a multimer, we investigated whether mutations disrupt the PTEN dimeric interface. Interestingly, the interface between the PTEN molecules represents a “coldspot”, with mutations occurring there at ~3.8 times lower frequencies than elsewhere, in both MSS and MSI tumors (p-value = 0.003, Fig. 1e, see Supplementary Table S4 for specific affected residues).

Overall selection for non-synonymous PTEN mutations

For most oncogenes and tumor suppressors with cancer driver activity, function is context dependent, with few universally important across a range of tissues and organs10,11. The driver status of PTEN mutations in CRC is ambiguous, with one study finding mutational inactivation of PTEN under selection in many tumor types, but not in CRC12. One approach to assess functional selection pressures in cancer genomes is to calculate the ratio of non-synonymous to synonymous mutations (dN/dS) in a tumor type-specific manner. Using this approach, PTEN mutations were identified as tumor drivers in 16/29 tumor types, including CRC11; however, this conclusion was drawn from the analysis of only 226 colorectal samples based on a complex statistical model. Using the extensive dataset in this study, we directly calculated dN/dS for PTEN benchmarked to the oncogenes KRAS and PIK3CA, and the tumor suppressors APC and TP53, considering non-truncating and truncating mutations separately (Fig. 2a). For KRAS and PI3K, there is a strong selection for non-truncating mutations and a negative selection against truncating mutations, reflecting the known requirement for specific missense mutations to activate these oncogenes. There is a very strong selection for APC truncating mutations, although non-truncating mutations are also positively selected. For both PTEN and TP53 there is similarly a very strong selection for both truncating and missense mutations; although the dN/dS is smaller for PTEN (suggesting a slightly higher burden of passenger mutations), nevertheless about 94% of non-truncating and 95% of truncating PTEN mutations are expected to be genuine driver mutations. To gain additional insight, we analyzed additional features of PTEN mutational landscape, including the hotspot profile, allele frequencies, and co-occurrence profiles, and impact on survival.

Fig. 2: Relationship of selection pressures and PTEN mutation patterns.

a Ratio of non-synonymous to synonymous (dN/dS) for non-truncating (blue) or truncating (green) mutations observed in tumors bearing mutations in KRAS, APC, PIK3CA, TP53, or PTEN. Error bars represent 95% confidence intervals. b Spectra of transitions and transversions for PTEN observed in the FMI and PAD subsets, compared to the frequency of these changes observed in all genes profiled by FMI in CRC, or previously published67 data on CRC specimens as baselines. c Spectra of PTEN transitions and transversions observed in the MT-L, MT-H and MSS-htmb subsets, based on specimens from the FMI dataset. d Frequency of observed (green) PTEN nucleotide changes in MT-L FMI subset versus values predicted based on nucleotide content (blue). The difference in the overall pattern of substitutions is statistically significant, p-value < 2.2e−16 (using a chi-square test). e Comparison of relative frequencies of top 5 predicted and top 5 observed single nucleotide substitution hotspots in the FMI MT-L subset. Frequency of the most abundant observed hotspot (R130) and the most abundant predicted hotspot (R308) was set as 1. Of the top 5 predicted hotspots, three (those in codons 130, 173, and 233) are also among the top 5 observed hotspots. Mutations in codons 234 and 308 are predicted to occur with high frequencies but have low counts in observed set. Observed hotspots in codons 136 and 68 occur at the locations predicted by the mutational spectra (albeit with much lower frequencies than those in codons 130, 173, and 233), but are not among the top 5 predicted hotspots. Supplementary Table S7 contains predicted and observed frequencies for all 403 codons of PTEN.

