Enzyme EC classification, v3: homology, residue profiles, a hierarchical head and a blended ranker
Submission for the Leak CURBER enzyme classification benchmark (DA3410 Deep Learning Lab, Project 2). Given an enzyme amino-acid sequence, the model predicts one fully specified EC number. It runs on CPU only and offline, inside the evaluator's 2 vCPU / 16 GB / 10 minute budget.
Local results
Five folds cover all 14,466 training clusters at 50% identity. Evaluation uses 14,103 non-redundant queries (one per cluster, best hit under 60% identity). Every score is out-of-fold for the reference, the head and the ranker. Matched reweights queries to the identity histogram of the real inputs; raw does not and is pessimistic. Treat the two as a range.
| system | matched L4 | L3 | L2 | L1 | raw L4 |
|---|---|---|---|---|---|
| DIAMOND bitscore vote | 0.719 | 0.884 | 0.900 | 0.921 | 0.669 |
| v1 as shipped (approximation) | 0.732 | 0.906 | 0.924 | 0.951 | 0.683 |
| v2 | 0.739 | 0.912 | 0.930 | 0.956 | 0.692 |
| v3 (this model) | 0.741 | 0.912 | 0.930 | 0.956 | 0.694 |
| perfect ranking of v2's candidates | 0.794 | 0.925 | 0.940 | 0.961 | 0.754 |
v3 against v2, paired on identical queries: L4 +0.0020, 95% CI [+0.0004, +0.0036]. On folds 3-4 alone, which never guided any v3 choice: +0.0024 [-0.0002, +0.0051].
These are local estimates, not leaderboard scores.
What changed from v2
- A blend of three rankers. Five-seed bags of LightGBM models trained with LambdaRank, XE-NDCG and a pointwise binary objective. Alone each scores like v2; averaged, their different errors partly cancel.
- A wider ESM-2 gate: inputs whose best bitscore is under 300 (v2: 150) get head evidence, embedded weakest first in three tiers so that a slower machine still covers the inputs that gain most.
- Reproducible ranking. Candidates with equal scores were ordered by Python's per-process hash seed in v2; v3 breaks ties by EC name.
- A deadline for the larger ranker: evidence gathering stops at 500 s (v2: 540 s).
The reference database, residue-profile features and hierarchical head are v2's, unchanged.
What was measured and not adopted
| idea | source | L4 vs v2 |
|---|---|---|
| Ranker hyperparameter search (60 configurations, confirmed on held-out folds) | β | +0.0000 |
| Query-relative (listwise) features | learning-to-rank practice | +0.0001 |
| More head-only candidates (2, 10, 15 instead of 5) | β | -0.0002 to -0.0008 |
| Residue-profile constants (6 variants) | HIT-EC / GraphEC residue focus | within Β±0.0008 |
| Up to 10 training queries per cluster instead of 3 | β | +0.0000 |
| Hard-negative contrastive batches + compositional label weights in the head | CLEAN; HIT-EC, CREEP/Horizyn | -0.0001 |
| Learned mix of all 13 ESM-2 layers in the head | β | worse on fold 0 |
| Rescue search with e-value 10 for no-hit inputs | β | best hit right 6% of the time |
| Embedding every input | β | below gate 300 |
How it works
- Homology. DIAMOND
--very-sensitiveagainst all 186,488 unique training sequences, 200 hits with full alignments, read through a pipe. - Residue profiles. Each candidate EC's aligned homologs form a Henikoff-weighted profile on the query's columns, compared with its competitors' (the fourth digit is decided by a few specificity-determining residues).
- Protein language model. ESM-2 (35M) and a head combining CLEAN's supervised contrastive projection with HIT-EC's hierarchical L1-L4 outputs: probabilities, centroid similarity, nearest neighbours and extra candidates.
- Ranking. 15 LightGBM models score each (query, candidate EC) pair on 47 features; z-scored raw scores are summed.
- Fallback. Anything with no evidence receives the most frequent training EC.
Reliability
- No file over 16 MiB written during inference: the DIAMOND index is built in the Dockerfile,
--hit-membufkeeps DIAMOND's data in memory, hits arrive through a pipe. Verified on all 11,367 inputs withRLIMIT_FSIZE= 16 MiB. - Runtime: 496 s for 11,367 inputs at 2 threads on an Apple M4 (limit 600 s). The embedding stage deliberately fills the time up to its deadline; a slower machine embeds fewer inputs.
- Failure tests, all with valid output for every input: the main search killed by its deadline
(
EC_TIME_BUDGET=130, 85% agreement with the full run), and an unusable DIAMOND index.
Usage
docker build --platform linux/amd64 -t my-enzyme-model .
docker run --rm --platform linux/amd64 --user 1000:1000 --network none --cpus 2 --memory 16g \
-v "$PWD/../../data:/input:ro" -v "$PWD/output:/output" \
my-enzyme-model python /app/predict.py \
--input /input/val_inputs.parquet --output /output/predictions.jsonl
EC_TIME_BUDGET (default 500 s) and EC_THREADS (default 2) override the defaults.
References
- Yu, Cui, Li, Luo, Jiang, Zhao. Enzyme function prediction using contrastive learning (CLEAN). Science 379(6639), 2023.
- Dumontet, Han, Lee, Oh, Kang. Trustworthy prediction of enzyme commission numbers using a hierarchical interpretable transformer (HIT-EC). Nature Communications 17(1), 2026.
- Song, Yuan, Chen, Zeng, Zhao, Yang. Accurately predicting enzyme functions through geometric graph learning on ESMFold-predicted structures (GraphEC). Nature Communications 15(1), 2024.
- Yang, Mora, Liu, Wittmann, Anandkumar, Arnold, Yue. CARE: a Benchmark Suite for the Classification and Retrieval of Enzymes. arXiv:2406.15669, 2025.
- Rocks et al. Dual-encoder contrastive learning accelerates enzyme discovery (Horizyn). PNAS 123(12), 2026; Mikhael, Chinn, Barzilay. CLIPZyme. ICML 2024.
- Buchfink, Reuter, Drost. Sensitive protein alignments at tree-of-life scale using DIAMOND. Nature Methods 18, 2021.
- Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model (ESM-2). Science 379, 2023.
- Ke et al. LightGBM. NeurIPS 2017.
EC vocabulary from ExPASy ENZYME (SIB, CC BY 4.0).