Use caseDNA and genomesResearch and clinical research

Assess somatic small-variant oncogenicity

Which methods help classify somatic SNVs and small indels while preserving uncertainty and the relevant gene mechanism?

What the evidence shows

Two predictor comparisons join the existing OncoVI evidence; their values cannot be pooled.

In Chen et al. 2020 (peer reviewed), 33 predictors scored somatic missense mutations labelled by OncoKB curation or by the authors' own cell viability assays, with likely neutral or assay-neutral somatic negatives. The authors co-developed one scored tool, CanDrA, and developed none of the others. With each tool's own default calls, the highest accuracy was 0.70 on OncoKB labels (DEOGEN2, sensitivity 0.87, specificity 0.51), and several tools called most neutral mutations damaging, for example M-CAP with sensitivity 0.97 and specificity 0.11, and CanDrA with specificity 0.17 and the lowest accuracy, 0.49.

On the assay labels the highest default accuracy was 0.65 (PrimateAI). When each tool's median score on the test draw was used as the threshold, which ranks tools but is not an operating point a reviewer could apply, the highest accuracy was 0.69 on OncoKB labels (PROVEAN and VEST4) and 0.70 on assay labels (CTAT-cancer and PrimateAI), and FATHMM-disease scored 0.50 on OncoKB labels. Chen et al. print each value as a mean with a range of two standard deviations over 100 draws, not a confidence interval, and do not say which OncoKB positive set fed these tables.

In a single-author bioRxiv preprint that is not peer reviewed (Lee 2026), 49 scores were compared on Cancer Gene Census missense variants whose negatives are germline-benign labels (ClinVar benign or likely benign, and gnomAD common variants), which this use case does not accept as somatic controls. In that table 42 of the 49 scores had a lower AUROC on oncogene variants than on tumour-suppressor variants, for example AlphaMissense 0.885 against 0.900 and CADD 0.821 against 0.894, while popEVE went the other way (0.900 against 0.842). The highest all-gene AUROC was 0.876 (MutScore); these AUROC values are rounded here to three decimals from the table's full precision.

None of these comparisons covers small indels, calibrated probabilities, criterion-level evidence strength, drug response or reviewer agreement.

You bring: Variant, genome build, transcript, allele fraction, quality and confidence in somatic origin
Tumour context, gene mechanism and dated oncogenicity-specific assertions
Functional studies and the applicable criteria specification
You want: An oncogenicity classification or unresolved assessment with constituent criteria, conflicting findings and evidence needs.
Evidence in this release: 7 comparisons shown · 47 tools · 84 configurations
Kept up to date: Evidence last checked 14 October 2026 · re-checked monthly, next due 14 November 2026

Baseline

ClinGen/CGC/VICC rules (OncoVI) and SIFT. Somatic oncogenicity is classified with the ClinGen/CGC/VICC points-based standard, which OncoVI applies as fixed rules. For the computational evidence, SIFT scores substitutions from sequence homology with a position-specific scoring matrix and is widely used to prioritise cancer mutations without a trained classifier.

Machine-learning methods are compared with the baseline on the same data in 3 tables: OncoKB-curated drivers versus likely neutral (Chen et al. 2020), Cell viability assay drivers versus neutral (Chen et al. 2020), Predictors by gene mechanism on CGC missense variants (Lee 2026 preprint). Baseline rows are marked in each table.

Tools compared

AlphaMissense, machine learning
BayesDel, machine learning, 2 configurations
bStatistic, specialist model
CADD, machine learning, 2 configurations
CanDrA, machine learning
CHASM, machine learning
ClinPred, machine learning
CTAT (cancer and population ensembles), specialist model, 2 configurations
DANN, machine learning, 2 configurations
DEOGEN2, machine learning, 2 configurations
Eigen, specialist model, 4 configurations
ESM-1b, foundation model
FATHMM, specialist model, 2 configurations
FATHMM-MKL, machine learning
FATHMM-XF, machine learning, 2 configurations
GenoCanyon, specialist model
GERP, specialist model, 2 configurations
gMVP, machine learning
Historical Erlangen MTB protein-function interpretation, conventional
Integrated_fitCons, specialist model
LIST-S2, specialist model
LRT, specialist model
M-CAP, machine learning, 2 configurations
MetaLR, machine learning, 2 configurations
MetaRNN, machine learning
MetaSVM, machine learning, 2 configurations
MisFit, specialist model, 2 configurations
MPC, machine learning, 2 configurations
MutationAssessor, specialist model, 2 configurations
MutationTaster, machine learning, 2 configurations
MutFormer, machine learning
MutPred, machine learning, 2 configurations
MutScore, machine learning
MVP, machine learning, 2 configurations
Oncogenicity Variant Interpreter (OncoVI)Baseline, conventional
PHACTboost, machine learning
phastCons, specialist model, 3 configurations
phyloP, specialist model, 3 configurations
PolyPhen-2, machine learning, 4 configurations
popEVE, specialist model
PrimateAI, machine learning, 2 configurations
PROVEAN, conventional, 2 configurations
REVEL, machine learning, 2 configurations
SIFTBaseline, conventional, 4 configurations
TransFIC, specialist model
VARITY, machine learning, 4 configurations
VEST4, machine learning, 2 configurations

Comparisons

Each table is one study or protocol with its own truth set and scoring. Compare tools within a table, not across tables.