Nucleotide substitution profile in PTEN in CRC: mutational signatures versus functional selection

To investigate the underlying selection factors for PTEN hotspots, we first cross-validated the changes observed in the PAD and FMI datasets, comparing the specific nucleotide substitutions observed in PTEN, versus cumulative substitutions observed in all genes analyzed in CRC specimens in FMI or public data (Fig. 2b). This showed overall congruence between the two PTEN datasets. The observed differences, while statistically significant due to the very high sample count, likely reflect the distinct percentage of MT-L, MT-H, and MSS-htmb in the FMI versus PAD datasets, as these CRC subtypes have highly divergent nucleotide substitution profiles (Fig. 2c). We next used the frequencies of each possible base substitution in each of 96 different trinucleotide contexts (ref. 13, Supplementary Table S5) to predict the nucleotide substitution spectrum and frequency across the PTEN coding sequence in MT-L and MT-H tumors. The transitions/transversions profile actually observed in PTEN MT-L cohort versus predicted based on the nucleotide content of the gene was significantly different (Fig. 2d, Supplementary Table S6), suggesting the presence of additional mutability variables. We also analyzed in detail the changes from the most abundant classes of PTEN mutations produced by mutational signatures, common in CRC14, (the SBS1 and ID1/ID2/ID5/ID7 signatures (collectively designated hereafter as IDT, for total); Supplementary Fig. S2a–c). This indicated that deletions in polynucleotide runs are much more frequent than insertions in MT-H cohort, as is typical for cells with deficiency in mismatch repair (MMR) proteins15, while the opposite was true for the MT-L subset.

Focusing on non-synonymous single nucleotide substitutions in PTEN for the combined cohort (excluding MSS-htmb samples, comprising ~0.7% of the total, as too small a cohort to yield significant conclusions), we compared the observed frequency of mutations, versus the predicted frequency based on the occurrence of mutational signatures common in CRC (Supplementary Table S6). As illustrated in Fig. 2e, this baseline prediction model correctly identified many of the most frequent hotspots. Mismatches between predicted and observed hotspot frequencies are likely due to selection pressure at the protein level, and/or reflecting the influence of additional local factors such as nucleosome positioning or protein binding sites, all of which have been reported to influence background mutability16.

To discriminate functional selection for the effect of specific mutations on protein function from the mesoscale features of the nucleotide sequence, we therefore also compared the predicted versus observed frequency of synonymous mutations, as these are not expected to affect protein function (Supplementary Fig. S2, Supplementary Table S7) while sharing the same coarse features (such as replication timing and chromatin state) with non-synonymous mutations. This analysis identified statistically significant differences in patterns of occurrence from predicted frequency for synonymous mutations (p-value 0.0015, Supplementary Fig. S2), suggesting the presence of additional DNA-based mutational processes driving such deviations. However, the deviation from the predicted pattern of non-synonymous mutations (Supplementary Table S7) could not be explained by the same non-selective processes identified in the analysis of the synonymous mutations, as only ~1% of models generated distributions that have metrics similar to that of the data, implying additional selective pressures at the protein level, and/or localized effects involving specific potentially mutable codons.

Loss of LPA activity and protein abundance contribute to selection for PTEN mutational hotspots

Recently published estimates of the damaging potential of almost all possible PTEN mutations17,18,19, allowed us to characterize most of the observed single amino acid substitutions in PTEN in terms of protein abundance and/or lipid phosphatase activity (LPA) (Supplementary Fig S3a, Supplementary Table S8). Using these data, as well as computational tools to predict the impact of in frame deletions on PTEN protein function20 and annotations for specific PTEN mutations in databases linking genomic variation with phenotypes, we were able to estimate the damaging potential for >90% of non-synonymous mutations in PTEN (non-hypermutated subset), dichotomizing them as functional wildtype, versus LoF (Fig. 3a, Supplementary Fig S3b, Supplementary Table S8). This analysis indicated that hotspot mutations were significantly more likely to be annotated as LoF than non-hotspot mutations, in both the MT-L and MT-H subsets (p-value < 2.2e−16). In general, mutations with greater loss of LPA or abundance are observed at relative frequencies much higher than are those with less effect on protein function (Fig. 3b, Supplementary Fig. S3c), again implying the presence of positive selection for function-damaging PTEN mutations.

