Under the hood

Structure prediction uses a combination of pre-computed structures and on-demand inference. Known proteins are typically returned in seconds; novel sequences are predicted on dedicated GPU hardware.

The output is a standard PDB file and a mean pLDDT score (predicted Local Distance Difference Test, 0–100). Regions above 70 are generally reliable; below 50 likely indicates disorder or low evolutionary constraint.

Mutation scoring uses a protein language model trained on hundreds of millions of sequences across the tree of life. It learns which amino acids are evolutionarily tolerated at each position — a strong proxy for functional importance.

For each residue, the model computes a log-probability score for all 20 standard amino acids. The wild-type score near zero means the residue is highly expected; strongly negative scores flag conserved, mutation-sensitive positions. The heatmap renders the full sequence — click any position to expand all 20 AA scores and the top 5 suggested substitutions.

If you provide a protein name, foldfunc queries a literature database to retrieve relevant published abstracts. These are passed into the interpretation step to ground the analysis in real published science rather than model priors alone.

Structure lookup and literature retrieval run concurrently to minimise total latency.

The sequence, pLDDT scores, mutation profile, and retrieved literature are synthesised by an AI language model into a structured biological interpretation. The output covers:

• Protein family classification and functional context
• Structural observations tied to high- and low-confidence regions
• Mutation-sensitive positions and their likely significance
• Open research questions suggested by the analysis
• A reliability note based on pLDDT confidence

The model reasons only from what is passed in; retrieved literature is the primary external knowledge source.