Protein → Music

Turn disease-causing protein mutations into sound — using ESM-2 embeddings to measure how "unusual" a mutation looks to a model trained only on evolution.

🎵 Try it live

https://huggingface.co/spaces/H9kkk/protein-music

Search any human protein, pick a real documented disease variant, and take the blind listening test — can you tell which audio track is the mutant, just by ear?

What it does

1. Pick any human protein (search via UniProt) and one of its real documented disease-causing variants
2. The app fetches both the healthy and mutant sequence
3. Both are run through ESM-2 (8M param, runs on CPU) to get per-residue embeddings
4. The embedding distance between healthy and mutant is computed at each position — this spikes sharply at the mutation site (e.g. p53 R175H: ~12x the local average)
5. That disruption signal is sonified — higher disruption = more dissonant — and rendered as a chart

How it works — honesty section

• The detection signal (embedding distance) is genuine, unmanipulated output from ESM-2. No disease labels were used in training.
• The sonification rules (disruption → dissonance, pitch mapping, etc.) are hand-designed, not learned. This is "pretrained model as feature extractor + deterministic mapping," not novel ML.
• This is an exploratory/educational tool, not a diagnostic one.

Code overview

The Python files in this repo implement the pipeline:

File	Role
fetcher.py	UniProt REST API — sequences, variants, mutations
embedder.py	ESM-2 embeddings, distance computation, music parameter mapping
composer.py	Converts music parameters into note sequences
synth.py	Renders notes to .wav directly (sine waves + harmonics)
server.py	Flask app — routes, ties everything together
templates/index.html	Frontend — search, charts (Chart.js), audio players

Tech stack

• Model: facebook/esm2_t6_8M_UR50D (Hugging Face transformers)
• Backend: Flask
• Data: UniProt REST API
• Audio: numpy + scipy (direct waveform synthesis, no soundfont)
• Charts: Chart.js
• Hosting: Hugging Face Spaces (Docker)

Ideas for extension

• Train a classifier on ESM-2 deltas + ClinVar labels to predict pathogenicity (would make this a genuine ML contribution beyond feature extraction)
• 3D structure visualization (AlphaFold) showing mutation site in context
• Crowd-sourced blind test results — track aggregate "% correctly identified"