Fig. 3: Functionality of mutationally altered PTEN at least partially defines the observed mutation spectrum.

a For missense mutations in the MT-L dataset, analysis using predicted abundance and lipid phosphatase activity (LPA) based on18. Prediction of the impact of inframe deletions on protein function were done using MutPred2. After further stratification by annotation in the clinical databases (CKB, OncoKB, Clinvar; see Methods for details), hotspot versus non-hotspot mutations were assigned as functionally wild type (WT; green shading) or loss of function (LoF; pink shading); NA (gray), not available, signified insufficient data for prediction. Image depicts count of mutations in each functionally assigned subset of mutations (LoF, WT, NA) for both hotspots (left side of the triangle) vs non-hotspots (right side of the triangle). LoF mutations are more frequent among the hotspots (p-value < 2.2e−16). b Comparison of predicted (red) vs observed (blue) LPA profile for PTEN missense mutations in the MT-L subset. For each predicted or observed mutation, the corresponding LPA values were retrieved from ref. 18, see Supplementary Table S8. c Difference between predicted and observed mutational profiles using chi-square statistics. Rand, random baseline; sign, prediction made on the basis of mutational signatures alone; LPA, with correction for LPA score alone; abund, with correction for PTEN protein abundance score alone; LPA + ab, with correction for both LPA score and abundance; LoF, with correction for overall loss of activity, i.e., also taking into account annotations in clinical databases, where available. d Comparison of relative frequencies of top 5 predicted (based on mutational signature) and top 5 observed hotspots in the FMI MT-L subset, with the correction for LPA score. The height of the lollipop stems for the hotspots was calculated in relation to the frequency of the most abundant hotspot (R130 for the observed data, R308 for the predicted data), which were set as 1. For each codon, the total predicted value includes all possible mutations (in proportion determined by relative probabilities for each substitution, Supplementary Table S6); these mutations may or may not result in the loss of function. The overall height of the stem is color-coded to indicate the fraction of the total which would result in the loss of LPA (black), vs no loss of LPA (light-gray). Of the top 5 predicted hotspots, 3 (in codons 130, 173, and 233) are also among the top 5 observed hotspots. Different patterns of predicted versus observed characterize distinct hotspots, correlated with impact on protein function. For example, only about ~2/3 of mutations leading to amino acid changes in codon 233 are predicted to cause loss of LPA; here, the actual number of mutations is closer to the predicted count of mutations with low LPA scores. Mutations in codons 234 and 308 are predicted to occur with high frequencies based on mutational signature, but none of these is predicted to affect LPA; these positions are infrequently mutated in the analyzed cohort. Observed hotspots in codons 133 and 319 occur at the frequencies higher than that predicted by the mutational spectra; notably, these overlap with a consensus motif increasing the mutability of CG dinucleotides (ref. 68, left inset), or an indel hotspot consensus motif (ref. 69, right inset). Supplementary Table S9 contains predicted and observed frequencies for all 403 positions of PTEN, as well as the fraction of functionally damaging mutations according to LPA scores, abundance measures, and the combinations thereof.

We therefore explored if adjusting for the functional status of PTEN mutations (applied as a coefficient calculated as fraction of function-damaging mutations in any given codon) would improve the ability to predict the hotspot profile. Correction for the LPA activity alone provided the best fit with the observed data, leading to a modest improvement in comparison with measures of protein abundance or overall LoF status; and resulted in the overall good congruency between predicted and observed data for single nucleotide substitution data (Fig. 3c, Supplementary Table S9). At least some of the remaining differences are likely due to fine scale mutational features, such as overlaps with mutation consensus sites, or sites with high DNA curvature (Fig. 3d insets, Supplementary Fig. S3d). Thus, as most critical selection processes, the profile of PTEN hotspots is first defined by mutational processes in CRC as defined by mutational signatures, and then further refined by secondary sequence features and positive selection for the loss of LPA.

