Clinically calibrated missense variant effect prediction that is robust across genes and ancestries
Much of the machinery exists (AlphaMissense, ClinGen calibration), and the remaining work is calibration, circularity control and equity, so difficulty (43) is lower than the structure problems. Application value 66 reflects clinical genetics decisions for millions of patients, broad impact on a deployed practice; leverage 47 because it reshapes guideline-level variant interpretation more than other computational fields.
The problem
Telling whether a single-letter DNA change in a patient is harmful, with probabilities doctors can trust.
- 1Patient has a protein-altering change
- 2Model scores damage
- 3Calibrate score to evidence
- 4Check on held-out experiments
- 5Report equal accuracy for all groups
Clinical context. The ACMG/AMP guidelines classify variants as pathogenic, likely pathogenic, uncertain, likely benign or benign by combining criteria. Computational predictions count as supporting evidence (PP3 for damaging, BP4 for benign). Tavtigian and colleagues expressed the framework as Bayesian: each evidence level corresponds to a likelihood ratio, with 'moderate' at about 4.3 to 1 odds of pathogenicity.
Built from
earlier problems → this oneThe problems this one is built from: progress or tools it needs (builds on), special cases that come first (generalizes), and weaker results on the way (implies). On the right, what builds on it.
- Start with
- Transferable ML potentials with non-local charge transfer, variable charge/spin states and long-range electrostaticsComputational Chemistry · D52
- One classical force field accurate for folded proteins, disordered proteins and weak protein–protein associationComputational Chemistry · D51.5
- Predicting fold-switching proteins and both of their folds from sequenceComputational Medicine · D58
- Step 2
- Boltzmann-weighted conformational ensembles of multi-state proteins from sequenceComputational Medicine · D64
- Step 3
- Prospective protein–ligand binding free energies at 1 kcal/mol RMSE across diverse, difficult targetsComputational Chemistry · D50
- Step 4
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Reverse-mode automatic differentiationAlgorithm
- Fast matrix multiplicationAlgorithm
- Step 2
- The transformer architectureAlgorithm
- Step 3
- Step 4
- Protein language models and generative protein designFrontier · host
- Then
- Clinically calibrated missense variant effect prediction that is robust across genes and ancestriesThe problem
State of the art
- AlphaMissense (Science 2023). Scores all 71 million possible human missense variants and classifies 89% as likely benign or likely pathogenic, trained without human clinical labels. Independent evaluations are mixed: in a 2024 study of 5,845 missense variants in 59 genes, a manual ACMG review yielded 4,085 VUS; AlphaMissense had sensitivity 92% and specificity 78% for pathogenicity, adding it as evidence changed the evidence quantification for 878 variants, 56 of which became likely pathogenic. A calibration study summarised by the Broad Institute found AlphaMissense can reach strong evidence for pathogenicity, though for few variants, and at best modest improvements over previously recommended tools in the trade-off between evidence strength and false positives.
- Calibration framework. Pejaver et al. (AJHG 2022) set ClinGen-endorsed calibrations for PP3/BP4 for several tools; only a limited set reach moderate or strong evidence.
- Zero-shot protein language models. On ProteinGym, a standard suite of deep mutational scanning assays and clinical substitution sets, leading zero-shot models reach only moderate rank correlation with measured effects (the leaderboard is live and no single figure is quoted here); performance degrades for higher-order epistasis and for orphan proteins with shallow alignments.
- Disparity. The burden of VUS differs across groups: a 2025 Genetics in Medicine analysis found the ratio of VUS to pathogenic variants varied about 3-fold by self-reported race and over 14-fold by testing indication. This describes reporting patterns, not predictor error by ancestry, which has not been established.
- Not shown. No predictor has met the calibration, gene-agnostic and ancestry-robust criteria in a held-out evaluation in sources reviewed.
Realistic applications
Reclassifying VUS would change clinical management for cancer predisposition, cardiomyopathy, epilepsy and other monogenic disorders, reduce diagnostic odysseys, and shrink the inequity in how many patients receive a definitive result. It also guides which variants to test experimentally.
What counts as solved
A ClinGen-endorsed (or equivalent) calibration of a single predictor showing at least moderate evidence strength for at least 50% of missense variants in known disease genes at the verified false-assertion rate, validated on post-cut-off expert-reviewed and MAVE data, with ancestry-stratified accuracy reported and within 3 percentage points.
How it gets verified
ReproductionA training run, experiment or computation rerun by someone else.
A single predictor must receive a ClinGen-endorsed (or equivalent) calibration giving at least moderate evidence strength for at least 50% of missense variants in known disease genes at the verified false-assertion rate, validated on post-cutoff expert-reviewed and MAVE data, with ancestry-stratified accuracy reported.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Calibration needs high-quality, non-circular clinical and functional ground truth across genes and ancestries, plus methodology for mechanism-aware predictions.
Bottlenecks
ranked by votes- #1EmpiricalNon-circular, ancestry-balanced ground truth
ClinVar is biased toward well-studied genes and European-ancestry cohorts; MAVE data cover only a fraction of genes.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Nov 1, 2024Independent AlphaMissense VUS studysource
Sensitivity 92% and specificity 78%; 56 of 4,085 VUS became likely pathogenic when AlphaMissense was added.
- Sep 19, 2023AlphaMissensesource
Predictions for 71 million variants released with 89% classified.
- Nov 3, 2022ClinGen calibration of predictorssource
Pejaver et al. map tool scores to evidence strengths and recommend thresholds.
- Mar 5, 2015ACMG/AMP guidelinessource
Standard five-tier framework including PP3/BP4 computational evidence criteria.