Rows marked Contested rest on results that other studies disagree with. The results are still shown. See the disagreements that involve this use case for what disagrees and what we are doing about it.

OncoVI three-class assessment: MTB selected expert reassessment

Proxy evidenceAuthor-reported

Oncogenicity category agreement, sensitivity and reported subgroup performance

Sorted by Percent agreement (higher is better). The best value in each column is highlighted.

Scroll sideways to see all 4 columns.

Tool	Percent agreement (prior mtb classification vs expert reassessment)	Percent agreement (vs expert reassessment)	Correlation (correlation type not stated)	Count (expert-concordant variants)
OncoVI 2026 publication configurationBaselineconventional	22%	80% (best in column)	0.7	Not reported
Historical Erlangen MTB protein-function interpretationconventional	Not reported	34%	Not reported	46

Caveats and method

Not a treatment-selection benchmark.
No human domain review of this intake; reference overlap and prospective workload remain unresolved.
Abstract/body accuracy discrepancies retained.

Protocol: OncoVI three-class assessment: MTB selected expert reassessment

OncoKB-curated drivers versus likely neutral (Chen et al. 2020)

Proxy evidenceIndependent studyAuthor-reported

Accuracy (higher is better), with the other metrics below each value. Showing the top 10 of 33 rows. The best value in each column is highlighted.

Scroll sideways to see all 2 columns.

Tool	Default categories	Median-score threshold
DEOGEN2machine learning	70% (best in column)Negative predictive value 0.79P 0.65R 87%Sp 0.51	66%Negative predictive value 0.66P 0.66R 66%Sp 0.66
PROVEANconventional	67%Negative predictive value 0.80P 0.61R 89%Sp 0.45	69% (best in column)Negative predictive value 0.70P 0.69R 70%Sp 0.68
FATHMM-XF Contested: other studies disagree with this study's ranking of this toolmachine learning	66%Negative predictive value 0.85P 0.60R 93%Sp 0.39	65%Negative predictive value 0.65P 0.65R 65%Sp 0.65
LRT Contested (2): other studies disagree with this study's ranking of this toolspecialist model	66%Negative predictive value 0.86P 0.60R 93%Sp 0.38	65%Negative predictive value 0.67P 0.63R 73%Sp 0.57
PrimateAImachine learning	66%Negative predictive value 0.71P 0.63R 77%Sp 0.55	66%Negative predictive value 0.66P 0.66R 66%Sp 0.66
Polyphen2_HVARmachine learning	64%Negative predictive value 0.76P 0.60R 88%Sp 0.39	62%Negative predictive value 0.62P 0.62R 63%Sp 0.61
Polyphen2_HDIVmachine learning	62%Negative predictive value 0.79P 0.58R 93%Sp 0.29	63%Negative predictive value 0.63P 0.62R 64%Sp 0.62
SIFTBaselineContested: other studies disagree with this study's ranking of this toolconventional	61%Negative predictive value 0.75P 0.57R 89%Sp 0.33	61%Negative predictive value 0.61P 0.60R 63%Sp 0.58
SIFT4GBaselineconventional	61%Negative predictive value 0.71P 0.58R 83%Sp 0.41	61%Negative predictive value 0.61P 0.61R 62%Sp 0.60
FATHMM-MKLmachine learning	60%Negative predictive value 0.93P 0.56R 98%Sp 0.21	63%Negative predictive value 0.63P 0.63R 63%Sp 0.63
MetaSVM Contested (2): other studies disagree with this study's ranking of this toolmachine learning	60%Negative predictive value 0.62P 0.59R 68%Sp 0.52	60%Negative predictive value 0.60P 0.60R 60%Sp 0.60
MutationTaster2machine learning	60%Negative predictive value 0.93P 0.56R 98%Sp 0.22	63%Negative predictive value 0.73P 0.59R 85%Sp 0.41
MutationAssessorspecialist model	58%Negative predictive value 0.59P 0.57R 66%Sp 0.50	59%Negative predictive value 0.59P 0.59R 61%Sp 0.57
MetaLR Contested: other studies disagree with this study's ranking of this toolmachine learning	57%Negative predictive value 0.59P 0.56R 67%Sp 0.47	57%Negative predictive value 0.57P 0.57R 57%Sp 0.57
M-CAPmachine learning	55%Negative predictive value 0.79P 0.53R 97%Sp 0.11	56%Negative predictive value 0.56P 0.56R 56%Sp 0.56
FATHMM-diseasespecialist model	52%Negative predictive value 0.53P 0.51R 63%Sp 0.42	50%Negative predictive value 0.50P 0.50R 50%Sp 0.50
CanDrAmachine learning	49%Negative predictive value 0.46P 0.49R 80%Sp 0.17	59%Negative predictive value 0.59P 0.59R 59%Sp 0.59
CADD Contested: other studies disagree with this study's ranking of this toolmachine learning	Not reported	64%Negative predictive value 0.64P 0.64R 64%Sp 0.64
CHASMmachine learning	Not reported	65%Negative predictive value 0.65P 0.65R 65%Sp 0.64
CTAT-cancerspecialist model	Not reported	63%Negative predictive value 0.63P 0.63R 63%Sp 0.63
CTAT-populationspecialist model	Not reported	66%Negative predictive value 0.66P 0.66R 66%Sp 0.66
DANNmachine learning	Not reported	59%Negative predictive value 0.59P 0.59R 59%Sp 0.59
Eigenspecialist model	Not reported	66%Negative predictive value 0.66P 0.66R 66%Sp 0.66
Eigen-PCspecialist model	Not reported	66%Negative predictive value 0.66P 0.66R 66%Sp 0.66
FATHMM-cancerspecialist model	Not reported	67%Negative predictive value 0.67P 0.67R 67%Sp 0.67
GenoCanyon Contested (2): other studies disagree with this study's ranking of this toolspecialist model	Not reported	64%Negative predictive value 0.72P 0.60R 82%Sp 0.47
Integrated_fitConsspecialist model	Not reported	51%Negative predictive value 0.51P 0.51R 56%Sp 0.46
MPCmachine learning	Not reported	68%Negative predictive value 0.68P 0.68R 68%Sp 0.68
MutPredmachine learning	Not reported	53%Negative predictive value 0.54P 0.52R 69%Sp 0.36
MVPmachine learning	Not reported	65%Negative predictive value 0.65P 0.65R 65%Sp 0.65
REVELmachine learning	Not reported	67%Negative predictive value 0.67P 0.67R 67%Sp 0.67
TransFICspecialist model	Not reported	59%Negative predictive value 0.59P 0.59R 59%Sp 0.59
VEST4machine learning	Not reported	69% (best in column)Negative predictive value 0.69P 0.69R 69%Sp 0.69