Evolutionary conservation and allele frequency of PTEN hotspot mutations suggests importance of non-canonical PTEN functions

Mutational hotspots overall are more frequent in evolutionarily conserved positions in protein-coding sequences, based on analysis of the TCGA Pancancer dataset21. We also analyzed the location of defined mutational hotspots in the context of the overall evolutionary conservation of both nucleotides and amino acid residues in PTEN (Fig. 4a–c). Across all CRC tumors, hotspot mutations were more likely than non-hotspots to occur in highly conserved amino acids of PTEN (mostly located in the phosphatase domain, Fig. 1c, Supplementary Fig. S3a). However, at the DNA level, hotspots did not occur in more highly conserved nucleotide positions (Supplementary Fig. S3e). These analyses suggested positive selective pressure for PTEN hotspot mutations at the level of protein function.

Fig. 4: Evolutionary conservation, but not allele frequencies, differ between hotspot and non-hotspot mutations of PTEN.

Structure of PTEN monomer (oriented with phosphatase domain to the top) (a) and primary amino acid sequence (b) color coded to indicate degree of conservation of residues. c Degree of conservation in mutations (missense and inframe indels) targeted by hotspot versus non-hotspot mutations indicates hotspot mutations target more highly conserved residues. Y-axis, Consurf score70 with the lower score indicating higher evolutionary constraint ***, p < 1.0E−12. d Allele Fraction (AF) of PTEN mutations, versus mutations in CRC driver genes, in MT-L CRC. e Allele frequency (AF) of PTEN mutations predicted as LoF (purple) or WT (blue) normalized to frequency of mutations in APC, KRAS, PI3KCA, or TP53 in MT-L CRC. No statistically significant differences (all p-values > 0.05) were found between WT and LOF profiles, using a KS test. f Allele frequency of mutations in indicated hotspot residues, or cumulatively for mutations defined as “other hotspots” (htsp), LoF, non-synonymous (non), synonymous (syn), or WT. ***, p < 0.005. Lower AF of hotspot in codon 319 (which arises as a result of loss of one half of the repeat ACTTACTT and overlaps a consensus indel site, see Fig. 3d and Supplementary Fig. S2a) indicates it is a relatively late event in tumor progression.

Driver mutations that alter protein activity typically occur earlier in tumor evolution, and thus have higher allele fraction (AF), in contrast to passenger mutations, which are expected to appear throughout cancer evolution and are frequently subclonal. For context, we compared the AF of PTEN mutations estimated to have LoF versus WT characteristics in tumors bearing APC, KRAS, PIK3CA, or TP53 mutations (Fig. 4d) in the MT-L subset, comprising the bulk of CRC samples. Considering the complete set of PTEN mutations, the AF of PTEN mutations is lower than that of APC or TP53 mutations, in agreement with the current view that alterations in PTEN are a later event in CRC carcinogenesis22; however, AFs are similar to those of KRAS, and higher than those of PIK3CA, suggesting selection. However, in the MT-L subset there was no significant difference in AF based on whether or not mutations impaired PTEN function (Fig. 4e). Further, comparison of the individual most abundant hotspots versus the set of all other hotspots, or mutations defined as LoF versus wt, or synonymous (Fig. 4f, Supplementary Table S10) in each case demonstrated a similar allele frequency in MT-L tumors. Comparison of AF of hotspots versus non-hotspots (Supplementary Fig. S3f, g) also demonstrated no significant differences. In contrast to earlier analysis (Fig. 3a–d), these data argued that any positive selection for function-damaging PTEN hotspot mutations is not reflected in their earlier appearance in the tumor evolutionary history.

Finally, we considered that mutations classified as wt-like based on LPA or protein stability may reflect disruption of alternative functions of PTEN, such as protein phosphatase activity, or interaction with substrate. Detailed analysis of the spatial distribution of these “wt” mutations of PTEN did not reveal any propensity to congregate in a particular PTEN domain, or in specific proximity to the sites of post-translational modifications (Supplementary Fig. S3h). Protein interactions have been reported for multiple domains of PTEN (Supplementary Fig. S3h); however, there is currently no available detail about whether the specific wt hotspot mutations would be critical mediators of such interactions.

