Functional characterization of all CDKN2A missense variants and comparison to in silico models of pathogenicity

eLife Assessment

This is a saturation mutagenesis screening of CDKN2A gene, successfully assessing the functionality of the missense variants. The work is solid and well-prosecuted. The manuscript was improved during the revision process and this work will serve as a valuable resource for diagnostic labs as well as cancer geneticists.

https://doi.org/10.7554/eLife.95347.4.sa0

Significance of the findings:

Valuable: Findings that have theoretical or practical implications for a subfield

Landmark
Fundamental
Important
Valuable
Useful

Strength of evidence:

Solid: Methods, data and analyses broadly support the claims with only minor weaknesses

Exceptional
Compelling
Convincing
Solid
Incomplete
Inadequate

During the peer-review process the editor and reviewers write an eLife Assessment that summarises the significance of the findings reported in the article (on a scale ranging from landmark to useful) and the strength of the evidence (on a scale ranging from exceptional to inadequate). Learn more about eLife Assessments

Abstract
Introduction
Results
Discussion
Methods
Data availability
References
Article and author information
Metrics

Abstract

Interpretation of variants identified during genetic testing is a significant clinical challenge. In this study, we developed a high-throughput CDKN2A functional assay and characterized all possible human CDKN2A missense variants. We found that 17.7% of all missense variants were functionally deleterious. We also used our functional classifications to assess the performance of in silico models that predict the effect of variants, including recently reported models based on machine learning. Notably, we found that all in silico models performed similarly when compared to our functional classifications with accuracies of 39.5–85.4%. Furthermore, while we found that functionally deleterious variants were enriched within ankyrin repeats, we did not identify any residues where all missense variants were functionally deleterious. Our functional classifications are a resource to aid the interpretation of CDKN2A variants and have important implications for the application of variant interpretation guidelines, particularly the use of in silico models for clinical variant interpretation.

Introduction

Genetic testing of patients with cancer to identify variants associated with an increased cancer risk and sensitivity to targeted therapies is becoming more common as broad testing criteria are integrated into clinical care guidelines (Goggins et al., 2020; Stoffel et al., 2019). The American College of Medical Genetics (ACMG) provides a framework to integrate multiple types of evidence, including variant characteristics, disease epidemiology, clinical information, and functional classifications, to interpret variants in any gene (Richards et al., 2015). In silico variant effect predictors are also integrated into ACMG variant interpretation guidelines as supporting evidence to aid classification of variants. While numerous models have been developed, varied accuracy, poor agreement between models, and inflated performance on publicly available data have been reported (Cubuk et al., 2021; Jaffe et al., 2011; Wilcox et al., 2022). Recently developed variant effect predictors aim to overcome these limitations by incorporating deep-learning-based protein structure predictions and by not training on human annotated datasets (Brandes et al., 2023; Cheng et al., 2023; Gao et al., 2023). However, post-development assessment of machine learning-based variant effect predictors, to determine accuracy on novel experimental datasets and suitability for clinical use, are limited.

Variants that cannot be classified as either pathogenic or benign are categorized as variants of uncertain significance (VUSs). However, while pathogenic and benign variants identified during genetic testing are clinically actionable, VUSs are the cause of deep uncertainty for patients and their health care providers as an unknown fraction are functionally deleterious and, therefore, likely pathogenic. For example, individuals with germline VUSs in a pancreatic cancer susceptibility gene are not be eligible for clinical surveillance programs that are associated with improved patient outcomes, unless they otherwise meet family history criteria (Goggins et al., 2020; Stoffel et al., 2019). Similarly, patients with breast or pancreatic cancer and a germline BRCA2 VUS would not be eligible for treatment with olaparib, a poly (ADP-ribose) polymerase inhibitor (Golan et al., 2019; Tutt et al., 2021). Reclassification of VUSs into pathogenic or benign strata has real-world, life-or-death consequences that necessitate a high degree of accuracy.

Germline VUSs in hereditary cancer genes are a common finding in patients with cancer and frequently can be reclassified as pathogenic on the basis of in vitro functional evidence (Kimura et al., 2022). In patients with pancreatic ductal adenocarcinoma (PDAC), germline CDKN2A VUSs affecting p16^INK4a, most often rare missense variants, are found in up to 4.3% of patients (Chaffee et al., 2018; Kimura et al., 2021; McWilliams et al., 2018; Roberts et al., 2016; Shindo et al., 2017; Zhen et al., 2015). As functional data from well-validated in vitro assays are incorporated into ACMG variant interpretation guidelines, we recently determined the functional consequence of 29 CDKN2A VUSs identified in patients with PDAC using an in vitro cell proliferation assay (Kimura et al., 2022; Richards et al., 2015). We found that over 40% of VUSs assayed were functionally deleterious and could reclassified as likely pathogenic.

Functional characterization, however, is time-consuming, expensive, and requires technical and scientific expertise. These limitations hinder assessment of in silico variant effect predictors and patient access to functional data that may allow reclassification of VUSs into clinically actionable strata. As CDKN2A VUSs will continue to be identified in patients with cancer undergoing genetic testing, we developed a high-throughput functional assay to provide a broad interpretation framework for CDKN2A variants. We characterized all possible CDKN2A missense variants and compared our functional classifications to recently developed in silico models based on machine learning to determine the accuracy of variant effect predictions.

Results

Functional characterization of CDKN2A missense variants

We utilized a codon-optimized CDKN2A sequence for our multiplexed functional assay. Expression of codon-optimized CDKN2A or the synonymous CDKN2A variants, p.L32L, p.G101G, and p.V126V, in PANC-1, a PDAC cell line with a homozygous deletion of CDKN2A, resulted in significant reduction in cell proliferation (p-value<0.0001; Figure 1—figure supplement 1A). There was no significant difference between codon-optimized CDKN2A and the three synonymous variants assayed. Conversely, expression of three pathogenic variants, p.L32P, p.G101W, and p.V126D, in PANC-1 cells did not result in any significant changes in cell proliferation. To determine if there were unappreciated selective effects during in vitro culture, we generated a CellTag library based on the pLJM1 plasmid that contained 20 nonfunctional 9 base pair barcodes of equal representation. We then transduced PANC-1 cells that stably expressed codon-optimized CDKN2A with the CellTag library (day 0) and determined representation of each barcode in the cell pool on day 9 and at confluency (day 45). We found no statistically significant changes in barcode representation, indicating that representation of a pool of functionally neutral variants is stable over a period of in vitro culture representing our assay time course (Figure 1—figure supplement 1B, Supplementary file 1).

We next determined whether we could identify functionally deleterious CDKN2A variants at a single residue when all amino acid variants were assayed simultaneously. We generated lentiviral expression plasmid libraries for all 156 CDKN2A amino acid residues, where each library contained all possible amino acids at a single residue. Twenty-seven variants (27 of 3120, 0.87%) were represented in the plasmid libraries at ≤1%. Expression plasmids for each of these 27 variants were individually generated by site-directed mutagenesis and added to the corresponding plasmid library to a calculated representation of 5% (Figure 1—figure supplement 2A and B, Supplementary file 2). Plasmid libraries were then individually amplified, and lentivirus produced. To confirm that the representation of each variant was maintained after transduction, we transduced three lentiviral libraries (amino acid residues p.R24, p.H66, and p.A127) individually into PANC-1 cells and determined the proportion of each variant in the amplified plasmid library and in the cell pool at day 9 post-transduction. The proportion of each variant in the amplified plasmid library and in the cell pool at day 9 were highly correlated (Figure 1—figure supplement 2C and D, Supplementary file 3).

For two CDKN2A amino acid residues that include pathogenic and benign variants, p.V126 and p.R144, we determined the representation of each variant in the transduced cell pool at day 9 and at confluency after a period of in vitro culture, day 23 and day 31 post-transfection, respectively (Figure 1A and B, Supplementary file 4, Supplementary file 5). Two synonymous variants, p.V126V and p.R144R, as well as a previously reported benign variant, p.R144C, either decreased or maintained their representation in the cell pool during in vitro culture as determined by the number of sequence reads supporting the variant. Representation of a previously reported pathogenic variant, p.V126D, increased in the cell pool. Notably, several other variants including p.V126R, p.V126W, p.V126K, and p.V126Y, also increased in representation in the cell pool, suggesting that additional amino acid changes at this residue are functionally deleterious (Figure 1A).

Figure 1 with 2 supplements see all

Download asset Open asset

Pooled analysis of *CDKN2A* variants at two residues with previously reported pathogenic and benign variants.

PANC-1 cell stably expressing 1 of 20 *CDKN2A* variants, 19 missense variants, and 1 synonymous variant, at residue p.V126 or p.R144 were cultured. Variant representation, as the percent of reads supporting the variant sequence, before and after a period in vitro cell proliferation determined by next-generation sequencing for the two residues, p.V126 (A) or p.R144 (B). *CDKN2A* variant p.V126D (*) was previously reported as pathogenic and increased representation during in vitro proliferation. CDKN2A variant p.R144C (**) was previously reported as benign variant and maintained representation during in vitro proliferation.

Figure 1—source data 1 Raw data in Figure 1.: https://cdn.elifesciences.org/articles/95347/elife-95347-fig1-data1-v1.xlsx
Download elife-95347-fig1-data1-v1.xlsx

To functionally characterize 2964 CDKN2A missense variants, PANC-1 cells were transduced with each of the 156 lentiviral expression libraries individually and representation of each CDKN2A variant in the resulting cell pool determined at day 9 after transduction and at confluency (days 16–40) (Supplementary file 5). Variant read counts were then analyzed using a gamma generalized linear model (GLM), that does not rely on annotation of pathogenic and benign variants to set classification thresholds, and variants with statistically significant p-values were classified as functionally deleterious (log₂ p-values≤–53.2). Variants with p-values that did not reach statistical significance were classified as either of indeterminate function (log₂ p-values>–53.2 and <–5.8) or functionally neutral (log₂ p-values≥–5.8).

