Boltz2ESI

Enzymes achieve catalytic selectivity through precise complementarity between an active site and its substrate, yet the substrate scope of most sequenced enzymes has never been measured. Sequence-based predictors infer compatibility from global similarity or protein language model embeddings and cannot resolve a single active-site substitution. Structure-aware methods sharpen this, but they depend on a predefined binding pocket — usually transferred from a homolog — and on rigid-body docking that treats the scaffold as static, missing the ligand-induced rearrangements that accompany turnover.

Boltz2ESI, posted as a bioRxiv preprint in August 2026 by researchers at Northeastern University with collaborators at the Broad Institute, KAIST, EPFL, and HITS Inc., replaces that pipeline with native co-folding. Given only an enzyme sequence and a substrate SMILES string, it folds the pair together in two stages — a full-length, MSA-guided co-fold that localizes the active site, then an MSA-free refold of the cropped pocket with the substrate that lets side chains and backbone adapt to the ligand — and reads the interaction out of the resulting complex. No pocket annotation is required.

It is a separately trained model rather than a mode of its backbone: Boltz-2, ESM3, and the substrate encoders are frozen feature extractors, and the trained component is a new geometry-conditioned interaction module. That distinction is load-bearing, because Boltz-2's own affinity predictions score near chance here — binding strength and catalytic compatibility are not the same quantity.

#Key Features

• Pocket-free two-stage co-folding: An MSA-guided full-length co-fold picks the highest-confidence of five complexes to localize the active site; the cropped pocket and substrate are then refolded without MSAs, keeping the pose with the best interface predicted TM-score.
• Multimodal fusion: Co-folding single, pair, and distogram representations are combined with per-residue ESM3-1.4B embeddings for evolutionary context and with Uni-Mol2 3D embeddings plus 2,048-bit Morgan fingerprints for substrate chemistry.
• Specificity beyond affinity: On the held-out split, raw Boltz-2 affinity reaches AUROC 0.5258 while Boltz2ESI reaches 0.7664, separating catalytic compatibility from tight but non-productive binding.
• Sub-family resolution: Discrimination holds as negatives are drawn from progressively closer Enzyme Commission classes, down to decoys sharing all four EC digits.
• Zero-shot pathway prioritization: Applied without pathway-specific training to 107 P450 candidates from a Withania somnifera transcriptome, it ranked CYP87G1 8th and CYP88C7 4th, ahead of EnzymeCAGE at 13th and 10th.

#Technical Details

The interaction module fuses ESM3 embeddings into enzyme tokens and the Uni-Mol2 plus Morgan features into substrate tokens through two-layer MLPs (both emitting 384 dimensions), broadcasts them into a pair tensor conditioned on the predicted distogram, and passes the result through a four-block Pairformer stack whose attention is masked to protein-ligand and ligand-ligand pairs. Mean pooling and an MLP head give a scalar interaction probability. The Pairformer is initialized from Boltz-2's pretrained affinity head; only this module is updated during training, with upstream features precomputed offline. Training used binary cross-entropy and AdamW at learning rate 1e-4, batch size 64 across eight NVIDIA L40S GPUs, at roughly 35 hours per fold.

Training data is the ESIBank benchmark: 323,783 quality-filtered pairs spanning 34,417 substrates and 8,124 enzymes, assembled from BRENDA and UniProt with EC-hierarchy negative sampling across five difficulty levels, plus curated sets for six enzyme families. Under four-fold cross-validation on the random split, Boltz2ESI reaches AUROC 0.9156 and AUPR 0.6221 against 0.8927/0.5995 for EZSpecificity and 0.6572/0.2057 for ESP; on the unknown-enzyme-and-substrate split it reaches AUROC 0.7664 against 0.7198 and 0.6778. Ablations show ESM3 embeddings are the largest single contributor (AUROC 0.6605 without them), and swapping co-folded active sites for precomputed docking poses collapses AUPR from 0.3331 to 0.2169. Optional family-specific fine-tuning improved AUPR in all five families tested and AUROC in four, most notably thiolases (0.6327 to 0.7457).

#Applications

The immediate use case is pathway de-orphaning: scoring candidate enzymes against a set of metabolites to shortlist which ones catalyze a step, as demonstrated on the withanolide cascade. The same scoring supports substrate-scope profiling for biocatalyst selection, triage of homologs that share an EC number but differ in reactivity, and prioritizing constructs for expression and assay in natural-product and metabolic-engineering programs. Sequence-and-SMILES inputs mean it applies to enzymes with no solved structure or characterized pocket.

#Impact

Boltz2ESI shows that a general co-folding model's representations carry information about catalytic specificity its affinity head does not expose, and that a small module trained on frozen structural and evolutionary features can recover it. The authors state the limits plainly: two-stage co-folding costs far more than sequence-only or docking pipelines, limiting proteome-scale throughput; ESIBank negatives drawn from incomplete BRENDA annotations include pseudo-negatives that depress measured performance for every method; and homology transfer still competes where good templates exist, as in the third withanolide step, where EnzymeCAGE ranked the true P450 sixth to Boltz2ESI's seventeenth. Code and cross-validation checkpoints are public, though the repository carries no declared license, and the preprint is not peer reviewed.