AI-assisted design of targeted mutations
A model that reads DNA and writes the smallest edit that produces a chosen effect — not just predicting what mutations do, but designing the ones you want. Scored 92% exact-match — seven points past the original 85% goal.
An encoder–decoder Transformer that takes a DNA sequence plus a natural-language target effect and generates the minimal base-pair changes, in HGVS form. Trained on ~260,000 real human variant sequences at 2,056 bp each, built from human-reviewed UniProt annotations, Ensembl cDNA, and Levenshtein-inferred codon changes.
The whole pipeline: multi-threaded UniProt/Ensembl scrapers with graceful HTTP retries, a custom-trained parser that turns free-text variant effects into training conditions, K-mer + SentencePiece tokenization, and an architecture that makes 2K-bp DNA tractable — RoPE, grouped-query attention, and an 8× CNN downsample. Fifteen full training runs on TPU v5e; the dataset later grew to 840,000 samples.