Show all 33 rowsShow the top 10 rows only

R recall, P precision, Sp specificity.

Caveats and method

What each column measures:

Default categories: Accuracy, sensitivity, specificity, PPV and NPV for 17 predictors on OncoKB benchmark (positive set not stated) using default categorical calls
Median-score threshold: Accuracy, sensitivity, specificity, PPV and NPV for 33 predictors on OncoKB benchmark (positive set not stated) using median-score threshold

Binary metrics are means over 100 random draws of 400 positives and 400 negatives; the bracketed range is the mean plus or minus two standard deviations, not a confidence interval.
Predictor scores date from dbNSFP v4.0 and 2019 web servers; current releases (for example AlphaMissense, CHASMplus, BoostDM) are not included.
Only 17 algorithms provide default categories, mostly germline deleteriousness tools; cancer-specific tools other than CanDrA and FATHMM-disease are absent.
OncoKB annotations are biased towards known, often actionable cancer genes (Discussion).
Some predictors were trained on known cancer mutations, so performance on literature-curated drivers may be inflated.
The OncoKB set has 421 or 497 likely neutral negatives (Methods and Results disagree) and 400 are drawn per repeat, so the draws overlap heavily and the spread understates sampling uncertainty. Results and Methods also disagree on the positive counts, and which positive set fed the binary metrics is not stated.
Two authors (H. Liang, G.B. Mills) co-developed CanDrA (reference 7), and the cell viability labels come from the same group's assays; CanDrA rows are author_reported.
No linked comparison reports calibrated probabilities, per-criterion evidence strength or indels; all variants are missense SNVs.
The threshold is each algorithm's median score on the benchmark draw, which is not available when classifying a new variant; with balanced draws it makes sensitivity, specificity, PPV and NPV nearly equal, so the table ranks tools rather than measuring a usable operating point.

Protocols: OncoKB benchmark (positive set not stated), default categorical calls (Chen et al. 2020 Additional file 10), OncoKB benchmark (positive set not stated), median-score threshold (Chen et al. 2020 Additional file 9)

Cell viability assay drivers versus neutral (Chen et al. 2020)

Proxy evidenceIndependent studyAuthor-reported

Accuracy (higher is better), with the other metrics below each value. Showing the top 10 of 33 rows. The best value in each column is highlighted.

Scroll sideways to see all 2 columns.