We found that 525 of 2964 missense variants (17.7%) were functionally deleterious in our assay (Figure 2A, Figure 2—figure supplement 1A, Supplementary file 4). In addition, 1784 variants (60.2%) were classified as functionally neutral, with the remaining 655 variants (22.1%) classified as indeterminate function (Figure 2A, Supplementary file 4). In general, our results were consistent with previously reported classifications. Of variants identified in patients with cancer and previously reported to be functionally deleterious in published literature and/or reported in ClinVar as pathogenic or likely pathogenic (benchmark pathogenic variants), 27 of 32 (84.4%) were functionally deleterious in our assay (Figure 2B, Figure 2—figure supplement 1B and C, Supplementary file 4; Chaffee et al., 2018; Chang et al., 2016; Horn et al., 2021; Hu et al., 2018; Kimura et al., 2022; McWilliams et al., 2018; Roberts et al., 2016; Zhen et al., 2015). Five benchmark pathogenic variants were characterized as indeterminate function, with log₂ p-values from –19.3 to –33.2. Of 156 synonymous variants and six missense variants previously reported to be functionally neutral in published literature and/or reported in ClinVar as benign or likely benign (benchmark benign variants), all were characterized as functionally neutral in our assay (Figure 2B, Figure 2—figure supplement 1B and C, Supplementary file 4; Kimura et al., 2022; McWilliams et al., 2018; Roberts et al., 2016). Of 31 VUSs previously reported to be functionally deleterious, 28 (90.3%) were functionally deleterious and 3 (9.7%) were of indeterminate function in our assay. Similarly, of 18 VUSs previously reported to be functionally neutral, 16 (88.9%) were functionally neutral and 2 (11.1%) were of indeterminate function in our assay (Figure 2B, Figure 2—figure supplement 1B and C, Supplementary file 4).

Figure 2 with 5 supplements see all

Download asset Open asset

Functional characterization of all possible *CDKN2A* missense variants.

(A) Functional classifications for 3120 *CDKN2A* variants, including 2964 missense variants and 156 synonymous variants. Variants were classified as functionally deleterious, indeterminate function, or neutral based on p-value using gamma generalized linear model (GLM). 525 (17.7%) variants were classified as functionally deleterious. (B) Log₂ p-value (gamma GLM) for 32 benchmark pathogenic variants, 6 benign variants, 31 variants of uncertain significance (VUSs) previously reported to have functionally deleterious effects, and 18 VUSs previously reported to have functionally neutral effects. (C) Heatmap with p-values (gamma GLM) for all 3120 *CDKN2A* variants assayed.

Figure 2—source data 1 Raw data in Figure 2B.: https://cdn.elifesciences.org/articles/95347/elife-95347-fig2-data1-v1.xlsx
Download elife-95347-fig2-data1-v1.xlsx

We next compared variant classifications using the gamma GLM to variant classifications using a normalized fold change method (Brenan et al., 2016; Giacomelli et al., 2018). Classification of missense variants using normalized fold change also differentiated benchmark pathogenic and benchmark benign variants (Figure 2—figure supplement 2A and B, Supplementary file 6). Using benchmark pathogenic variants and benchmark benign variants to set thresholds for classification, we classified all variants as either functionally deleterious (log₂ normalized fold change ≤0.24), indeterminate function (log₂ normalized fold change >0.24 and <1.09), or functionally neutral (log₂ normalized fold change ≥1.09). Using these thresholds, 12 of 18 VUSs (66.7%) previously reported to be functionally neutral were classified as functionally neutral, while 6 (33.3%) were of indeterminate function. Similarly, of 31 VUSs previously reported to be functionally deleterious, 30 (96.8%) were functionally deleterious and 1 (3.2%) was of indeterminate function (Figure 2—figure supplement 2A and B, Supplementary file 6). Overall, 632 of 2964 missense variants were functionally deleterious (21.3%), 674 variants were indeterminate function (22.7%), and 1658 variants were functionally neutral (55.9%) using log₂ normalized fold change to classify variants (Figure 1—figure supplement 2C, Supplementary file 6). Notably, 517 of 525 variants (98.5%) classified as functionally deleterious and 1586 of 1784 variants (88.9%) classified as functionally neutral using the gamma GLM were similarly classified using log₂ normalized fold change (Figure 2—figure supplement 2D).

To confirm the reproducibility of our variant classifications, 28 amino acid residues were assayed in duplicate, and variants classified using the gamma GLM. The majority of missense variants, 452 of 560 (80.7%), had the same functional classification in each of the two replicates (Figure 2—figure supplement 3A and B, Supplementary file 4). Of variants with discordant classifications, 6 (1.1%) were functionally deleterious in one replicate and of indeterminate function in another. While 102 variants (18.2%) were functionally neutral in one replicate and of indeterminate function in another. Importantly, no variant that was functionally deleterious in one replicate and functionally neutral in another (Supplementary file 4). Furthermore, the correlation coefficient between duplicate assay results was similar using the gamma GLM and log₂ normalized fold change (Figure 2—figure supplement 3A and C). We also determined whether underrepresentation in the cell pool at day 9 affected variant functional classifications. Fifty-three of 2964 missense variants (1.8%) were present in the cell pool at day 9 of the first assay replicate (experiment 1) at <2%, as determined by the number of sequence reads supporting the variant (Figure 2—figure supplement 4A, Supplementary file 4). There was no statistically significant difference in the proportion of variants classified as functionally deleterious for variants present in less than 2% of the cell pool at day 9 (12 of 53 variants; 22.6%), and variants present in more than 2% of the cell pool (496 of 2911 variants; 17.0%) (p-value=0.28) (Figure 2—figure supplement 4). We also found no significant differences in the proportion of variants classified as functionally deleterious for variants present in more than 2% of the cell pool at day 9 when variants were binned in 1% intervals (Figure 2—figure supplement 4B).

Comparison to in silico prediction algorithms

As in silico predictions of variant effect are integrated into ACMG variant interpretation guidelines as supporting evidence, we compared the ability of different algorithms, including recently described algorithms that incorporate deep-learning models of protein structure, to predict the functional consequence of CDKN2A missense variants. We compared our functional classifications to predictions from Combined Annotation Dependent Depletion (CADD), Polymorphism Phenotyping v2 (PolyPhen-2), Sorting Intolerant From Tolerant (SIFT), Variant Effect Scoring Tool score (VEST), AlphaMissense, ESM1b, and PrimateAI-3D. In silico predictions for all missense variants were available for PolyPhen-2, SIFT, VEST, AlphaMissense, and ESM1b. For CADD and PrimateAI-3D, 910 (152 functionally deleterious, 196 indeterminate, and 562 functionally neutral) and 904 (152 functionally deleterious, 196 indeterminate, and 556 functionally neutral) missense variants had in silico predictions available respectively (Supplementary file 7). In silico variant effect predictors performed similarly across a broad range of performance characteristics (Supplementary file 8). Accuracy of in silico model predictions were 39.5–85.4% (CADD – 45.1%; PolyPhen-2 – 39.5%; SIFT – 60.9%; VEST – 71.9%; AlphaMissense – 71.6%; ESM1b – 59.2%; and PrimateAI-3D – 85.4%) (Figure 3). We also assessed sensitivity, specificity, positive predictive value, and negative predictive value for each model. We found that sensitivity was 0.25–0.98 (CADD – 0.97; PolyPhen-2 – 0.98; SIFT – 0.79; VEST – 0.91; AlphaMissense – 0.94; ESM1b – 0.95; and PrimateAI-3D – 0.25), specificity was 0.27–0.98 (CADD – 0.35; PolyPhen-2 – 0.27; SIFT – 0.57; VEST – 0.68; AlphaMissense – 0.67; ESM1b – 0.51; and PrimateAI-3D – 0.98), positive predictive value was 0.22–0.68 (CADD – 0.23; PolyPhen-2 – 0.22; SIFT – 0.28; VEST – 0.38; AlphaMissense – 0.38; ESM1b – 0.3; and PrimateAI-3D – 0.68), and negative predictive value was 0.87–0.98 (CADD – 0.98; PolyPhen-2 – 0.98; SIFT – 0.93; VEST – 0.97; AlphaMissense – 0.98; ESM1b – 0.98; and PrimateAI-3D – 0.87).

Figure 3 with 2 supplements see all

Download asset Open asset

Comparison of functional classifications and in silico variant effect predictions for all possible CDKN2A missense variants.

Variant effect predictions for *CDKN2A* missense variants using CADD, PolyPhen-2, SIFT, VEST, AlphaMissense, ESM1b, and PrimateAI-3D. Predicted deleterious, damaging, or pathogenic effects (black box) and predicted neutral, tolerated, benign, or ambiguous effects (white box) presented as percent of missense variants with an available prediction. Number of missense variants with an available prediction for each in silico model given in parentheses. Accuracy shown as a red line. CADD: Combined Annotation Dependent Depletion; PolyPhen-2: Polymorphism Phenotyping v2; SIFT: Sorting Intolerant From Tolerant; VEST: Variant Effect Scoring Tool score.

Figure 3—source data 1 Raw data in Figure 3.: https://cdn.elifesciences.org/articles/95347/elife-95347-fig3-data1-v1.xlsx
Download elife-95347-fig3-data1-v1.xlsx

We also tested the effect of combining multiple in silico predictors. 904 missense variants had in silico predictions from all seven algorithms. The remaining 2060 missense variants had in silico predictions from five algorithms. Of variants with in silico predictions from all seven algorithms, 378 (41.8%) had predictions of deleterious or pathogenic effect from a majority of algorithms (≥4), and of these, 137 (36.2%) were functionally deleterious in our assay. Similarly, of 2060 missense variants that had in silico predictions from five algorithms, 1107 (53.7%) had predictions of deleterious or pathogenic effect from a majority of algorithms (≥3), of which, 361 (32.6%) were functionally deleterious in our assay (Supplementary file 7).

Distribution of functionally deleterious variants

Analysis of functionally deleterious variants may highlight critical and non-critical resides for CDKN2A function. We found that functionally deleterious missense variants were not distributed evenly across CDKN2A. CDKN2A contains four ankyrin repeats that mediate protein-protein interactions, ankyrin repeat 1 at codon 11–40, ankyrin repeat 2 at codon 44–72, ankyrin repeat 3 at codon 77–106, and ankyrin repeat 4 at codon 110–139 (Goldstein, 2004; Ruas and Peters, 1998; Sun et al., 2010; Figure 2—figure supplement 5A). Functionally deleterious variants were enriched in ankyrin repeat 1 (21.0%, adjusted p-value=0.01), ankyrin repeat 2 (26.2%, adjusted p-value=1.0 × 10^–10), and ankyrin repeat 3 (26.3%, adjusted p-value=2.6 × 10^–11), while depleted in ankyrin repeat 4 (6.5%, adjusted p-value=3.2 × 10^–13) and non-ankyrin repeat regions (6.8%, adjusted p-value=0) (Figure 2—figure supplement 5B). Moreover, functionally deleterious variants were further enriched within 10 residue subregions of ankyrin repeats 1–3, with 37.0% of variants in residues 16–25 of ankyrin repeat 1, 40.0% of variants in residues 46–55 of ankyrin repeat 2, and 48.0% of variants in residues 80–89 of ankyrin repeat 3 being classified as functionally deleterious (Figure 2C, Supplementary file 4).

