DeepRES

Between 20% and 50% of the reactions catalogued in metabolic pathway databases such as KEGG have never been assigned to a gene sequence. These are orphan enzymes: reactions with confirmed activity but no known protein to carry them out. The mirror image sits in UniProt, where millions of proteins carry no functional annotation. No existing tool searched across the gap — enzyme classifiers assign proteins to fixed schemas such as EC numbers and cannot express a reaction outside that schema, while reaction-query tools such as E-zyme 2.0 accept any reaction but retrieve only already-annotated proteins.

DeepRES, developed by Keisuke Hirota and Takuji Yamada at the Institute of Science Tokyo and posted to bioRxiv in July 2025, is a two-stage framework built to close that gap. The first stage, EnzymeCNN, is a dilated convolutional network that decides whether a protein is an enzyme at all, reading amino acid sequence and Foldseek 3Di structural tokens through parallel convolutional stacks before merging them. The second stage, EnzymeCLIP, borrows the contrastive language-image pre-training recipe from computer vision: separate protein and reaction encoders project into a shared 256-dimensional latent space, and the cosine similarity between the two embeddings becomes a catalytic-activity score. Because reactions enter as Reaction SMILES strings rather than class labels, DeepRES is independent of any reaction classification schema — the authors describe it as the first method designed specifically to associate orphan enzymes with proteins of unknown function.

#Key Features

• Schema-free enzyme-reaction retrieval: Reactions are represented as Reaction SMILES rather than EC numbers or reaction classes, so screening is not bounded by an annotation vocabulary.
• Sequence and structure together: EnzymeCNN consumes amino acid and 3Di sequences jointly, and ablations showed it beating sequence-only and structure-only variants, with structural tokens contributing most of the signal.
• Pretrained encoders adapted with LoRA: EnzymeCLIP builds on SaProt and RXNFP, a BERT-based reaction encoder, fine-tuning both backbones through low-rank adaptation rather than training from scratch.
• Robust to low homology: With similar training data stripped out by MMseqs2 and Foldseek at progressively tighter identity and TM-score thresholds, homology-search baselines collapsed while EnzymeCNN held its accuracy and annotated every input rather than returning no-hits.
• Trained once, applied at scale: A single fixed pipeline screened 13.9 million metagenomic proteins with no per-target retraining, yielding a reusable catalogue of putative enzymes.

#Technical Details

The protein dataset was built from Swiss-Prot release 2024_03 clustered at 90% identity with MMseqs2; after dropping entries lacking an AlphaFold DB structure or exceeding 850 residues, 294,522 proteins were split 8:1:1 into training, validation and test sets of 235,617, 29,452 and 29,453 examples, with enzymes (defined by the presence of an EC annotation) making up roughly 49% of each split. Structures came from AlphaFold DB v4, converted to 3Di with Foldseek; held-out benchmark proteins were folded with AlphaFold 2. Mapping the enzyme subset onto Rhea release 133 produced 139,156 enzyme-reaction pairs, split by reaction into 106,174 training and 18,869 test pairs.

Forty-eight EnzymeCLIP variants were trained for 10 epochs across five seeds and ranked by enrichment factor over the top 10% of predictions. SaProt-based encoders consistently outperformed ESM-2-based ones, fine-tuning the backbones beat freezing them, and a cyclic-consistency regularizer improved results while a soft contrastive loss degraded them. Against CLEAN, the leading EC-number annotation tool, EnzymeCLIP was competitive on top-1 and better on top-10 accuracy over newly released Swiss-Prot and Rhea entries. Code is MIT-licensed; both checkpoints are on Zenodo under CC BY 4.0.

#Applications

DeepRES targets genome and metagenome annotation where homology transfer runs out of signal. In the paper's application, 13.9 million unknown-function proteins from the ESMFold-derived ESM Metagenomic Atlas were reduced to 595,340 predicted enzymes, of which 97,074 were linked to 897 of 1,255 orphan enzymes drawn from the EnteroPathway gut-microbiome database. Those candidates then served as a BLAST reference against 4,744 human gut metagenome-assembled genomes from MGnify, surfacing biosynthetic gene cluster candidates for 15 pathway modules, including anthocyanin degradation clusters concentrated in Bacteroides. That homology step needs no GPU, so the catalogue slots into conventional pipelines, and the framework applies equally to eukaryotes and viruses.

#Impact

DeepRES extends enzyme annotation from classification into cross-modal retrieval, showing that a CLIP-style objective over pretrained biological encoders can search protein and reaction space jointly at genome scale. Its practical output — a reference set of putative orphan-enzyme genes with genomic and taxonomic context — is usable by groups that run no deep learning themselves. The authors are candid about the limits: roughly 40% of the proteins EnzymeCNN called enzymes could not be linked to any reaction, RXNFP's treatment of SMILES handles chirality poorly despite the stereoselectivity of real enzymes, and CLIP's one-to-one pairing assumption fits awkwardly with the many-to-many relationships created by homology, enzyme multifunctionality and convergent evolution. All predictions are computational candidates awaiting experimental confirmation, and the work remains a preprint pending peer review.