Tool	Default categories	Median-score threshold
PrimateAImachine learning	65% (best in column)Negative predictive value 0.70P 0.63R 79%Sp 0.52	70% (best in column)Negative predictive value 0.70P 0.70R 70%Sp 0.70
DEOGEN2machine learning	62%Negative predictive value 0.72P 0.58R 85%Sp 0.38	64%Negative predictive value 0.64P 0.64R 64%Sp 0.64
PROVEANconventional	62%Negative predictive value 0.70P 0.57R 78%Sp 0.48	59%Negative predictive value 0.59P 0.58R 59%Sp 0.58
MutationAssessorspecialist model	60%Negative predictive value 0.59P 0.60R 60%Sp 0.60	60%Negative predictive value 0.60P 0.60R 60%Sp 0.60
CanDrAmachine learning	59%Negative predictive value 0.69P 0.56R 87%Sp 0.30	67%Negative predictive value 0.67P 0.67R 67%Sp 0.67
FATHMM-XF Contested: other studies disagree with this study's ranking of this toolmachine learning	59%Negative predictive value 0.72P 0.56R 89%Sp 0.28	63%Negative predictive value 0.63P 0.63R 63%Sp 0.63
Polyphen2_HDIVmachine learning	59%Negative predictive value 0.70P 0.56R 88%Sp 0.29	58%Negative predictive value 0.58P 0.58R 58%Sp 0.57
Polyphen2_HVARmachine learning	59%Negative predictive value 0.65P 0.57R 80%Sp 0.38	59%Negative predictive value 0.59P 0.59R 59%Sp 0.59
LRT Contested (2): other studies disagree with this study's ranking of this toolspecialist model	58%Negative predictive value 0.71P 0.56R 93%Sp 0.20	65%Negative predictive value 0.68P 0.63R 74%Sp 0.56
SIFT4GBaselineconventional	58%Negative predictive value 0.63P 0.54R 69%Sp 0.48	60%Negative predictive value 0.61P 0.58R 67%Sp 0.52
M-CAPmachine learning	57%Negative predictive value 0.80P 0.55R 96%Sp 0.18	58%Negative predictive value 0.58P 0.58R 58%Sp 0.58
MetaLR Contested: other studies disagree with this study's ranking of this toolmachine learning	56%Negative predictive value 0.55P 0.56R 53%Sp 0.59	56%Negative predictive value 0.56P 0.56R 56%Sp 0.56
MetaSVM Contested (2): other studies disagree with this study's ranking of this toolmachine learning	56%Negative predictive value 0.55P 0.57R 51%Sp 0.61	55%Negative predictive value 0.55P 0.55R 55%Sp 0.55
SIFTBaselineContested: other studies disagree with this study's ranking of this toolconventional	56%Negative predictive value 0.66P 0.53R 80%Sp 0.35	56%Negative predictive value 0.56P 0.55R 59%Sp 0.53
MutationTaster2machine learning	55%Negative predictive value 0.79P 0.53R 97%Sp 0.12	61%Negative predictive value 0.74P 0.57R 87%Sp 0.36
FATHMM-MKLmachine learning	53%Negative predictive value 0.71P 0.52R 97%Sp 0.08	62%Negative predictive value 0.62P 0.62R 62%Sp 0.62
FATHMM-diseasespecialist model	50%Negative predictive value 0.52P 0.47R 37%Sp 0.62	46%Negative predictive value 0.46P 0.46R 46%Sp 0.46
CADD Contested: other studies disagree with this study's ranking of this toolmachine learning	Not reported	59%Negative predictive value 0.59P 0.59R 59%Sp 0.59
CHASMmachine learning	Not reported	69%Negative predictive value 0.69P 0.69R 69%Sp 0.69
CTAT-cancerspecialist model	Not reported	70% (best in column)Negative predictive value 0.70P 0.70R 70%Sp 0.70
CTAT-populationspecialist model	Not reported	64%Negative predictive value 0.64P 0.64R 64%Sp 0.64
DANNmachine learning	Not reported	58%Negative predictive value 0.58P 0.58R 58%Sp 0.58
Eigenspecialist model	Not reported	62%Negative predictive value 0.62P 0.62R 62%Sp 0.62
Eigen-PCspecialist model	Not reported	63%Negative predictive value 0.63P 0.63R 63%Sp 0.63
FATHMM-cancerspecialist model	Not reported	66%Negative predictive value 0.66P 0.65R 66%Sp 0.65
GenoCanyon Contested (2): other studies disagree with this study's ranking of this toolspecialist model	Not reported	58%Negative predictive value 0.66P 0.55R 84%Sp 0.31
Integrated_fitConsspecialist model	Not reported	49%Negative predictive value 0.49P 0.49R 53%Sp 0.46
MPCmachine learning	Not reported	66%Negative predictive value 0.66P 0.66R 66%Sp 0.66
MutPredmachine learning	Not reported	66%Negative predictive value 0.66P 0.66R 66%Sp 0.66
MVPmachine learning	Not reported	64%Negative predictive value 0.64P 0.64R 64%Sp 0.64
REVELmachine learning	Not reported	64%Negative predictive value 0.64P 0.64R 65%Sp 0.64
TransFICspecialist model	Not reported	61%Negative predictive value 0.61P 0.61R 61%Sp 0.61
VEST4machine learning	Not reported	62%Negative predictive value 0.62P 0.62R 62%Sp 0.62

Show all 33 rowsShow the top 10 rows only

R recall, P precision, Sp specificity.

