Skip to main content
Back to timeline
bioRxivSource publication:

EnzRef reads reactants and products separately before touching the enzyme, lifting R² by 19.8% on unseen enzymes

Synopsis

Predicting enzyme turnover numbers (kcat) requires resolving activity differences between reactions of the same enzyme and between enzymes acting on the same reaction. EnzRef splits reaction chemistry into reactant and product conditions, interacts each separately with a frozen protein language model, expresses every conditioned enzyme readout as a deviation from that enzyme's own mean residue response, and recomposes the two responses. On the CatPred-DB source held-out set, mean single-model R² rises from 0.553 for MCKcat to 0.614; on the TurNuP-DB enzyme-unseen subset, R² rises from 0.259 to 0.310, a 19.8% relative gain without target-domain adaptation.

Source-provided article image: Decompose and Recompose: Reference-Centered Multimodal Adaptation of Protein Language Models for Enzyme Turnover Prediction
Figure 1 · bioRxiv · Page 1

Interpretation

Turnover prediction models now read the same enzyme under reactants and products as two separate conditions, instead of merging both chemical sides into one reaction vector before enzyme interaction. Earlier PMAK and MCKcat encode reactants and products jointly through RXNFP before residue-level interaction, so the enzyme's separate responses to the two chemical roles, and the relation between them, stay implicit in the joint representation. On the CatPred-DB source held-out set of 2,049 records, mean single-model R² over ten independent runs rises from 0.553 for MCKcat to 0.614, and SRCC from 0.732 to 0.758; EnzRef's mean single-model R² also exceeds MCKcat's ensemble score of 0.599.

Each chemically conditioned enzyme readout is restated as a deviation from that enzyme's own residue background, so differences between samples are no longer swamped by a shared enzyme background. A cross-attention readout carries both a chemically conditioned component and an enzyme-specific residue background that is shared across chemical conditions within one enzyme but varies between enzymes, so comparing raw readouts mixes the two kinds of variation. The variant without reference centering, keeping the other modules and the same parameter count, reaches mean source held-out R² of 0.553 against 0.614 for full EnzRef; paired activity-difference RMSE falls 6.47% for reactions sharing an enzyme and 5.79% for enzymes sharing a reaction, with both errors lower in all ten paired runs.

The two referenced responses are recomposed through their mean deviation and a directed role contrast before the prediction head, making the chemical role relation explicit in the enzyme's context. Keeping separate role conditions without this recomposition gives mean source held-out R² of 0.601 and SRCC of 0.753. At equal parameter count, adding symmetry-aware semantic recomposition raises mean R² to 0.614 and SRCC to 0.758, while paired activity-difference RMSE drops from 1.023 to 1.005 for shared enzymes and from 1.162 to 1.135 for shared reactions.

The whole adaptation trains only a small set of parameters beyond the frozen encoder and is applied directly to external data after source training. With ESM2-650M frozen, the chemical encoder and downstream modules hold 1.72M trainable parameters, fewer than the overhead the compared methods carry beyond the shared frozen encoder. On 1,008 external records whose enzyme sequences are absent from source training and validation, fixed source-trained models reach mean R² of 0.310 against 0.259 for MCKcat, a 19.8% relative gain; with cached inputs, inference takes 0.775±0.035 seconds per checkpoint, 3.14× faster than MCKcat.

Perspective

The demonstrated scope is the CatPred-DB source held-out set plus two nested TurNuP-DB subsets, 2,746 pair-unseen and 1,008 enzyme-unseen records, with external evaluation applying fixed source-trained models and no target-domain training or tuning. The method requires known reactant and product annotations, so the full formulation is limited when products are unknown or incompletely specified; extending enzyme-referenced role modeling to partially specified reactions is the direction the authors propose.

External transfer is evaluated only on two nested TurNuP-DB subsets with models fixed after source training; the full formulation depends on known reactant and product annotations and is limited when products are unknown or incompletely specified. Source curation corrected a unit-conversion error in CatPred-DB's released preprocessing of SABIO-RK records, changing 1,918 labels, so numbers are not fully comparable with the original release.