Distinct co-segregation of PTEN mutations in MT-L and MT-H tumors

In complementary analysis, we asked if specific classes of PTEN alteration selectively co-occurred with common driver mutations that themselves were associated with significant survival effects (Fig. 5a, b). APC, KRAS, TP53, and BRAF mutations have markedly different distribution patterns in MT-L and MT-H cohorts, reflecting the distinct underlying genetic causes of these tumors, and distribution of these mutations has been used to develop a 10-group classification system associated with distinct clinical outcomes10. The frequency of PTEN alterations varied greatly between these groups in MT-L, but not MT-H tumors. In MT-L tumors, most PTEN alterations are observed in the highly prevalent groups of tumors characterized by mutations in APC + KRAS, APC + TP53, or APC + KRAS + TP53. However, the highest frequency of PTEN alterations is associated with tumors bearing BRAF mutations (BRAF & BRAF + TP53 groups), while the lowest frequency is observed in tumors bearing APC and TP53 alterations in the absence of a KRAS or BRAF driver mutation. These patterns suggest biological selection for PTEN mutations in the context of specific patterns of driver mutations in the MT-L tumors. Intriguingly, more granular analysis of the relationship between PTEN and other driver mutations in MT-L tumors (Fig. 5c) indicates that the mutual exclusion between PTEN and TP53 is driven entirely by the interaction of TP53 mutations and PTEN LoF mutations; in contrast, TP53 deletions tend to co-occur with PTEN deletions and other alterations. In contrast, with the exception of under-representation in tumors with KRAS + TP53 alterations, similar frequencies of PTEN mutations are observed regardless of the profile of driver mutations in MT-H tumors.

Fig. 5: Correlation of PTEN alterations with alterations in other CRC driver genes, and with survival.

PTEN alteration frequencies in the MT-L (a) and MT-H (b) cohorts, subdivided as in ref. 10 into functionally distinct groups according to the presence of mutations in APC (A), TP53 (P), BRAF (B), KRAS (K), in combinations as indicated, or none (N) of these mutations. On the vertical axis, the height of each bar represents the fraction of MT-L tumors containing mutations in the indicated gene. On horizontal axis, fraction of non-synonymous PTEN alterations of any type. c Co-occurrence of PTEN LoF or WT-like mutations with alterations in APC, TP53, PIK3CA, SMAD4, KRAS, based on the merged FMI-PAD dataset for MT-L CRC. Error bars represent 95% confidence intervals. d Pattern of homozygous PTEN deletions in TCGA cohorts, summarized from all Pancancer studies. Schematic of the chromosome 10, coordinates 87650K-88600K, is shown. Top table, counts of all homozygous deletions in TCGA Pancancer set, encompassing PTEN and the genes indicated. Lines indicate the deletions encompassing PTEN specifically in the TCGA CRC cohort. e Analysis of survival in the complete AACR-GENIE Biopharma Collaborative Releases (BPC) CRC cohort, based on PTEN status. Red, homozygous PTEN deletions (n = 23); green, PTEN missense or indel mutations (n = 92); blue, PTEN wt (n = 1323). Shaded areas represent 95% confidence intervals. f Survival analysis for MT-L subset, homozygous PTEN deletion (red) vs PTEN wt (blue). Shaded areas represent 95% confidence intervals. g Survival analysis for MT-L subset, PTEN mutated vs PTEN wt. Shaded areas represent 95% confidence intervals. h Survival analysis for MT-H CRCs, PTEN mutated vs PTEN wt. Shaded areas represent 95% confidence intervals. Note a threshold of 0.005 is used in this study as a threshold for statistical significance (see Methods). i Frequencies of PTEN mutations by microsatellite stability cohorts and cancer stages (combined stages I & II, vs combined stages III & IV). Error bars represent 95% confidence intervals. See Supplementary Table S12 for details.

