Wiki · Bio × AI Methods · concept
Variant-effect prediction
Predicting whether a genetic mutation is harmful or harmless — often from evolution and protein language models alone.
Variant-effect prediction asks a blunt clinical question: will this mutation break the protein or not? The commonest case is a missense variant — a single DNA change that swaps one amino acid for another. Some swaps are silent passengers the body tolerates fine (benign); others cripple a protein and cause disease (pathogenic). The goal is to tell those two apart before anyone gets sick.
How it works
The oldest and still strongest signal is evolutionary conservation: a position that has stayed unchanged across millions of years of related species is probably load-bearing, so mutating it is risky. Modern methods learn this automatically. A protein language model trained on natural sequences assigns every possible amino acid a probability at every position — and a variant the model finds surprising (low probability) is more likely to be damaging. That gives a prediction with no disease labels at all, straight from sequence: ESM-1v does exactly this “zero-shot.” DeepMind’s AlphaMissense goes further, combining the structural context of an AlphaFold-style model with language-model likelihoods to score missense changes across the human proteome.
Why it matters (for bio × AI)
Sequencing has outrun interpretation. Millions of human variants sit in databases tagged “variant of uncertain significance” (VUS) — seen in a patient, but nobody knows if they matter. A model that scores all of them at once lets clinical geneticists triage: focus the expensive follow-up on the variants most likely to be real. That is genuine leverage. But the honest caveat matters: these scores are hypotheses, not diagnoses, and the benchmarks can be circular — a model trained on what evolution conserves, then tested against pathogenicity labels that also reflect conservation, can look better than it is. Useful for prioritizing; not a verdict.
Related
See protein language model for the likelihoods that power zero-shot scoring, and AlphaFold for the structural context AlphaMissense builds on.