LANCE: learning enzyme kinetics

Bach Pham, Kinjal Mondal & Jeffery B. Klauda
University of Maryland | Manuscript under mentor review | Not yet published

Enzyme measurements are sparse and depend on experimental conditions. LANCE predicts k_cat, K_m, and K_i with one shared architecture and three prediction heads.

One model, three predictions

The core idea of my architecture is to implement the JEPA-style objectives.

Joint-Embedding Predictive Architecture (JEPA)-style learning guides two latent routes with different roles. Molecule-conditioned cross-attention first connects molecular features with enzyme residue features. Two shared 64-dimensional action vectors, adapted for each task, then update separate binding and catalytic representations:

• Reaction JEPA uses available products during training to help the catalytic route predict a product representation from the reactant representation rather than reconstructing the product.
• Binding JEPA uses a product-free target to guide the binding route toward a representation combining the molecule and enzyme interface.

Products are not required at inference. LANCE-7R also adds a second stage with just 406 fitted parameters. It retrieves stored training measurements for an exactly matching reactant, excluding the query's sequence-99 cluster, to adjust k_cat predictions while leaving K_m and K_i unchanged. Without an eligible exact match, the original neural prediction is preserved. This gives turnover a selective correction without retraining the shared backbone.

LANCE-7R vs. CatPred

We retrained CatPred v1.0.0 on the same CatPred-DB split using ten independent runs per task and evaluated both systems on exactly matched rows.

LANCE-7R has favorable point estimates in 18 of 18 Heldout/OOD99 comparisons and 41 of 48 comparisons across all supplied Heldout and sequence-OOD cohorts.

On the 348-row turnover OOD99 cohort, R^2 improves from 0.323 to 0.395 and MAE falls from 1.024 to 0.979. On 1,193 held-out inhibition measurements, MAE falls from 0.865 to 0.788. These results measure the complete system; they do not establish how much of the gain comes from either JEPA objective alone.

CatPred: mean of ten outer ensembles. LANCE-7R: one selected ten-member ensemble. R^2 and MAE use log10 targets; P1mag (within 10x) is the percentage with absolute log10 error below 1. OOD labels are the supplied target-specific sequence-identity cohorts; Km and Ki have no supplied OOD20 rows. Counts summarize metric directions across nested cohorts, not statistical significance.