Across all single residues, the mean percent of functionally deleterious missense variants was 17.7% (95% confidence interval: 12.7–20.9%) (Figure 2—figure supplement 5C, Supplementary file 4). At five amino acid residues, p.G23, p.G55, p.H83, p.D84, and p.G89, 17 of 19 (89.5%) possible missense variants were functionally deleterious. Notably, these residues are conserved between human and murine p16 (Byeon et al., 1998). And p.H83 has been reported to stabilize peptide loops connecting the helix-turn-helix structure of the four ankyrin repeats (Byeon et al., 1998), whereas p.D84 and p.G89 are located in a 20-residue region reported to interact with CDK4 and CDK6 (Fåhraeus et al., 1996). Conversely, 18 residues were tolerant of amino acid substitutions, with no missense variant characterized as functionally deleterious in our assay (Figure 2—figure supplement 5C, Supplementary file 4).

We also determined whether the location of variants in protein domains correlated with in silico predictions for the 904 missense variants with predictions from all seven algorithms (Figure 3—figure supplement 1A–H) and the 2060 missense variants with predictions from five algorithms (Figure 3—figure supplement 2A–H). Notably, Ank2 and Ank3 domains had more variants predicted to have deleterious or pathogenic effect by the majority of algorithms compared to Ank1, Ank4, and non-Ank domains (Figure 3—figure supplement 1C, Figure 3—figure supplement 2C). We also found increasing agreement between in silico predictions of deleterious or pathogenic effect and functionally deleterious classification in our assay as the number of algorithms predicting deleterious or pathogenic effects increased (Figure 3—figure supplement 1B, Figure 3—figure supplement 2B). This was true for all CDKN2A protein domains assessed (Figure 3—figure supplement 1D–H, Figure 3—figure supplement 2D–H).

Functional effect of CDKN2A somatic mutations

Somatic alterations in CDKN2A are a frequent finding in many types of cancer. However, not all somatic alterations are unequivocally deleterious to protein function. Missense somatic mutations are particularly challenging to functionally interpret and the presence of a functionally neutral somatic mutation may impact patient care (Tung et al., 2020). To understand the functional effect of missense somatic mutations in CDKN2A, we functionally classified mutations reported in the Catalogue Of Somatic Mutations In Cancer (COSMIC) (Forbes et al., 2010), The Cancer Genome Atlas (TCGA) (Muddabhaktuni and Koyyala, 2021), patients with cancer undergoing sequencing at The Johns Hopkins University School of Medicine (JHU), and the Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets Clinical Sequencing Cohort (MSK-IMPACT) (Cheng et al., 2015). Overall, 355 unique missense somatic mutations were reported, of which 119 (33.5%) were functionally deleterious in our assay (Supplementary file 9). The percent of missense somatic mutations that were classified as functionally deleterious was greater than the percent of all possible CDKN2A missense variants classified as functionally deleterious, suggesting enrichment of functionally deleterious missense changes among somatic mutations (Figure 2A, Supplementary file 4, Supplementary file 9). The proportion of missense somatic mutations that were functionally deleterious was similar in COSMIC, TCGA, JHU, and MSK-IMPACT. We found that 34.2–53.4% of unique missense somatic mutations classified as functionally deleterious, with 61.4–67.6% of patients having a functionally deleterious somatic mutation (Figure 4A, Supplementary file 9). As with functionally deleterious variants, functionally deleterious missense somatic mutations were also not distributed evenly across CDKN2A, being enriched within the ankyrin repeat 3 (Figure 4B, Supplementary file 9). We found that 32.4–50.0% of all functionally deleterious missense somatic mutations occurred within ankyrin repeat 3, with 48.0–58.0% of patients in each cohort having a functionally deleterious missense somatic mutation in this domain. Notably, 65.7–76.0% of functionally deleterious missense somatic mutations in this domain were in residues 80–89 (Supplementary file 9).

Figure 4 with 2 supplements see all

Download asset Open asset

Functional classification of missense somatic mutations in *CDKN2A*.

(A) Somatic missense variants in *CDKN2A* reported in COSMIC, TCGA, JHU, or MSK-IMPACT, by functional classification (deleterious – black box; indeterminate – gray box; neutral – white box). (B) Distribution of functionally deleterious missense somatic mutations *CDKN2A* reported in COSMIC, TCGA, JHU, or MSK-IMPACT by ankyrin (ANK) repeat. (C) Percent of missense somatic mutations in *CDKN2A* that were classified as functionally deleterious (black box), indeterminate function (gray box), or functionally neutral (white box) group by tumor type. Missense somatic mutations reported in COSMIC, TCGA, JHU, and MSK-IMPACT were combined. The number of missense somatic mutations for each tumor type given in parentheses. COSMIC; the Catalogue Of Somatic Mutations In Cancer, TCGA; The Cancer Genome Atlas, JHU; The Johns Hopkins University School of Medicine, MSK-IMPACT; Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets.

Figure 4—source data 1 Raw data in Figure 4A and B.: https://cdn.elifesciences.org/articles/95347/elife-95347-fig4-data1-v1.xlsx
Download elife-95347-fig4-data1-v1.xlsx
Figure 4—source data 2 Raw data in Figure 4C.: https://cdn.elifesciences.org/articles/95347/elife-95347-fig4-data2-v1.xlsx
Download elife-95347-fig4-data2-v1.xlsx

When considering unique missense somatic mutations, 26 of 355 (7.3%) would be classified as pathogenic or likely pathogenic by ACMG classification guidelines and these were found in 263 of 1176 (22.4%) patients in COSMIC, 45 of 185 (24.3%) patients in TCGA, 40 of 184 (21.7%) patients in JHU, and 46 of 174 (26.4%) patients in MSK-IMPACT (Figure 4—figure supplement 1A and B). In each cohort, the most prevalent of these somatic mutations were p.His83Tyr and p.Asp84Asn, with more than half of the patients with a somatic mutation that could be classified as pathogenic or likely pathogenic having either the p.His83Tyr or p.Asp84Asn alteration (Figure 4—figure supplement 1C). In our functional assays, these somatic mutations were both classified as functionally deleterious.

We were also able to determine the functional classification of CDKN2A missense somatic mutations in COSMIC, TCGA, JHU, and MSK-IMAPCT by cancer type. We found that 22.2–100% of CDKN2A missense somatic mutations were functionally deleterious depending on cancer type (Figure 4—figure supplement 2A–D). When considering missense somatic mutation reported in any database, there was a statistically significant depletion of functionally deleterious mutations in colorectal adenocarcinoma (20.4%; adjusted p-value=5.4 × 10^–9) (Figure 4C). As the proportion of missense somatic mutations that were functionally deleterious was less in colorectal carcinoma compared to other types of cancer, we assessed whether somatic mutations in mismatch repair genes (MLH1, MLH3, MSH2, MSH6, PMS1, and PMS2) were associated with the functional status of CDKN2A missense somatic mutations. Thirty-five patients in COSMIC had a CDKN2A missense somatic mutation, of which 12 (34.3%) had a somatic mutation in a mismatch repair gene. We found that no patients with a somatic mutation in a mismatch repair gene had a functionally deleterious CDKN2A missense somatic mutation compared to 6 of 23 samples (26.1%) without a somatic mutation in a mismatch repair gene (p-value=0.062).

CDKN2A variants in variant databases

The Genome Aggregation Database (gnomAD) v4.1.0 reports 287 missense variants in CDKN2A, including the 13 pathogenic, 4 likely pathogenic, 3 likely benign, 3 benign, and 264 VUSs classified using ACMG variant interpretation guidelines (Figure 5A and B, Supplementary file 10). Of the 264 missense VUSs, 177 were functionally neutral (67.0%), 56 (21.2%) were indeterminate function, and 31 (11.7%) were functionally deleterious in our assay using the gamma GLM for classification (Figure 5C). Similarly, ClinVar reports 395 CDKN2A missense VUSs, of which 256 (64.8%) were functionally neutral, 94 (23.8%) were indeterminate function, and 45 (11.4%) were functionally deleterious in our assay (Figure 5D, Supplementary file 11).

Figure 5

Download asset Open asset

*CDKN2A* synonymous and missense variants reported in gnomAD and ClinVar.

(A) Synonymous and missense variants in *CDKN2A* reported in gnomAD. (B) 287 *CDKN2A* missense variants reported in gnomAD, by American College of Medical Genetics (ACMG) guideline classification. (C) 264 missense variants in *CDKN2A* reported in gnomAD, by functional classification (deleterious – black box; indeterminate – gray box; neutral – white box). (D) 395 missense variants in *CDKN2A* reported in ClinVar, by functional classification (deleterious – black box; indeterminate – gray box; neutral – white box).

Discussion

VUSs in hereditary cancer susceptibility genes, predominantly rare missense variants, are a frequent finding in patients undergoing genetic testing and the cause of significant uncertainty. ACMG variant interpretation guidelines incorporate functional data, as well as other evidence such as in silico predictions of variant effect, to aid classification of variants as either pathogenic or benign. CDKN2A VUSs are a frequent finding in patients with PDAC. We previously found that over 40% of CDKN2A VUSs identified in patients with PDAC were functionally deleterious and therefore could be reclassified as likely pathogenic. In this study, we developed, a high-throughput, in vitro assay and functionally characterized 2964 CDKN2A missense variants, representing all possible single amino acid variants. We found that 525 missense variants (17.7%) were functionally deleterious. These predefined functional characterizations are resource for the scientific community and can be integrated into variant interpretation schema to aid classification of CDKN2A germline variants and somatic mutations.