Caveats and method

What each column measures:

Default categories: Accuracy, sensitivity, specificity, PPV and NPV for 17 predictors on cell viability drivers versus neutral using default categorical calls
Median-score threshold: Accuracy, sensitivity, specificity, PPV and NPV for 33 predictors on cell viability drivers versus neutral using median-score threshold

Binary metrics are means over 100 random draws of 400 positives and 400 negatives; the bracketed range is the mean plus or minus two standard deviations, not a confidence interval.
Predictor scores date from dbNSFP v4.0 and 2019 web servers; current releases (for example AlphaMissense, CHASMplus, BoostDM) are not included.
Only 17 algorithms provide default categories, mostly germline deleteriousness tools; cancer-specific tools other than CanDrA and FATHMM-disease are absent.
Functional calls come from the authors' own Ba/F3 and MCF10A screens; a growth effect in either cell model counts as positive.
Activating and inactivating variants are pooled, so oncogene and tumour-suppressor mechanisms are not separated.
The combined cell viability set has 376 positives (Results), yet Methods say 400 positives were drawn per repeat, so positives were drawn with replacement or the counts differ; the source does not say.
Two authors (H. Liang, G.B. Mills) co-developed CanDrA (reference 7), and the cell viability labels come from the same group's assays; CanDrA rows are author_reported.
No linked comparison reports calibrated probabilities, per-criterion evidence strength or indels; all variants are missense SNVs.
The threshold is each algorithm's median score on the benchmark draw, which is not available when classifying a new variant; with balanced draws it makes sensitivity, specificity, PPV and NPV nearly equal, so the table ranks tools rather than measuring a usable operating point.

Protocols: Cell viability drivers versus neutral, default categorical calls (Chen et al. 2020 Additional file 22), Cell viability drivers versus neutral, median-score threshold (Chen et al. 2020 Additional file 21)

Predictors by gene mechanism on CGC missense variants (Lee 2026 preprint)

Proxy evidenceIndependent study

AUROC (higher is better), with the other metrics below each value. Showing the top 10 of 49 rows. The best value in each column is highlighted.

Scroll sideways to see all 3 columns.

