ESM-ECForest

A newly sequenced bacterial genome delivers a few thousand predicted proteins and almost no functional labels. Assigning Enzyme Commission (EC) numbers to them is what turns those sequences into metabolism, since an EC number links a gene product to a specific biochemical reaction. The task is hard in a particular way: the EC system is hierarchical, so a wrong call at the first digit poisons all four; annotations are wildly uneven, with ligases and translocases holding a fraction of the examples available for oxidoreductases; and alignment-based transfer degrades among distant homologs whose biochemistry has drifted.

ESM-ECForest answers this by refusing to learn the sequence-to-function map from EC labels at all. Protein sequences are encoded once by a frozen ESM-2 checkpoint, and every parameter the study actually fits sits in scikit-learn Random Forests over those embeddings. The pipeline runs in two stages: stage A is a binary forest separating enzymes from non-enzymes, and only sequences it passes reach stage B, a multi-label bank of forests that assigns one or more EC numbers. Rare EC classes then need only enough data to draw a decision boundary in an embedding space that masked-language-model pretraining over unlabeled sequences has already organized.

The framework was released in August 2026 by Xiao Hua and Ghjuvan M. Grimaud at the Division of Biotechnology and Applied Microbiology, Lund University. It sits alongside a cluster of recent EC predictors that take different routes to the same hierarchy: HIT-EC stacks one transformer encoder per EC level, EZpred pools embeddings across unlabeled homologs, and FuncSeek casts assignment as retrieval in a fused embedding space. ESM-ECForest is the minimal-machinery member of that group.

#Key Features

• Identify-then-classify staging: Screening out non-enzymes first keeps the multi-label stage from inventing catalytic annotations for structural or regulatory proteins. Stage A reached 0.976 F1 externally, correctly calling 97.7% of enzymes and 96.6% of non-enzymes.
• A deliberately small frozen backbone: Embeddings come from esm2_t6_8M_UR50D, the 6-layer, roughly 8M-parameter ESM-2 variant rather than the 650M or 3B one. The final-layer beginning-of-sequence token gives a 320-dimensional vector per protein, chunked and averaged for sequences past 1,000 residues.
• Multi-label EC assignment: Stage B trains 3,021 one-vs-rest forests, one per EC label seen in training, and retains any label scoring at least 0.7. Multifunctional enzymes — 20.2% of training enzymes and 50.7% of annotated benchmark proteins — receive every applicable number.
• Homology-aware evaluation: Splits were made at the level of MMseqs2 clusters at 40% identity and 80% coverage, and benchmark proteins share at most 38.4% identity with any training sequence.
• CPU-scale inference: Predicting the 25,778-sequence benchmark took 11.6 minutes and 13.4 GB of RAM on 24 CPU threads, no GPU.

#Technical Details

Training data came from UniProtKB/Swiss-Prot release 2026_01 — entries scoring 4 or 5 for annotation quality, 50–5,000 residues long, from a phylogenetically spread selection of prokaryotic and unicellular eukaryotic taxa: 104,648 enzymes and 80,366 non-enzymes, 185,014 sequences partitioned 8:1:1 at cluster level. Stage A is a single forest of 400 trees at maximum depth 40 with balanced class weights; stage B is the 3,021-classifier bank. On the independent 25,778-protein benchmark spanning 15 species, weighted F1 ran 0.944, 0.929, 0.920, and 0.902 across EC Levels 1–4, ahead of DeepECTransformer (0.925–0.857), ProteInfer (0.918–0.821), and DeepEC (0.660–0.548). Macro F1, which weights rare labels equally, fell further with depth — 0.911 at Level 1 to 0.790 at Level 4 — and the classwise picture is consistent: EC 1–3 scored 0.965, 0.932, and 0.956, while lyases through translocases landed between 0.861 and 0.897.

#Applications

The intended use is bulk functional annotation ahead of metabolic reconstruction: predicted EC numbers become gene-protein-reaction associations in genome-scale metabolic models, candidates for gap-filling missing reactions, and evidence for revising existing associations. The authors frame it as a prioritization tool rather than a definitive annotation system, since an EC prediction describes catalytic potential encoded in a sequence, not realized flux in a cell. The released classifiers are downloaded and pointed at a FASTA file, which suits groups annotating draft genomes or metagenome-assembled contigs without ML infrastructure.

#Impact

The result worth carrying away is how little supervised machinery the accuracy required: the smallest ESM-2 checkpoint plus off-the-shelf forests matched or beat end-to-end supervised transformers at every level of the hierarchy, with the largest margins in the least-annotated classes. Several caveats bound that claim. The work is a preprint that has not been peer reviewed, and the authors ran no ablations, so the separate contributions of the embeddings, the forests, and the two-stage design remain unseparated. Because stage B covers only EC labels present in training, previously unseen terminal EC numbers cannot be assigned, though higher-level digits may still be recovered. Training taxa were restricted to prokaryotes and unicellular eukaryotes, and the released repository contains inference scripts only — the training code is not part of the release.