We classified CDKN2A missense variants using a gamma GLM as either functionally deleterious, indeterminate functional, or functionally neutral. However, we did not classify variants that may have gain-of-function effects, resulting in decreased representation in the cell pool. Future studies are necessary to determine the prevalence and significance of CDKN2A gain-of-function variants.

Importantly, variant classifications using a gamma GLM were not biased by assay outputs for previously reported – benchmark – pathogenic or begin variants to determine thresholds. Classification thresholds were determined using the change in representation of 20 nonfunctional barcodes in a pool of PANC-1 cells stably expressing CDKN2A after a period of in vitro proliferation. Even so, CDKN2A missense variant classifications were remarkably similar using a gamma GLM or normalized fold change with thresholds determined using benchmark pathogenic and begin variants. Of missense variants classified as functionally deleterious using a gamma GLM, 98.5% were similarly classified using normalized fold change.

We repeated our functional assay twice for 28 CDKN2A residues. For the remaining 128 residues of CDKN2A, the functional assay was completed once. While we found general agreement between functional classifications from each replicate for the 28 residues assayed in duplicate, additional repeats for each residue are necessary to determine variability in variant functional classifications.

Our characterization of all possible CDKN2A missense variants allowed us to assess the ability of in silico algorithms – including recently published predictors based on machine learning AlphaMissense, ESM1b, and PrimateAI-3D – to predict the pathogenicity or functional effect of CDKN2A missense variants. We found that all in silico variant effect predictors assessed performed similarly. Highest accuracy was observed with PrimateAI-3D at 85.4%, followed by VEST at 71.9% and AlphaMissense at 71.6%. Importantly, even in silico predictors performing best in one metric may perform poorly in others. For example, PrimateAI-3D had the highest specificity (0.98) and positive predictive values (0.68), but the lowest sensitivity (0.25) and negative predictive value (0.87). Given that reclassification of VUSs in hereditary cancer genes into inappropriate strata has significant implications for patients, use of in silico models for clinical variant interpretation, including those utilizing machine learning, may be premature. Ultimately, our data support current ACMG guidelines that include in silico predictions of variant effect as supporting evidence of pathogenicity or benign impact.

Our study also provides other insights for the implementation of variant interpretation guidelines. ACMG guidelines include presence of a missense variant at a residue with a previously reported pathogenic variant as moderate evidence of pathogenicity. We found that functionally deleterious missense variants were not evenly distributed across CDKN2A. We found enrichment of functionally deleterious missense variants in ankyrin repeats 1–3 and depletion in ankyrin repeat 4. Notably, no CDKN2A residue was completely intolerant of amino acid changes. Suggesting, at least for CDKN2A, that the presence of a pathogenic missense variant at a residue should be used with caution when classifying other missense variants at the same residue.

We characterized variants based upon a broad cellular phenotype, cell proliferation, in a single PDAC cell line. It is possible that CDKN2A variant functional classifications are cell-specific and assay-specific. Our assay may not encompass all cellular functions of CDKN2A and an alternative assay of a specific CDKN2A function, such as CDK4 binding, may result in different variant functional classifications. Furthermore, CDKN2A variants may have different effects if alternative cell lines are used for the functional assay. However, cell-specific effects appear to be limited. In our previous study, we characterized 29 CDKN2A VUSs in three PDAC cell lines, using cell proliferation and cell cycle assays, and found agreement between all functional classifications (Kimura et al., 2022).

This study supports the utility of our in vitro functional assay. In general, we found that benchmark pathogenic variants, benchmark benign variants, and VUSs previously reported to be functionally deleterious had congruent functional classifications in our assay. Moreover, we found that functionally deleterious effects were enriched among somatic missense mutations, and depleted in missense VUSs in gnomAD, compared to all CDKN2A missense variants. Importantly, our functionally assay provides evidence to reclassify 301 of 395 (76.2%) missense VUSs reported in ClinVar and 208 of 264 (78.8%) missense VUSs reported in gnomAD. These include 45 (11.4%) VUSs in ClinVar and 31 missense VUSs in gnomAD that could be reclassified as likely pathogenic variants.

In this study, we determined functional classifications for all possible CDKN2A missense variants. Comparison of our functional classifications to in silico variant effect predictors, including recently described algorithms based on machine learning, provides performance benchmarks and supports current recommendations integrating data computational data into variant interpretation guidelines.

Methods

Key resources table

Reagent type (species) or resource	Designation	Source or reference	Identifiers	Additional information
Gene (Homo sapiens)	CDKN2A	GenBank	Gene ID: 1029 NM_000077.5 NP_000068.1
Cell line (H. sapiens)	PANC-1	American Type Culture Collection	Cat. #: CRL-1469 RRID:CVCL_0480
Cell line (H. sapiens)	293T	American Type Culture Collection	Cat. #: CRL-3216 RRID:CVCL_0063
Recombinant DNA reagent	pHAGE-CDKN2A (plasmid)	Addgene	RRID:Addgene_116726	Lentiviral vector expressing CDKN2A
Recombinant DNA reagent	pLentiV_ Blast (plasmid)	Addgene	RRID:Addgene_111887
Recombinant DNA reagent	pLJM1-Empty (plasmid)	Addgene	RRID:Addgene_91980
Recombinant DNA reagent	psPAX2 (plasmid)	Addgene	RRID:Addgene_12260
Recombinant DNA reagent	pCMV-VSV-G (plasmid)	Addgene	RRID:Addgene_8454
Recombinant DNA reagent	pLJM1-CDKN2A (plasmid)	Twist Bioscience		Lentiviral vector expressing codon-optimized CDKN2A
Recombinant DNA reagent	pLJM1-CDKN2A-Leu32Leu (plasmid)	This paper		Lentiviral vector expressing CDKN2A-Leu32Leu
Recombinant DNA reagent	pLJM1-CDKN2A-Leu32Pro (plasmid)	This paper		Lentiviral vector expressing CDKN2A- Leu32Pro
Recombinant DNA reagent	pLJM1-CDKN2A- Gly101Gly (plasmid)	This paper		Lentiviral vector expressing CDKN2A- Gly101Gly
Recombinant DNA reagent	pLJM1-CDKN2A- Gly101Trp (plasmid)	This paper		Lentiviral vector expressing CDKN2A- Gly101Trp
Recombinant DNA reagent	pLJM1-CDKN2A- Val126Asp (plasmid)	This paper		Lentiviral vector expressing CDKN2A- Val126Asp
Recombinant DNA reagent	pLJM1-CDKN2A- Val126Val (plasmid)	This paper		Lentiviral vector expressing CDKN2A- Val126Val
Recombinant DNA reagent	pLentiV-Blast-CellTag (plasmid)	This paper		Lentiviral vector expressing CellTAg
Sequence-based reagent	PCR primers	This paper		See Supplementary file 7
Commercial assay or kit	GenePrint 10 System	Promega Corporation	Cat. #: B9510
Commercial assay or kit	PCR-based MycoDtect kit	Greiner Bio-One	Cat. #: 463 060
Commercial assay or kit	Q5 Site-Directed Mutagenesis kit	New England Biolabs	Cat. #: E0552
Commercial assay or kit	PureLink Genomic DNA Mini Kit	Invitrogen	Cat. #: K1820-01
Commercial assay or kit	KAPA HiFi HotStart PCR Kit	Kapa Biosystems	Cat. #: KK2501
Commercial assay or kit	Qubit dsDNA HS assay kit	Invitrogen	Cat. #: Q33230
Commercial assay or kit	MiSeq Reagent Kit v2 (300 cycles)	Illumina	Cat. #: MS-102-2002
Chemical compound, drug	Dulbecco’s modified Eagle’s medium	Gibco/Thermo Fisher	Cat. #: 11995-065
Chemical compound, drug	Lipofectamine 3000 Transfection Reagent	Thermo Fisher Scientific	Cat. #: L3000008
Chemical compound, drug	Q5 Hot Start High-Fidelity 2X Master Mix	New England Biolabs	Cat. #: M0494S
Chemical compound, drug	Agencourt AMPure XP system	Beckman Coulter, Inc	Cat. #: A63881
Chemical compound, drug	Lenti-X Concentrator	Clontech	Cat. #: 631231
Chemical compound, drug	FxCycle Violet Ready Flow Reagent	Invitrogen	Cat. #: R37166
Software, algorithm	TC20 Automated Cell Counter	Bio-Rad Laboratories	Cat. #: 1450102
Software, algorithm	MiSeq System	Illumina
Software, algorithm	MiSeq control software	Illumina		Version 2.5.0.5
Software, algorithm	R	The R Foundation	RRID:SCR_001905	Version 4.2.0
Software, algorithm	JMP	SAS	RRID:SCR_014242	Version 11
Software, algorithm	Python statsmodel package	The Python Software Foundation	RRID:SCR_016074	Version 0.14.0

Cell lines

PANC-1 (American Type Culture Collection, Manassas, VA, USA; catalog no. CRL-1469), a human PDAC cell line with a homozygous deletion of CDKN2A (Caldas et al., 1994) and 293T (American Type Culture Collection; catalog no. CRL-3216), a human embryonic kidney cell line, were maintained in Dulbecco’s modified Eagle’s medium (Thermo Fisher Scientific Inc, Waltham, MA, USA; catalog no. 11995-065) supplemented with 10% fetal bovine serum (Thermo Fisher Scientific Inc; catalog no. 26140-079). Cell line authentication and mycoplasma testing were performed using the GenePrint 10 System (Promega Corporation, Madison, WI, USA; catalog no. B9510) and the PCR-based MycoDtect kit (Greiner Bio-One, Monroe, NC, USA; catalog no. 463 060) (Genetics Resource Core Facility, The Johns Hopkins University, Baltimore, MD, USA).

CDKN2A somatic mutation data

CDKN2A (p16^INK4; NP_000068.1) missense somatic mutation data was obtained from the Catalogue Of Somatic Mutations In Cancer (Forbes et al., 2010), The Cancer Genome Atlas (Muddabhaktuni and Koyyala, 2021), patients with cancer undergoing sequencing at The Johns Hopkins University School of Medicine (Baltimore, MD, USA), Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets Clinical Sequencing Cohort (Cheng et al., 2015). CDKN2A variant data was obtained from gnomAD v.4.1.0 and ClinVar (Landrum et al., 2014).

