Chapter one
The taxonomy of evidence
The claim
PROTEA does not use one model to predict function. It gathers several independent kinds of evidence and lets a final judge weigh them. One kind is a compact numerical fingerprint of the protein, learned by a protein language model (a neural network trained on millions of sequences). Another is a set of per-candidate clues the judge reads directly: how similar a match is, how common a term is, how a candidate term relates to what we already know. A third is grafted in from an outside catalogue, the InterPro2GO mapping, which links known protein domains to functions; it adds 0.0179 to the sealed score and is what tips seven of the nine cells into first place. Why so many kinds? Because none of them wins everywhere. And the judge is frugal: of the sixty-four clues it is offered, sixteen carry ninety-five percent of the weight and twenty-eight carry none at all.
Pull the marker to see the receipt and the operation that regenerates the number.
The evidence
| Gain share | ||
|---|---|---|
| go_term_frequency | 0.255 | base-rate prior |
| anc2vec_query_known_maxcos | 0.194 | |
| lineage_is_ancestor_of_known | 0.127 | |
| k_position | 0.117 | |
| anc2vec_neighbor_cos | 0.093 | |
| vote_count | 0.034 | |
| anc2vec_neighbor_maxcos | 0.034 | |
| identity_nw | 0.022 |
Our caveats
We state these before anyone states them for us.
- Gain measures how much the trees lean on a column, not causation on held-out proteins.
- The single largest driver, go_term_frequency at twenty-five percent, is a base-rate prior. We surface it rather than hide it.
- Of the twenty-eight zero-gain features, the emb_pca_* columns are populated but worthless, while the interpro_* columns read zero only because their producer tables are empty. That is a broken producer, not a proof of uselessness, and the graft above is the evidence InterPro can matter.