PTEN is a negative regulator of PI3K, which is composed of a catalytic subunit, including PIK3CA, and negative regulatory subunits, including PIK3R1. In CRC, mutations activating PIK3CA or inactivating PIK3R1 occur at a low but appreciable frequency; these mutations significantly co-occurred with both LoF and WT mutations in PTEN (Fig. 5c). Similarly, BRAF and KRAS mutations significantly co-occurred with all classes of PTEN mutations. Additional analysis of the FMI and PAD subsets (Supplementary Fig. S4a, b), as well as analysis of the FMI subset for co-occurrence of mutations in key CRC driver genes, with either individual PTEN mutational hotspots, or cumulatively for belonging to other hotspots, or non-hotspots, also revealed no significant differences in co-occurrence frequency in MT-L (Supplementary Fig. S4c, d) or MT-H (Supplementary Fig. S4e) tumors. Considering co-occurrence separately for the different classes of LOF (e.g., by abundance or loss of LPA), in some cases there are differences (e.g., BRAF). This may reflect the importance of protein-protein interactions in which PTEN plays a non-catalytic function, which may be retained for PTEN mutant forms with reduced LPA, but reduced along with protein abundance.

PTEN mutation versus deletion; relation to prognosis

Although most analyses of PTEN focus on mutation or epigenetic loss of expression of the gene, in some tumors, PTEN is only lost by deletion. Deletions of the PTEN gene were observed in the CRC data set, mostly in MT-L tumors, making up about 40% of all alterations (1228/3311 specimens with PTEN alterations in the combined FMI + PAD dataset). PTEN deletions most commonly co-occurred with the loss of the neighboring genes (Fig. 5d), as well as with the mutation of BRAF, KRAS, and/or PIK3R, and deletions in TP53, and were mutually exclusive with TP53 mutations and PIK3CA mutations (Supplementary Table S11).

We evaluated the impact of PTEN mutation versus loss on overall survival of CRC patients (Fig. 5e). An initial analysis of pooled CRC specimens suggested opposing effects of non-synonymous PTEN mutations and PTEN deletions, with PTEN deletions associated with shorter survival (median overall survival (OS) 31.3 versus 63.2 months), but PTEN mutations with better survival (median OS not reached but >150 months). Independent analysis of the MT-L (Fig. 5f, g) and MT-H (Fig. 5h) cohorts confirmed that PTEN deletions were associated with worse outcome in MT-L patients (median OS 31.3 versus 60.0 months) (Fig. 5f). In contrast, non-synonymous mutations had no effect on survival in MT-L tumors (Fig. 5g), even though ~90% of them caused loss of PTEN LPA function. Intriguingly, a positive survival benefit of PTEN mutations was confined to MT-H tumors; however, this result did not reach statistical significance, likely due to inadequate sample size (Fig. 5h). This was unlikely to reflect specific cosegregation of PTEN mutations with specific driver mutations, which was not observed in MT-H tumors (Fig. 5b). TNM status could potentially influence outcome, if PTEN mutations were selectively associated with late-stage tumors. While stage information is not available for FMI specimens, we analyzed PAD data for 3221 MT-L samples and 796 MT-H samples (Fig. 5i, Supplementary Table S12). This indicated frequency of PTEN mutations was comparable in early (stage I and II) versus late (stage III and IV) tumors, excluding this interpretation.

In contrast, specific association of PTEN deletions with poor outcome in MT-L tumors may partially reflect the effect of alterations in other driver genes co-segregating with these deletions (Fig. 5a; e.g., mutation of either BRAF, KRAS, or PIK3R1, or deletion of TP53 results in shorter survival (48 vs 68 months)). However, it may also reflect the loss of genes linked to PTEN in addition to or instead of PTEN itself. Analysis of the breakpoints of PTEN-encompassing deletions in CRC and other tumor types (Fig. 5d) indicated that genes commonly co-deleted with PTEN include KLLN, ATAD1, PAPSS2, RNLS, and LIPJ. Separate analysis of deletion patterns of these genes shows they are never in the absence of a co-occurring deletion of PTEN, and that mutations inactivating these genes in CRC are extremely uncommon, affecting <1% of tumors. However, it is possible that loss of one or more of these genes in combination with PTEN loss may worsen disease presentation; for example, loss of expression of PAPSS2 has been linked to colitis and colon cancer23, and downregulation of KLLN exacerbates the presentation of Cowden syndrome24.