PROTEA fullgo: first place on the LAFA continuous benchmark
This repository holds the trained models behind PROTEA's first-place result on the LAFA continuously updated public benchmark for protein function prediction. On the sealed September 2025 to March 2026 target frame (7,401 proteins), scored with the benchmark's own harness, PROTEA leads the public leaderboard.
| Method | No knowledge | Limited knowledge | Partial knowledge | Overall |
|---|---|---|---|---|
| PROTEA (this work) | 0.477 | 0.482 | 0.215 | 0.391 |
| TransFew (previous leader) | 0.428 | 0.485 | 0.230 | 0.381 |
| FunBind | 0.441 | 0.451 | 0.205 | 0.366 |
| PROTEA, similarity search only | 0.412 | 0.394 | 0.165 | 0.324 |
The score is the benchmark's IA-weighted micro-F, averaged over the three branches of the Gene Ontology, for the three standard settings.
How it works, in plain terms
A similarity search finds the most related proteins whose functions are already known and borrows their annotations (strong when nothing else is known). On top of that, a learned combiner weighs three additional signals: a direct predictor that proposes functions across the whole catalogue, the protein's own weaker prior annotations, and, for proteins that are already partly characterised, the functions that habitually accompany what is already known across different branches of the ontology. Every ingredient was chosen on an earlier period of data and the system was measured once on the later, sealed period.
Files
ensemble_gbm_NK.txt,ensemble_gbm_LK.txt,ensemble_gbm_PK.txt: the per-setting LightGBM combiners (the final stage).classifier_6plm_asl.pt,classifier_m2_anc2vec.pt: the direct full-catalogue predictor over six concatenated frozen protein representations.feature_spec.json: the exact feature set and the sealed per-setting scores.
Reproduce
Full recipe, data periods, and commands: the fullgo/ module of the
protea-reranker-lab repository
(REPRODUCE.md). The whole system uses frozen, pre-computed protein representations and
fits on a single 12 GB GPU.
Author
Francisco Miguel Pérez Canales. Part of the PROTEA doctoral thesis project.