HIT-EC

Enzyme Commission numbers are a four-level hierarchical code — class, subclass, sub-subclass, and serial number — that specify what reaction an enzyme catalyses. Assigning them computationally is a multi-label classification problem with thousands of classes and a brutal long tail: in one widely used benchmark, EC numbers with 25 or fewer training sequences account for 41% of all classes, and prior deep learning methods scored around 0.70 F1 on them. A second problem is annotation quality. As of September 2022, 15.4% of Swiss-Prot enzyme entries carried incomplete EC numbers, where the first two or three levels are known but the rest is not — data that most training pipelines simply discard.

HIT-EC, from University of Nevada, Las Vegas and Sun Moon University, addresses both. It mirrors the EC hierarchy in the network itself, stacking four transformer encoders so that each one predicts a level and passes its representation onward. A dedicated loss handles partially annotated sequences instead of dropping them. And it is trained as an evidential model, so each prediction comes with a measure of how much evidence supports it rather than a bare softmax score.

Unlike much recent enzyme annotation work, HIT-EC does not sit on a pretrained protein language model. Its embedding layer is a learned matrix over the 23 sequence tokens, trained from scratch alongside the encoders.

#Key Features

• Hierarchy-aligned encoders: Four transformer encoders correspond to the four EC levels; each produces a level-specific local prediction, and the original sequence embedding is reinjected into every encoder to prevent the deeper levels from drifting.
• Local and global prediction flows: A classification token accumulated through the encoder stack drives a global linear prediction, which is aggregated with the four local predictions to form the final multi-label output.
• Learning from incomplete annotations: Sequences with partial EC numbers contribute supervision at the levels they do specify; ablation shows this yields at least a 2% F1 gain on both averaged metrics and on underrepresented classes.
• Evidential predictions: The model reports evidence rather than raw probabilities, so low-support calls are distinguishable from confident ones — relevant given that a compared baseline always emits at least one EC number and therefore has a false discovery rate of 1 on non-enzymes.
• Attention-based interpretation: Per-token relevance scores accumulated across the four attention layers highlight the residues driving a prediction, recovering conserved motifs and substrate recognition sites in the CYP106A2 cytochrome P450 family.

#Technical Details

The cross-validation corpus draws roughly 200,000 sequences from Swiss-Prot and the PDB released before September 2022, covering 1,938 EC numbers with at least ten sequences each, split 80/10/10 by stratified sampling. Sequences are truncated or padded to 1,024 residues; the tuned configuration uses an embedding dimension of 1,024, two attention heads, and dropout 0.1, optimised with Optuna.

HIT-EC records a micro-averaged F1 of 0.93 ± 0.01 against 0.88 for CLEAN, 0.81 for ECPICK, and 0.79 for DeepECtransformer, and a macro-averaged F1 of 0.84 ± 0.02 against 0.80, 0.75, and 0.58. On classes with fewer than 25 sequences it reaches 0.77 ± 0.02, ahead of CLEAN at 0.73. External validation on New-392 — Swiss-Prot enzymes registered after the training cutoff — gives 0.69 ± 0.04 micro-F1 across all sequences and 0.94 ± 0.05 within the models' EC coverage. Across fourteen complete microbial genomes from KEGG it leads in eight species and averages 0.81 micro-F1 against 0.78 for ECPICK and 0.77 for CLEAN. The distributed checkpoint is trained more broadly still: Swiss-Prot, PDB, curated KEGG entries, and TrEMBL sequences realigned with DIAMOND, totalling about 450,000 sequences and covering 4,255 EC numbers.

#Applications

The natural use is bulk functional annotation of newly sequenced microbial genomes and metagenomes, where many predicted proteins have no characterised homolog and rare EC classes matter most. Metabolic engineering and pathway reconstruction workflows benefit from the evidential scores, which let a curator triage which automated calls warrant follow-up, while residue-level relevance maps point at the regions the model associates with catalytic identity. A public web platform accepts FASTA uploads and returns predictions with contribution scores, and a demo notebook loads the released v2.0.0 checkpoint for local inference.

#Impact

HIT-EC's clearest result is on the tail of the EC distribution, where hierarchical supervision and the incomplete-annotation strategy together close a gap that has limited enzyme annotation tools for years. It also makes a case that a task-specific transformer trained from scratch remains competitive with contrastive-learning approaches built on pretrained protein embeddings, at least where a large labelled corpus exists. Code, the pretrained checkpoint, and the interpretation notebook are released under the MIT license. The work appeared as a preprint in February 2025 and was published in Nature Communications in January 2026 under the shortened title "Trustworthy prediction of enzyme commission numbers using a hierarchical interpretable transformer".