Tool	All genes	Oncogenes (GOF)	Tumour suppressors (LOF)
MutScoremachine learning	0.876 (best in column)Average precision 0.754	0.877Average precision 0.809	0.898Average precision 0.720
VARITY_Rmachine learning	0.874Average precision 0.745	0.894Average precision 0.804	0.894Average precision 0.690
VARITY_R_LOOmachine learning	0.873Average precision 0.739	0.891Average precision 0.798	0.893Average precision 0.683
ClinPredmachine learning	0.869Average precision 0.741	0.888Average precision 0.789	0.898Average precision 0.717
AlphaMissensemachine learning	0.868Average precision 0.750	0.885Average precision 0.804	0.900Average precision 0.749
VARITY_ERmachine learning	0.867Average precision 0.717	0.886Average precision 0.780	0.886Average precision 0.658
VARITY_ER_LOOmachine learning	0.866Average precision 0.714	0.884Average precision 0.775	0.886Average precision 0.656
BayesDel_addAF Contested: other studies disagree with this study's ranking of this toolmachine learning	0.864Average precision 0.712	0.878Average precision 0.764	0.892Average precision 0.684
popEVEspecialist model	0.860Average precision 0.717	0.900 (best in column)Average precision 0.834	0.842Average precision 0.614
MetaRNNmachine learning	0.853Average precision 0.733	0.887Average precision 0.793	0.889Average precision 0.715
gMVPmachine learning	0.850Average precision 0.675	0.860Average precision 0.739	0.913 (best in column)Average precision 0.731
MutationTastermachine learning	0.847Average precision 0.676	0.868Average precision 0.779	0.897Average precision 0.616
MutPred2machine learning	0.846Average precision 0.664	0.876Average precision 0.773	0.896Average precision 0.661
ESM1bfoundation model	0.842Average precision 0.641	0.841Average precision 0.719	0.858Average precision 0.594
MisFit_Dspecialist model	0.839Average precision 0.657	0.846Average precision 0.723	0.883Average precision 0.662
VEST4machine learning	0.833Average precision 0.648	0.840Average precision 0.713	0.885Average precision 0.626
PHACTboostmachine learning	0.833Average precision 0.693	0.850Average precision 0.745	0.889Average precision 0.707
Polyphen2_HVARmachine learning	0.826Average precision 0.550	0.820Average precision 0.602	0.852Average precision 0.482
CADD_rawmachine learning	0.826Average precision 0.596	0.821Average precision 0.642	0.894Average precision 0.646
BayesDel_noAFmachine learning	0.825Average precision 0.649	0.827Average precision 0.688	0.867Average precision 0.630
MPCmachine learning	0.823Average precision 0.613	0.842Average precision 0.716	0.873Average precision 0.626
PROVEANconventional	0.819Average precision 0.588	0.813Average precision 0.646	0.859Average precision 0.548
REVEL Contested: other studies disagree with this study's ranking of this toolmachine learning	0.816Average precision 0.652	0.843Average precision 0.739	0.859Average precision 0.618
DEOGEN2machine learning	0.815Average precision 0.653	0.858Average precision 0.732	0.841Average precision 0.596
Eigen-raw_codingspecialist model	0.814Average precision 0.589	0.809Average precision 0.633	0.874Average precision 0.606
Polyphen2_HDIVmachine learning	0.814Average precision 0.470	0.805Average precision 0.541	0.847Average precision 0.409
fathmm-XF_codingmachine learning	0.813Average precision 0.568	0.810Average precision 0.644	0.857Average precision 0.476
SIFT4GBaselineconventional	0.808Average precision 0.541	0.786Average precision 0.572	0.847Average precision 0.471
MVPmachine learning	0.807Average precision 0.570	0.815Average precision 0.657	0.803Average precision 0.502
Eigen-PC-raw_codingspecialist model	0.799Average precision 0.557	0.795Average precision 0.607	0.865Average precision 0.574
PrimateAImachine learning	0.797Average precision 0.508	0.817Average precision 0.618	0.872Average precision 0.505
MutationAssessorspecialist model	0.794Average precision 0.554	0.769Average precision 0.604	0.824Average precision 0.535
SIFTBaselineconventional	0.778Average precision 0.442	0.764Average precision 0.520	0.817Average precision 0.383
LIST-S2specialist model	0.768Average precision 0.521	0.756Average precision 0.562	0.832Average precision 0.536
phyloP100way_vertebratespecialist model	0.764Average precision 0.470	0.766Average precision 0.567	0.818Average precision 0.422
DANNmachine learning	0.762Average precision 0.389	0.749Average precision 0.468	0.823Average precision 0.378
MetaSVMmachine learning	0.762Average precision 0.507	0.766Average precision 0.563	0.794Average precision 0.500
M-CAPmachine learning	0.758Average precision 0.481	0.775Average precision 0.544	0.803Average precision 0.478
MetaLRmachine learning	0.748Average precision 0.544	0.750Average precision 0.543	0.788Average precision 0.522
phyloP470way_mammalianspecialist model	0.729Average precision 0.347	0.714Average precision 0.437	0.786Average precision 0.310
MisFit_Sspecialist model	0.729Average precision 0.549	0.803Average precision 0.695	0.846Average precision 0.635
MutFormermachine learning	0.693Average precision 0.446	0.768Average precision 0.582	0.655Average precision 0.286
GERP_92_mammalsspecialist model	0.691Average precision 0.260	0.662Average precision 0.333	0.742Average precision 0.198
GERPpp_RSspecialist model	0.687Average precision 0.285	0.658Average precision 0.349	0.757Average precision 0.253
phastCons100way_vertebratespecialist model	0.678Average precision 0.268	0.671Average precision 0.357	0.720Average precision 0.206
phastCons470way_mammalianspecialist model	0.648Average precision 0.244	0.636Average precision 0.330	0.693Average precision 0.190
phastCons17way_primatespecialist model	0.630Average precision 0.265	0.680Average precision 0.405	0.667Average precision 0.204
phyloP17way_primatespecialist model	0.588Average precision 0.211	0.592Average precision 0.294	0.610Average precision 0.152
bStatisticspecialist model	0.438Average precision 0.160	0.425Average precision 0.245	0.384Average precision 0.0951

Show all 49 rowsShow the top 10 rows only

Caveats and method

What each column measures:

All genes: AUROC and average precision (labelled AUPRC) of 49 predictor and conservation scores on CGC missense variants, all genes
Oncogenes (GOF): AUROC and average precision (labelled AUPRC) of 49 predictor and conservation scores on CGC missense variants, oncogenes (gain of function)
Tumour suppressors (LOF): AUROC and average precision (labelled AUPRC) of 49 predictor and conservation scores on CGC missense variants, tumour suppressors (loss of function)