Plasmids

pHAGE-CDKN2A (Addgene, Watertown, MA, USA; plasmid no. 116726) was created by Gordon Mills & Kenneth Scott (Ng et al., 2018). pLJM1 (Addgene; plasmid no. 91980) was created by Joshua Mendell (Golden et al., 2017). pLentiV_Blast (Addgene, plasmid no. 111887) was created by Christopher Vakoc (Tarumoto et al., 2020). psPAX2 (Addgene, plasmid no. 12260) was created by Didier Trono, and pCMV-VSV-G (Addgene, plasmid no. 8454) was created by Bob Weinberg (Stewart et al., 2003).

CDKN2A expression plasmid libraries

Codon-optimized CDKN2A cDNA using p16^INK4A amino acid sequence (NP_000068.1) was designed (Supplementary file 12) and pLJM1 containing codon-optimized CDKN2A (pLJM1-CDKN2A) generated by Twist Bioscience (South San Francisco, CA, USA). 156 plasmid libraries were then synthesized by using pLJM1-CDKN2A, such that each library contained all possible 20 amino acids variants (19 missense and 1 synonymous) at a given position, generating 500 ng of each plasmid library (Twist Bioscience, South San Francisco, CA, USA). The proportion of variant in each library was shown in Supplementary file 2. Variants with a representation of less than 1% in a plasmid library were individually generated using the Q5 Site-Directed Mutagenesis kit (New England Biolabs, Ipswich, MA, USA; catalog no. E0552), and added to each library to a calculated proportion of 5%. Primers used for site-directed mutagenesis are given in Supplementary file 13. Each library was then amplified to generate at least 5 µg of plasmid DNA using QIAGEN Plasmid Midi Kit (QIAGEN, Germantown, MD, USA; catalog no. 12143).

Single variant CDKN2A expression plasmids

Individual pLJM1-CDKN2A expression constructs for CDKN2A missense variants, p.L32L, p.L32P, p.G101G, p.G101W, p.V126D, and p.V126V, were generated using the Q5 Site-Directed Mutagenesis kit (New England Biolabs, Ipswich, MA, USA; catalog no. E0552). Primers used for site-directed mutagenesis are given in Supplementary file 13. Integration of each CDKN2A variant was confirmed using Sanger sequencing (Genewiz, Plainsfield, NJ, USA) using the CMV Forward sequencing primer (CGCAAATGGGCGGTAGGCGTG). The manufacturer’s protocol was followed unless otherwise specified.

CellTag plasmid library

Twenty nonfunctional 9 base pair barcodes ‘CellTags’ were subcloned into pLentiV_Blast using the Q5 Site-Directed Mutagenesis kit (New England Biolabs, Ipswich, MA, USA; catalog no. E0552) (Biddy et al., 2018). Primers used to generate each CellTag plasmid are given in Supplementary file 13. Integration of each CellTag was confirmed using Sanger sequencing (Genewiz) (sequencing primer: AACTGGGAAAGTGATGTCGTG). The manufacturer’s protocol was followed unless otherwise specified. CellTag plasmids were then pooled to form a CellTag plasmid library with equal representation of each CellTag plasmid.

Lentivirus production

Lentivirus production was performed as previously described with the following modifications (Kimura et al., 2022). pLJM1 lentiviral expression vectors (plasmid libraries and single variant expression plasmids) and lentiviral packaging vectors (psPAX2 and pCMV-VSV-G) were transfected into 293T cells using Lipofectamine 3000 Transfection Reagent (Thermo Fisher Scientific, Waltham, MA, USA; catalog no. L3000008). Media was collected at 24 hr and 48 hr, pooled, and lentiviral particles concentrated using Lenti-X Concentrator (Clontech, Mountain View, CA, USA; catalog no. 631231) using the manufacturer’s protocol.

Lentiviral transduction

PANC-1 cells were used for CDKN2A plasmid library and single variant CDKN2A expression plasmid transductions. PANC-1 cells previously transduced with pLJM1-CDKN2A (PANC-1^CDKN2A) and selected with puromycin were used for CellTag library transductions. Briefly, 1×10⁵ cells were cultured in media supplemented with 10 µg/mL polybrene and transduced with 4×10⁷ transducing units per mL of lentivirus particles. Cells were then centrifuged at 1200 × g for 1 hr. After 48 hr of culture at 37°C and 5% CO₂, transduced cells were selected using 3 µg/mL puromycin (CDKN2A plasmid libraries and single variant CDKN2A expression plasmids) or 5 µg/mL blasticidin (CellTag plasmid library) for 7 days. Expected MOI was one. After selection, cells were trypsinized and 5×10⁵ cells were seeded into T150 flasks. DNA was collected from remaining cells and this sample was named as day 9. T150 flasks were cultured until confluent and then DNA was collected. The time for cells to become confluent varied for each amino acid residue (days 16–40, Supplementary file 5). DNA was extracted from PANC-1 cells using the PureLink Genomic DNA Mini Kit (Invitrogen, Carlsbad, CA, USA; catalog no. K1820-01). The assay for CellTag library was repeated in triplicate. We repeated our CDKN2A assay in duplicate for 28 residues. For the remaining 128 CDKN2A residues the assay was completed once.

Generation of sequence libraries

Library preparation and sequencing was performed as previously described with the following modifications (Kinde et al., 2011). For the first stage PCR, three target specific primers were designed to amplify CDKN2A amino acid positions 1–53, 54–110, and 111–156 (Supplementary file 13). Forward and reverse first stage primers contained 5’ M13F (GTAAAACGACGGCCAGC) and M13R (CAGGAAACAGCTATGAC) sequence, respectively, to enable amplification and ligation of Illumina adapter sequences in a second stage PCR (Supplementary file 13). DNA was amplified with Q5 Hot Start High-Fidelity 2X Master Mix (New England Biolabs; catalog no. M0494S). For the first stage PCR, each DNA sample was amplified in three reactions each containing 66 ng of DNA for 18 cycles. First stage PCR products for each sample were then pooled and purified using the Agencourt AMPure XP system (Beckman Coulter, Inc, Brea, CA, USA; catalog no. A63881), eluting into 50 µL of elution buffer. Purified PCR product was amplified in a second stage PCR to add Illumina adaptor sequences and indexes (Supplementary file 13). Second stage PCR Amplification was performed with KAPA HiFi HotStart PCR Kit (Kapa Biosystems, Wilmington, MA, USA; catalog no. KK2501) in 25 µL reactions containing 5X KAPA HiFi Buffer – 5 µL, 10 mM KAPA dNTP Mix – 0.75 µL, 10 μM forward primer – 0.75 µL, 10 μM reverse primer – 0.75 µL. For the first stage PCR, 66 ng of template DNA and 12.5 µL, Q5 Hot Start High-Fidelity 2X Master Mix was used with the following cycling conditions: 98°C for 30 s; 18 cycles of 98°C for 10 s, 72°C for 30 s, 72°C for 25 s; 72°C for 2 min. For the second stage PCR, 0.25 µL of first stage PCR product and 0.5 µL of 1 U/μL KAPA HiFi HotStart DNA Polymerase was used with the following cycling conditions: 95°C for 3 min; 25 cycles of 98°C for 20 s, 62°C for 15 s, 72°C for 1 min. Second stage PCR products were purified with the Agencourt AMPure XP system (Beckman Coulter, Inc; catalog no. A63881) into 30 µL of elution buffer. Samples were quantified by Qubit using dsDNA HS assay kit (Invitrogen; catalog no. Q33230).

Sequencing and analysis

Sequence libraries were pooled in equimolar amounts into groups of 16 samples and sequenced on the Illumina MiSeq System (Illumina, San Diego, CA, USA) with the MiSeq Reagent Kit v2 (300 cycles) (Illumina catalog no. MS-102-2002) to generate 150 base pair paired-end reads. Samples were demultiplexed and FASTQ sequence read files were generated with MiSeq control software 2.5.0.5 (Illumina). Paired sequence reads were then combined into a single contiguous sequence using Paired-End Read Merger (Zhang et al., 2014). Reads supporting each variant at a given amino acid position were counted using Perl.

Functional characterization of CDKN2A variants using a gamma GLM

We determined if a variant has a fitness advantage by assessing the significance of the observed ratio $r_{v, c f}$ at confluence between the number of cells with a missense variant $v$ and the number of cells with a synonymous variant at a given amino acid position. Using the missense variant as a benchmark variant, we assumed that the distribution of $r_{v, c f}$ can be explained by two key covariates: $r_{v, i n i t}$ , which represent the missense variant-to-synonymous variant ratio at day 9, and $p_{v, i n i t}$ , the proportion of the missense variant cells among other variants, including the synonymous variant, at the studied position. More specifically, given the variables $r_{v, i n i t} \land p_{v, i n i t}$ , the ratio at confluence follows a distribution:

r_{v, c f} \sim Γ (α, β_{v})

where the mean $u_{v}$ of the Gamma distribution is such that:

u_{v} = \frac{α}{β_{v}} = r_{v, i n i t}^{a} p_{v, i n i t}^{b} .

Here, the parameters of the null model to estimate are $α, a, a n d b$ , where $α$ is the shape parameter of the Gamma distribution and is assumed to be the same for all variants. This model is a gamma GLM over the response variable $r_{v, c f}$ with a log-link function and covariates $l o g (r_{v, i n i t})$ and $l o g (p_{v, i n i t}) .$ Estimating the parameters will provide a null distribution of $r_{v, c f}$ , generating a p-value for every observed $r_{v, c f}$ for any variant at a given position.

To estimate the parameters $α, a$ , and $b$ , we utilized three control experiments where the CellTag plasmid library was transduced into PANC-1^CDKN2Aco cells so that each CellTag represented a neutral variant. For a single experiment, every variant can be considered as wild-type, and we test the other 19 variants against it, knowing that they are neutral and therefore follow the null distribution. This provides us with 19×20 triplets $(r_{v, c f}, p_{v}^{i n i t}, r_{v}^{i n i t})$ , for every experiment, yielding 1140 data points when considering all three experiments together. To estimate the parameters using these 1140 data points, we fit the GLM corresponding GLM model using the sklearn.linear_model module.

