MutaScope — Predicting the effect of every protein mutation
MutaScope is a live web platform that predicts how every possible single amino-acid change affects a protein’s function. Pick one of 217 proteins, or paste your own sequence, and you see the predicted effect of each mutation on an interactive heatmap linked to the protein’s 3D structure. Everything is benchmarked on ProteinGym, the standard public benchmark: about 696,000 measured mutations from 217 deep mutational scanning experiments.
- Zero-shot scoring with protein language models. Adding structure information (SaProt) clearly beats sequence alone (ESM-2): Spearman 0.46 vs 0.41. I reproduced the benchmark’s published scores almost exactly, and in doing so found two errors in them.
- Supervised models (ridge, LightGBM, MLP) trained on each protein’s own lab data, tested under three cross-validation schemes with leakage checks. I deployed ridge because it predicts best at positions the model has never seen (+0.09 Spearman over zero-shot), and every prediction comes with a calibrated 90% interval (90.8% coverage on held-out data).
- Live scoring of any sequence through a FastAPI service on Google Cloud Run, deployed automatically by a CI/CD pipeline. Profiling the service under load led to a fix that made it 60 times faster for page reads.
- Fully reproducible: a Nextflow pipeline, checksum-pinned data and models, and experiment tracking in MLflow.