Negatives are ClinVar benign or likely benign and gnomAD common variants, that is germline-benign labels; the use case does not accept these as automatic somatic negatives.
Many supervised tools were trained on ClinVar labels, so the ClinVar-benign negatives overlap their training data.
Single-author bioRxiv preprint (v1, posted 2026-07-23), not peer reviewed; the analysis code is in the linked repository.
Scores are dbNSFP 5.3.1a rank scores; tool coverage differs, so each row has its own n and positives.
No uncertainty is printed in this table. The preprint Results give VARITY_R 0.874 and AlphaMissense 0.868, matching the table; the OncoCal section gives 0.871 and 0.870, probably on the unique protein-level set used to train OncoCal.
The table cannot be regenerated from the repository alone: its script (src/evaluation/part_b.py) needs dbNSFP 5.3.1a and COSMIC v104, which are not redistributed. It was added in the first commit (b17e658d6b) and is unchanged at the pinned commit, whose message credits an AI coding assistant but does not touch the table or its script.
No linked comparison reports calibrated probabilities, per-criterion evidence strength or indels; all variants are missense SNVs.
Gene role uses CGC gene-level roles where variant-level CIViC mechanism is missing, so some variants may be assigned the wrong mechanism.

Protocols: CGC missense oncogenicity, all genes (Lee 2026, OncoCal tool_performance.tsv), CGC missense oncogenicity, oncogenes (gain of function) (Lee 2026, OncoCal tool_performance.tsv), CGC missense oncogenicity, tumour suppressors (loss of function) (Lee 2026, OncoCal tool_performance.tsv)

OncoVI three-class assessment: Erlangen MTB variants

Proxy evidenceAuthor-reported

Oncogenicity category agreement, sensitivity and reported subgroup performance

Single reported result.

OncoVI 2026 publication configurationBaselineconventional

Percent agreement: 78%
Percent agreement (class: benign): 43%
Percent agreement (class: VUS): 97%
Percent agreement (gene role: oncogene): 82.5%
Percent agreement (gene role: tumour suppressor): 76.0%
Percent agreement (oncogenic vs pathogenic): 79%
Percent agreement (variant type: missense): 78.2%
Percent agreement (variant type: truncating): 49.4%

Caveats and method

Not a treatment-selection benchmark.
No human domain review of this intake; reference overlap and prospective workload remain unresolved.
Abstract/body accuracy discrepancies retained.

Protocol: OncoVI three-class assessment: Erlangen MTB variants

OncoVI three-class assessment: ClinVar April 2025 oncogenicity assertions

Direct evidenceAuthor-reported

Oncogenicity category agreement, sensitivity and reported subgroup performance

Single reported result.

OncoVI 2026 publication configurationBaselineconventional

Percent agreement (gene role: oncogene): 76.1%
Percent agreement (gene role: tumour suppressor): 85.7%
Percent agreement (variant type: missense): 81.1%
Percent agreement (variant type: truncating): 87.5%
Accuracy: 83%
Recall (class: oncogenic): 84%

Caveats and method

Not a treatment-selection benchmark.
No human domain review of this intake; reference overlap and prospective workload remain unresolved.
Abstract/body accuracy discrepancies retained.

Protocol: OncoVI three-class assessment: ClinVar April 2025 oncogenicity assertions

OncoVI three-class assessment: Guideline SOP variants

Direct evidenceAuthor-reported

Oncogenicity category agreement, sensitivity and reported subgroup performance

Single reported result.

OncoVI 2026 publication configurationBaselineconventional

Accuracy: 81%
Accuracy (gene role: oncogene): 82.0%
Accuracy (gene role: tumour suppressor): 87.1%
Accuracy (variant type: missense): 83.1%
Accuracy (variant type: truncating): 77.7%
Correlation (correlation type not stated): 0.87
Recall (class: benign): 83%
Recall (class: oncogenic): 88%

Caveats and method

Not a treatment-selection benchmark.
No human domain review of this intake; reference overlap and prospective workload remain unresolved.
Abstract/body accuracy discrepancies retained.

Protocol: OncoVI three-class assessment: Guideline SOP variants

Not covered yet

OncoVI direct SOP/ClinVar classification exists; prospective reviewer-effort, calibration and independent model-versus-expert evaluation remain missing.
MTB old labels measure protein-function impact, not oncogenicity; selected reassessment shares experts/resources and is disagreement-enriched.
Abstract/body accuracy conflicts remain flagged. Supplemental Table S6 download returned challenge HTML; only body-printed strata are extracted.
Qualified human scientific review remains outstanding; automated source transcription does not establish experimental replication.
No linked independent comparison includes CHASMplus, BoostDM or AlphaMissense on somatic driver labels with somatic or functional negatives; the 2026 preprint that includes AlphaMissense uses germline-benign negatives.
No linked comparison reports calibration of predictor scores to oncogenicity probability or to ClinGen/CGC/VICC SOP evidence strength.
All linked predictor comparisons score missense SNVs; none covers small indels.
Per-gene results (for example TP53 transactivation in Chen et al. 2020 benchmark 3) are not transcribed in this intake.
Binary metrics in Chen et al. 2020 use each tool's median benchmark score as the threshold, which is not available when classifying a new variant.
No machine-learning method has been compared with the OncoVI criteria-based classification on the same data in the linked evidence.