After the estimation of parameters $α, a$ , and $b$ , every observation for a tested variant $v$ at a given position of the triplet $(r_{v, c f}, p_{v}^{i n i t}, r_{v}^{i n i t})$ yields a p-value, defined as the probability of observing a ratio at confluence that is at least $r_{v, c f}$ given $p_{v, i n i t}, r_{v, i n i t}$ under the null Gamma model. As some variants were tested in repeated experiments, we combined their associated p-values into a single p-value using Fisher’s method. Finally, to determine if a variant presents a fitness advantage, we apply a Benjamini-Hochberg estimator on all the tested variants p-values, fixing the false discovery rate at a level of 0.05.

Functional characterization of CDKN2A variants using log₂ normalized fold change

Fold change for each variant was calculated using the proportion of total reads representing a variant at confluency (days 16–40, Supplementary file 5) to the proportion of total reads representing a variant on day 9 after transfection. Fold change was then normalized to the synonymous variant at each residue and then log₂ normalized fold change values calculated (Supplementary file 4, Supplementary file 6). Variants with log₂ normalized fold change values greater than or equal to the minimum value of benchmark pathogenic variants were characterized as functionally deleterious, while variants with values smaller than or equal to the maximum value of benchmark benign variants were characterized as functionally neutral (Supplementary file 6). Log₂ normalized fold change values between these defined thresholds were classified as indeterminate. Mean values were used for replicated variants.

Data visualization

Heatmap of individual variant p-values by amino position was generated using R with the heatmaply package (Galili et al., 2018).

Cell proliferation assay

Cell proliferation assay was performed as previously described with the following modifications (Kimura et al., 2022). 1×10⁵ cells were seeded into in vitro culture (day 0). Cells were counted on day 14 using a TC20 Automated Cell Counter (Bio-Rad Laboratories, Hercules, CA, USA; catalog no. 1450102). Relative cell proliferation value was calculated as cell number normalized to empty vector control. Assays were repeated in triplicate. Mean cell proliferation value and standard deviation (s.d.) were calculated.

Variant effect predictions

Publicly available algorithms were used to predict the consequence of CDKN2A missense variants. Prediction algorithms used included: CADD (Kircher et al., 2014), PolyPhen-2 (Adzhubei et al., 2010), SIFT (Kumar et al., 2009), VEST (Carter et al., 2013), AlphaMissense (Cheng et al., 2023), ESM1b (Brandes et al., 2023), and PrimateAI-3D (Gao et al., 2023; Supplementary file 7). PolyPhen-2, SIFT, VEST, AlphaMissense, and ESM1b prediction were available for all missense variants. CADD scores were available for 910 missense variants and where multiple CADD scores were possible, mean values were used. PrimateAI-3D prediction scores were available for 904 assayed missense variants.

Statistical analyses

Statistical analyses were performed using JMP v.11 (SAS, Cary, NC, USA) and Python statsmodel package (version 0.14.0). Student’s t-tests was used to compare mean cell proliferation values. A chi-square test was used to compare the proportion of functionally deleterious variants for variants present in <2% and ≥2% of the cell pool at day 9. A Fisher’s exact test was used to compare prevalence of functionally deleterious CDKN2A variants in colorectal cancer cases from COSMIC with and without somatic mutations in mismatch repair genes. Z-tests with multiple test correction performed with the Bonferroni method was used in the following comparisons: (1) proportion of functionally deleterious variants present in <2% of the cell pool and ≥2% of the cell pool at day 9 binned in 1% intervals, (2) proportion of variants in each domain predicted to have deleterious or pathogenic effect by the majority of algorithms, (3) proportion of functionally deleterious variants in each domain, and (4) proportion of functionally deleterious missense variants and somatic mutations.

Data availability

All data generated or analysed during this study are included in the manuscript and supporting files; source data files have been provided.

References

(2010) A method and server for predicting damaging missense mutations
Nature Methods 7:248–249.

https://doi.org/10.1038/nmeth0410-248
- PubMed
- Google Scholar
1. Biddy BA
2. Kong W
3. Kamimoto K
4. Guo C
5. Waye SE
6. Sun T
7. Morris SA
(2018) Single-cell mapping of lineage and identity in direct reprogramming
Nature 564:219–224.

https://doi.org/10.1038/s41586-018-0744-4
- PubMed
- Google Scholar
1. Brandes N
2. Goldman G
3. Wang CH
4. Ye CJ
5. Ntranos V
(2023) Genome-wide prediction of disease variant effects with a deep protein language model
Nature Genetics 55:1512–1522.

https://doi.org/10.1038/s41588-023-01465-0
- PubMed
- Google Scholar
1. Brenan L
2. Andreev A
3. Cohen O
4. Pantel S
5. Kamburov A
6. Cacchiarelli D
7. Persky NS
8. Zhu C
9. Bagul M
10. Goetz EM
11. Burgin AB
12. Garraway LA
13. Getz G
14. Mikkelsen TS
15. Piccioni F
16. Root DE
17. Johannessen CM
(2016) Phenotypic characterization of a comprehensive set of MAPK1/ERK2 missense mutants
Cell Reports 17:1171–1183.

https://doi.org/10.1016/j.celrep.2016.09.061
- PubMed
- Google Scholar
1. Byeon I-JL
2. Li J
3. Ericson K
4. Selby TL
5. Tevelev A
6. Kim H-J
7. O’Maille P
8. Tsai M-D
(1998) Tumor suppressor p16INK4A: determination of solution structure and analyses of its interaction with cyclin-dependent kinase 4
Molecular Cell 1:421–431.

https://doi.org/10.1016/S1097-2765(00)80042-8
- Google Scholar
1. Caldas C
2. Hahn SA
3. da Costa LT
4. Redston MS
5. Schutte M
6. Seymour AB
7. Weinstein CL
8. Hruban RH
9. Yeo CJ
10. Kern SE
(1994) Frequent somatic mutations and homozygous deletions of the p16 (MTS1) gene in pancreatic adenocarcinoma
Nature Genetics 8:27–32.

https://doi.org/10.1038/ng0994-27
- PubMed
- Google Scholar
(2013) Identifying mendelian disease genes with the variant effect scoring tool
BMC Genomics 14 Suppl 3:S3.

https://doi.org/10.1186/1471-2164-14-S3-S3
- PubMed
- Google Scholar
1. Chaffee KG
2. Oberg AL
3. McWilliams RR
4. Majithia N
5. Allen BA
6. Kidd J
7. Singh N
8. Hartman AR
9. Wenstrup RJ
10. Petersen GM
(2018) Prevalence of germline mutations in cancer genes
Genetics in Medicine 20:119–127.

https://doi.org/10.1038/gim.2017.85
- Google Scholar
1. Chang MT
2. Asthana S
3. Gao SP
4. Lee BH
5. Chapman JS
6. Kandoth C
7. Gao JJ
8. Socci ND
9. Solit DB
10. Olshen AB
11. Schultz N
12. Taylor BS
(2016) Identifying recurrent mutations in cancer reveals widespread lineage diversity and mutational specificity
Nature Biotechnology 34:155–163.

https://doi.org/10.1038/nbt.3391
- PubMed
- Google Scholar
1. Cheng DT
2. Mitchell TN
3. Zehir A
4. Shah RH
5. Benayed R
6. Syed A
7. Chandramohan R
8. Liu ZY
9. Won HH
10. Scott SN
11. Brannon AR
12. O’Reilly C
13. Sadowska J
14. Casanova J
15. Yannes A
16. Hechtman JF
17. Yao J
18. Song W
19. Ross DS
20. Oultache A
21. Dogan S
22. Borsu L
23. Hameed M
24. Nafa K
25. Arcila ME
26. Ladanyi M
27. Berger MF
(2015) Memorial sloan kettering-integrated mutation profiling of actionable cancer targets (MSK-IMPACT): a hybridization capture-based next-generation sequencing clinical assay for solid tumor molecular oncology
The Journal of Molecular Diagnostics 17:251–264.

https://doi.org/10.1016/j.jmoldx.2014.12.006
- PubMed
- Google Scholar
1. Cheng J
2. Novati G
3. Pan J
4. Bycroft C
5. Žemgulytė A
6. Applebaum T
7. Pritzel A
8. Wong LH
9. Zielinski M
10. Sargeant T
11. Schneider RG
12. Senior AW
13. Jumper J
14. Hassabis D
15. Kohli P
16. Avsec Ž
(2023) Accurate proteome-wide missense variant effect prediction with AlphaMissense
Science 381:eadg7492.

https://doi.org/10.1126/science.adg7492
- PubMed
- Google Scholar
1. Cubuk C
2. Garrett A
3. Choi S
4. King L
5. Loveday C
6. Torr B
7. Burghel GJ
8. Durkie M
9. Callaway A
10. Robinson R
11. Drummond J
12. Berry I
13. Wallace A
14. Eccles D
15. Tischkowitz M
16. Whiffin N
17. Ware JS
18. Hanson H
19. Turnbull C
(2021) Clinical likelihood ratios and balanced accuracy for 44 in silico tools against multiple large-scale functional assays of cancer susceptibility genes
Genetics in Medicine 23:2096–2104.

https://doi.org/10.1038/s41436-021-01265-z
- Google Scholar
1. Fåhraeus R
2. Paramio JM
3. Ball KL
4. Laín S
5. Lane DP
(1996) Inhibition of pRb phosphorylation and cell-cycle progression by a 20-residue peptide derived from P16CDKN2/INK4A
Current Biology 6:84–91.

https://doi.org/10.1016/s0960-9822(02)00425-6
- PubMed
- Google Scholar
1. Forbes SA
2. Tang G
3. Bindal N
4. Bamford S
5. Dawson E
6. Cole C
7. Kok CY
8. Jia M
9. Ewing R
10. Menzies A
11. Teague JW
12. Stratton MR
13. Futreal PA
(2010) COSMIC (the catalogue of somatic mutations in cancer): a resource to investigate acquired mutations in human cancer
Nucleic Acids Research 38:D652–D657.

https://doi.org/10.1093/nar/gkp995
- PubMed
- Google Scholar
(2018) heatmaply: an R package for creating interactive cluster heatmaps for online publishing
Bioinformatics 34:1600–1602.

https://doi.org/10.1093/bioinformatics/btx657
- PubMed
- Google Scholar
1. Gao H
2. Hamp T
3. Ede J
4. Schraiber JG
5. McRae J
6. Singer-Berk M
7. Yang Y
8. Dietrich ASD
9. Fiziev PP
10. Kuderna LFK
11. Sundaram L
12. Wu Y
13. Adhikari A
14. Field Y
15. Chen C
16. Batzoglou S
17. Aguet F
18. Lemire G
19. Reimers R
20. Balick D
21. Janiak MC
22. Kuhlwilm M
23. Orkin JD
24. Manu S
25. Valenzuela A
26. Bergman J
27. Rousselle M
28. Silva FE
29. Agueda L
30. Blanc J
31. Gut M
32. de Vries D
33. Goodhead I
34. Harris RA
35. Raveendran M
36. Jensen A
37. Chuma IS
38. Horvath JE
39. Hvilsom C
40. Juan D
41. Frandsen P
42. de Melo FR
43. Bertuol F
44. Byrne H
45. Sampaio I
46. Farias I
47. do Amaral JV
48. Messias M
49. da Silva MNF
50. Trivedi M
51. Rossi R
52. Hrbek T
53. Andriaholinirina N
54. Rabarivola CJ
55. Zaramody A
56. Jolly CJ
57. Phillips-Conroy J
58. Wilkerson G
59. Abee C
60. Simmons JH
61. Fernandez-Duque E
62. Kanthaswamy S
63. Shiferaw F
64. Wu D
65. Zhou L
66. Shao Y
67. Zhang G
68. Keyyu JD
69. Knauf S
70. Le MD
71. Lizano E
72. Merker S
73. Navarro A
74. Bataillon T
75. Nadler T
76. Khor CC
77. Lee J
78. Tan P
79. Lim WK
80. Kitchener AC
81. Zinner D
82. Gut I
83. Melin A
84. Guschanski K
85. Schierup MH
86. Beck RMD
87. Umapathy G
88. Roos C
89. Boubli JP
90. Lek M
91. Sunyaev S
92. O’Donnell-Luria A
93. Rehm HL
94. Xu J
95. Rogers J
96. Marques-Bonet T
97. Farh KK-H
(2023) The landscape of tolerated genetic variation in humans and primates
Science 380:eabn8153.

https://doi.org/10.1126/science.abn8197
- PubMed
- Google Scholar
1. Giacomelli AO
2. Yang X
3. Lintner RE
4. McFarland JM
5. Duby M
6. Kim J
7. Howard TP
8. Takeda DY
9. Ly SH
10. Kim E
11. Gannon HS
12. Hurhula B
13. Sharpe T
14. Goodale A
15. Fritchman B
16. Steelman S
17. Vazquez F
18. Tsherniak A
19. Aguirre AJ
20. Doench JG
21. Piccioni F
22. Roberts CWM
23. Meyerson M
24. Getz G
25. Johannessen CM
26. Root DE
27. Hahn WC
(2018) Mutational processes shape the landscape of TP53 mutations in human cancer
Nature Genetics 50:1381–1387.

https://doi.org/10.1038/s41588-018-0204-y
- PubMed
- Google Scholar
1. Goggins M
2. Overbeek KA
3. Brand R
4. Syngal S
5. Del Chiaro M
6. Bartsch DK
7. Bassi C
8. Carrato A
9. Farrell J
10. Fishman EK
11. Fockens P
12. Gress TM
13. van Hooft JE
14. Hruban RH
15. Kastrinos F
16. Klein A
17. Lennon AM
18. Lucas A
19. Park W
20. Rustgi A
21. Simeone D
22. Stoffel E
23. Vasen HFA
24. Cahen DL
25. Canto MI
26. Bruno M
27. International Cancer of the Pancreas Screening (CAPS) Consortium
(2020) Management of patients with increased risk for familial pancreatic cancer: updated recommendations from the International Cancer of the Pancreas Screening (CAPS) Consortium
Gut 69:7–17.

https://doi.org/10.1136/gutjnl-2019-319352
- PubMed
- Google Scholar
1. Golan T
2. Hammel P
3. Reni M
4. Van Cutsem E
5. Macarulla T
6. Hall MJ
7. Park J-O
8. Hochhauser D
9. Arnold D
10. Oh D-Y
11. Reinacher-Schick A
12. Tortora G
13. Algül H
14. O’Reilly EM
15. McGuinness D
16. Cui KY
17. Schlienger K
18. Locker GY
19. Kindler HL
(2019) Maintenance olaparib for germline BRCA-mutated metastatic pancreatic cancer
The New England Journal of Medicine 381:317–327.

https://doi.org/10.1056/NEJMoa1903387
- PubMed
- Google Scholar
1. Golden RJ
2. Chen B
3. Li T
4. Braun J
5. Manjunath H
6. Chen X
7. Wu J
8. Schmid V
9. Chang TC
10. Kopp F
11. Ramirez-Martinez A
12. Tagliabracci VS
13. Chen ZJ
14. Xie Y
15. Mendell JT
(2017) An argonaute phosphorylation cycle promotes microRNA-mediated silencing
Nature 542:197–202.

https://doi.org/10.1038/nature21025
- PubMed
- Google Scholar
1. Goldstein AM
(2004) Familial melanoma, pancreatic cancer and germline CDKN2A mutations
Human Mutation 23:630.

https://doi.org/10.1002/humu.9247
- PubMed
- Google Scholar
1. Horn IP
2. Marks DL
3. Koenig AN
4. Hogenson TL
5. Almada LL
6. Goldstein LE
7. Romecin Duran PA
8. Vera R
9. Vrabel AM
10. Cui G
11. Rabe KG
12. Bamlet WR
13. Mer G
14. Sicotte H
15. Zhang C
16. Li H
17. Petersen GM
18. Fernandez-Zapico ME
(2021) A rare germline CDKN2A variant (47T>G; p16-L16R) predisposes carriers to pancreatic cancer by reducing cell cycle inhibition
The Journal of Biological Chemistry 296:100634.

https://doi.org/10.1016/j.jbc.2021.100634
- PubMed
- Google Scholar
1. Hu C
2. Hart SN
3. Polley EC
4. Gnanaolivu R
5. Shimelis H
6. Lee KY
7. Lilyquist J
8. Na J
9. Moore R
10. Antwi SO
11. Bamlet WR
12. Chaffee KG
13. DiCarlo J
14. Wu Z
15. Samara R
16. Kasi PM
17. McWilliams RR
18. Petersen GM
19. Couch FJ
(2018) Association between inherited germline mutations in cancer predisposition genes and risk of pancreatic cancer
JAMA 319:2401–2409.

https://doi.org/10.1001/jama.2018.6228
- PubMed
- Google Scholar
1. Jaffe A
2. Wojcik G
3. Chu A
4. Golozar A
5. Maroo A
6. Duggal P
7. Klein AP
(2011) Identification of functional genetic variation in exome sequence analysis
BMC Proceedings 5 Suppl 9:9–13.

https://doi.org/10.1186/1753-6561-5-S9-S13
- PubMed
- Google Scholar
(2021) The role of inherited pathogenic CDKN2A variants in susceptibility to pancreatic cancer
Pancreas 50:1123–1130.

https://doi.org/10.1097/MPA.0000000000001888
- PubMed
- Google Scholar
(2022) Functional CDKN2A assay identifies frequent deleterious alleles misclassified as variants of uncertain significance
eLife 11:e1137.

https://doi.org/10.7554/eLife.71137
- Google Scholar
(2011) Detection and quantification of rare mutations with massively parallel sequencing
PNAS 108:9530–9535.

https://doi.org/10.1073/pnas.1105422108
- PubMed
- Google Scholar
1. Kircher M
2. Witten DM
3. Jain P
4. O’Roak BJ
5. Cooper GM
6. Shendure J
(2014) A general framework for estimating the relative pathogenicity of human genetic variants
Nature Genetics 46:310–315.

https://doi.org/10.1038/ng.2892
- PubMed
- Google Scholar
1. Kumar P
2. Henikoff S
3. Ng PC
(2009) Predicting the effects of coding non-synonymous variants on protein function using the SIFT algorithm
Nature Protocols 4:1073–1081.

https://doi.org/10.1038/nprot.2009.86
- PubMed
- Google Scholar
1. Landrum MJ
2. Lee JM
3. Riley GR
4. Jang W
5. Rubinstein WS
6. Church DM
7. Maglott DR
(2014) ClinVar: public archive of relationships among sequence variation and human phenotype
Nucleic Acids Research 42:D980–D985.

https://doi.org/10.1093/nar/gkt1113
- PubMed
- Google Scholar
1. McWilliams RR
2. Wieben ED
3. Chaffee KG
4. Antwi SO
5. Raskin L
6. Olopade OI
7. Li D
8. Highsmith WE
9. Colon-Otero G
10. Khanna LG
11. Permuth JB
12. Olson JE
13. Frucht H
14. Genkinger J
15. Zheng W
16. Blot WJ
17. Wu L
18. Almada LL
19. Fernandez-Zapico ME
20. Sicotte H
21. Pedersen KS
22. Petersen GM
(2018) CDKN2A germline rare coding variants and risk of pancreatic cancer in minority populations
Cancer Epidemiology, Biomarkers & Prevention 27:1364–1370.

https://doi.org/10.1158/1055-9965.EPI-17-1065
- PubMed
- Google Scholar
1. Muddabhaktuni BMC
2. Koyyala VPB
(2021) The cancer genome atlas
Indian Journal of Medical and Paediatric Oncology 42:353–355.

https://doi.org/10.1055/s-0041-1735440
- Google Scholar
1. Ng PKS
2. Li J
3. Jeong KJ
4. Shao S
5. Chen H
6. Tsang YH
7. Sengupta S
8. Wang Z
9. Bhavana VH
10. Tran R
11. Soewito S
12. Minussi DC
13. Moreno D
14. Kong K
15. Dogruluk T
16. Lu H
17. Gao J
18. Tokheim C
19. Zhou DC
20. Johnson AM
21. Zeng J
22. Ip CKM
23. Ju Z
24. Wester M
25. Yu S
26. Li Y
27. Vellano CP
28. Schultz N
29. Karchin R
30. Ding L
31. Lu Y
32. Cheung LWT
33. Chen K
34. Shaw KR
35. Meric-Bernstam F
36. Scott KL
37. Yi S
38. Sahni N
39. Liang H
40. Mills GB
(2018) Systematic functional annotation of somatic mutations in cancer
Cancer Cell 33:450–462.