Contribute evidence or propose a correction

Question, sources and review

Who this is for, setting and exclusions

Choose methods that support evidence-based oncogenicity review of somatic small variants and identify when the available evidence remains insufficient.

Who this is for: Somatic variant curators
Molecular pathologists evaluating interpretation methods
Biological setting: SNVs and small indels, with separate protocols for oncogene missense and tumour-suppressor loss-of-function mechanisms.
Clinical research scope: Clinical research on somatic interpretation support. Tumour-only sequencing does not establish somatic origin, and therapeutic relevance needs a separate context-specific review.

Outside this use case

Inherited predisposition, fusions, rearrangements and copy-number variants require separate interpretation protocols.
Oncogenicity and molecular activity do not establish drug response or clinical actionability.
Germline-benign labels are not automatically valid somatic negative controls.

Evidence collection plan

Collecting evidence

Already on record: 7 comparisons shown · 47 tools, in the evaluated evidence above. The status above describes only the specific comparison in this plan, which remains open; it does not mean no evidence has been collected. The plan defines a comparison to investigate; it does not establish model performance or suitability.

Comparison question

Does a proposed method improve independent oncogenicity classification or review efficiency over explicit conventional criteria while retaining calibrated unresolved results?

Baselines, outcomes and validation requirements

Baselines to include

ClinGen/CGC/VICC criteria-based review under a pinned specification
A relevant specialised predictor assessed only within its actual output scope

Outcomes to measure

Category precision/recall and criteria fidelity
Calibration, abstention and unresolved-case coverage
Expert review effort, measured separately from treatment outcomes

Validation requirements

Use oncogenicity-specific labels with evidence, review status and dates.
Hold out allelic series, residues, studies and genes where unseen-gene transfer is intended.
Audit training and assertion overlap, including predictor contributions to labels.
Preserve tumour, assay and gene-mechanism context.
Use independent qualified somatic-variant adjudication and justified negative or uncertain examples.

Next collection task

Close the remaining endpoint, independence and comparator gaps documented in the 30 September 2026 audit and linked article issue; obtain qualified scientific review.

Sources, review and downloads

Automated source review · 2026-09-30 · Codex source curation, independent worker cross-review and root integration review

Source-backed literature curation with independent automated transcription and scope checks. No new model execution, independent experimental replication, qualified human scientific review or clinical validation.

Rewire clinical use-case priorities: workflow definitions and evidence plans · Original source ↗C4 — Somatic small-variant oncogenicity
Comprehensive assessment of computational algorithms in predicting cancer driver mutations · Original source ↗Results 'Benchmark 2' and 'Benchmark 5'; Methods 'OncoKB annotation benchmark', 'In vitro cell viability assay benchmark' and 'Calculation of five evaluation metrics based on categorical predictions'; reference 7; Competing interests
Chen et al. 2020, Additional file 1: Default prediction categories of 17 algorithms · Original source ↗Default prediction categories of 17 algorithms
Chen et al. 2020, Additional file 9: Performance metrics of 33 algorithms, median-score threshold, benchmark 2 (OncoKB) · Original source ↗Benchmark 2, median-score threshold; see data/omics/use-case-coverage-somatic-oncogenicity-20261009/claims.csv
Chen et al. 2020, Additional file 10: Performance metrics of 17 algorithms, default categories, benchmark 2 (OncoKB) · Original source ↗Benchmark 2, default categories; see data/omics/use-case-coverage-somatic-oncogenicity-20261009/claims.csv
Chen et al. 2020, Additional file 21: Performance metrics of 33 algorithms, median-score threshold, benchmark 5 (cell viability) · Original source ↗Benchmark 5, median-score threshold; see data/omics/use-case-coverage-somatic-oncogenicity-20261009/claims.csv
Chen et al. 2020, Additional file 22: Performance metrics of 17 algorithms, default categories, benchmark 5 (cell viability) · Original source ↗Benchmark 5, default categories; see data/omics/use-case-coverage-somatic-oncogenicity-20261009/claims.csv
An openly licensed benchmark and per-gene calibration map for missense pathogenicity predictors on activating cancer drivers · Original source ↗Methods 'Labels' and 'Features'; Results first section; Competing interests
OncoCal repository, tables/tool_performance.tsv (per-tool AUROC and AUPRC by gene role) · Original source ↗tables/tool_performance.tsv at commit 40b7770f2a768ba800c4f4c6b48e2bfe967ed14a; see data/omics/use-case-coverage-somatic-oncogenicity-20261009/claims.csv

Release 2026-11-02-a7f03c9e55a5 · question use-case-somatic-small-variant-oncogenicity

Use-case input digest 43a80aef5473f1a49c013b6d46b720b850ff5a0d18dcbb741dcdeed1ed082db1

Download questions and review metadata (JSON) (gzip) · Verify release checksums (gzip)