https://doi.org/10.1016/j.ccell.2018.01.021
- PubMed
- Google Scholar
(2015) Standards and guidelines for the interpretation of sequence variants: A joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology
Genetics in Medicine 17:405–424.

https://doi.org/10.1038/gim.2015.30
- PubMed
- Google Scholar
1. Roberts NJ
2. Norris AL
3. Petersen GM
4. Bondy ML
5. Brand R
6. Gallinger S
7. Kurtz RC
8. Olson SH
9. Rustgi AK
10. Schwartz AG
11. Stoffel E
12. Syngal S
13. Zogopoulos G
14. Ali SZ
15. Axilbund J
16. Chaffee KG
17. Chen Y-C
18. Cote ML
19. Childs EJ
20. Douville C
21. Goes FS
22. Herman JM
23. Iacobuzio-Donahue C
24. Kramer M
25. Makohon-Moore A
26. McCombie RW
27. McMahon KW
28. Niknafs N
29. Parla J
30. Pirooznia M
31. Potash JB
32. Rhim AD
33. Smith AL
34. Wang Y
35. Wolfgang CL
36. Wood LD
37. Zandi PP
38. Goggins M
39. Karchin R
40. Eshleman JR
41. Papadopoulos N
42. Kinzler KW
43. Vogelstein B
44. Hruban RH
45. Klein AP
(2016) Whole genome sequencing defines the genetic heterogeneity of familial pancreatic cancer
Cancer Discovery 6:166–175.

https://doi.org/10.1158/2159-8290.CD-15-0402
- PubMed
- Google Scholar
1. Ruas M
2. Peters G
(1998) The p16INK4a/CDKN2A tumor suppressor and its relatives
Biochimica et Biophysica Acta (BBA) - Reviews on Cancer 1378:F115–F177.

https://doi.org/10.1016/S0304-419X(98)00017-1
- Google Scholar
1. Shindo K
2. Yu J
3. Suenaga M
4. Fesharakizadeh S
5. Cho C
6. Macgregor-Das A
7. Siddiqui A
8. Witmer PD
9. Tamura K
10. Song TJ
11. Navarro Almario JA
12. Brant A
13. Borges M
14. Ford M
15. Barkley T
16. He J
17. Weiss MJ
18. Wolfgang CL
19. Roberts NJ
20. Hruban RH
21. Klein AP
22. Goggins M
(2017) Deleterious germline mutations in patients with apparently sporadic pancreatic adenocarcinoma
Journal of Clinical Oncology 35:3382–3390.

https://doi.org/10.1200/JCO.2017.72.3502
- PubMed
- Google Scholar
1. Stewart SA
2. Dykxhoorn DM
3. Palliser D
4. Mizuno H
5. Yu EY
6. An DS
7. Sabatini DM
8. Chen ISY
9. Hahn WC
10. Sharp PA
11. Weinberg RA
12. Novina CD
(2003) Lentivirus-delivered stable gene silencing by RNAi in primary cells
RNA 9:493–501.

https://doi.org/10.1261/rna.2192803
- PubMed
- Google Scholar
1. Stoffel EM
2. McKernin SE
3. Brand R
4. Canto M
5. Goggins M
6. Moravek C
7. Nagarajan A
8. Petersen GM
9. Simeone DM
10. Yurgelun M
11. Khorana AA
(2019) Evaluating susceptibility to pancreatic cancer: ASCO provisional clinical opinion
Journal of Clinical Oncology 37:153–164.

https://doi.org/10.1200/JCO.18.01489
- PubMed
- Google Scholar
1. Sun P
2. Nallar SC
3. Raha A
4. Kalakonda S
5. Velalar CN
6. Reddy SP
7. Kalvakolanu DV
(2010) GRIM-19 and p16(INK4a) synergistically regulate cell cycle progression and E2F1-responsive gene expression
The Journal of Biological Chemistry 285:27545–27552.

https://doi.org/10.1074/jbc.M110.105767
- PubMed
- Google Scholar
1. Tarumoto Y
2. Lin S
3. Wang J
4. Milazzo JP
5. Xu Y
6. Lu B
7. Yang Z
8. Wei Y
9. Polyanskaya S
10. Wunderlich M
11. Gray NS
12. Stegmaier K
13. Vakoc CR
(2020) Salt-inducible kinase inhibition suppresses acute myeloid leukemia progression in vivo
Blood 135:56–70.

https://doi.org/10.1182/blood.2019001576
- PubMed
- Google Scholar
1. Tung NM
2. Robson ME
3. Ventz S
4. Santa-Maria CA
5. Nanda R
6. Marcom PK
7. Shah PD
8. Ballinger TJ
9. Yang ES
10. Vinayak S
11. Melisko M
12. Brufsky A
13. DeMeo M
14. Jenkins C
15. Domchek S
16. D’Andrea A
17. Lin NU
18. Hughes ME
19. Carey LA
20. Wagle N
21. Wulf GM
22. Krop IE
23. Wolff AC
24. Winer EP
25. Garber JE
(2020) TBCRC 048: phase II study of olaparib for metastatic breast cancer and mutations in homologous recombination-related genes
Journal of Clinical Oncology 38:4274–4282.

https://doi.org/10.1200/JCO.20.02151
- PubMed
- Google Scholar
1. Tutt ANJ
2. Garber JE
3. Kaufman B
4. Viale G
5. Fumagalli D
6. Rastogi P
7. Gelber RD
8. de Azambuja E
9. Fielding A
10. Balmaña J
11. Domchek SM
12. Gelmon KA
13. Hollingsworth SJ
14. Korde LA
15. Linderholm B
16. Bandos H
17. Senkus E
18. Suga JM
19. Shao Z
20. Pippas AW
21. Nowecki Z
22. Huzarski T
23. Ganz PA
24. Lucas PC
25. Baker N
26. Loibl S
27. McConnell R
28. Piccart M
29. Schmutzler R
30. Steger GG
31. Costantino JP
32. Arahmani A
33. Wolmark N
34. McFadden E
35. Karantza V
36. Lakhani SR
37. Yothers G
38. Campbell C
39. Geyer CE
(2021) Adjuvant olaparib for patients with BRCA1 - or BRCA2 -mutated breast cancer
The New England Journal of Medicine 384:2394–2405.

https://doi.org/10.1056/NEJMoa2105215
- Google Scholar
1. Wilcox EH
2. Sarmady M
3. Wulf B
4. Wright MW
5. Rehm HL
6. Biesecker LG
7. Abou Tayoun AN
(2022) Evaluating the impact of in silico predictors on clinical variant classification
Genetics in Medicine 24:924–930.

https://doi.org/10.1016/j.gim.2021.11.018
- PubMed
- Google Scholar
(2014) PEAR: a fast and accurate illumina paired-end reAd mergeR
Bioinformatics 30:614–620.

https://doi.org/10.1093/bioinformatics/btt593
- PubMed
- Google Scholar
1. Zhen DB
2. Rabe KG
3. Gallinger S
4. Syngal S
5. Schwartz AG
6. Goggins MG
7. Hruban RH
8. Cote ML
9. McWilliams RR
10. Roberts NJ
11. Cannon-Albright LA
12. Li D
13. Moyes K
14. Wenstrup RJ
15. Hartman A-R
16. Seminara D
17. Klein AP
18. Petersen GM
(2015) BRCA1, BRCA2, PALB2, and CDKN2A mutations in familial pancreatic cancer: A PACGENE study
Genetics in Medicine 17:569–577.

https://doi.org/10.1038/gim.2014.153
- PubMed
- Google Scholar

Article and author information

Author details

Hirokazu Kimura

Department of Pathology, the Johns Hopkins University School of Medicine, Baltimore, United States

Contribution
Conceptualization, Data curation, Formal analysis, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing

Competing interests
No competing interests declared
Kamel Lahouel
1. Division of Integrated Genomics, Translational Genomics Research Institute, Phoenix, United States
2. Department of Computational and Quantitative Medicine, Beckman Research Institute, City of Hope, Duarte, United States
Contribution
Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing – review and editing

Competing interests
No competing interests declared
Cristian Tomasetti
1. Division of Integrated Genomics, Translational Genomics Research Institute, Phoenix, United States
2. Department of Computational and Quantitative Medicine, Beckman Research Institute, City of Hope, Duarte, United States
Contribution
Conceptualization, Resources, Data curation, Formal analysis, Investigation, Methodology, Writing – original draft, Writing – review and editing

Competing interests
Receives royalties from Exact Science; is a founder of C2T, a scientific advisor for PrognomiQ, and a consultant for Bayer
Nicholas Jason Roberts
1. Department of Pathology, the Johns Hopkins University School of Medicine, Baltimore, United States
2. Department of Oncology, the Johns Hopkins University School of Medicine, Baltimore, United States
Contribution
Conceptualization, Resources, Data curation, Formal analysis, Investigation, Methodology, Writing – original draft, Project administration, Writing – review and editing

For correspondence
nrobert8@jhmi.edu

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0002-8709-0664

Funding

National Cancer Institute (P50CA62924)

Nicholas Jason Roberts

Susan Wojcicki and Dennis Troper

Nicholas Jason Roberts

The Sol Goldman Pancreatic Cancer Research Center

Nicholas Jason Roberts

Rolfe Pancreatic Cancer Foundation

Nicholas Jason Roberts

Japanese Society of Gastroenterology (Support for Young Gastroenterologists Studying in the United States)

Hirokazu Kimura

Japan Society for the Promotion of Science (Overseas Research Fellowships)

Hirokazu Kimura

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

National Institutes of Health grant P50CA62924 (NJR). Susan Wojcicki and Dennis Troper (NJR). The Sol Goldman Pancreatic Cancer Research Center (NJR). The Rolfe Pancreatic Cancer Foundation (NJR). The Japanese Society of Gastroenterology Support for Young Gastroenterologists Studying in the United States (HK). The Japan Society for the Promotion of Science Overseas Research Fellowships (HK).

Version history

Preprint posted: December 28, 2023
Sent for peer review: December 28, 2023
Reviewed Preprint version 1: February 19, 2024
Reviewed Preprint version 2: November 27, 2024
Reviewed Preprint version 3: April 3, 2025
Version of Record published: April 16, 2025

Cite all versions

You can cite all versions using the DOI https://doi.org/10.7554/eLife.95347. This DOI represents all versions, and will always resolve to the latest one.

Copyright

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.