UniKP
Summary¶
Yu, Deng, He, Keasling & Luo (2023) present UniKP, a unified framework that predicts three enzyme kinetic parameters — turnover number (kcat), Michaelis constant (Km), and catalytic efficiency (kcat/Km) — from a single architecture combining pretrained protein and molecule language model embeddings with a tree-ensemble regressor. An extension, EF-UniKP, further incorporates environmental factors (pH, temperature) via a two-layer framework, and the authors demonstrate UniKP-guided enzyme discovery and directed-evolution applications.
Method¶
- Representations: pretrained protein language model embeddings for the enzyme sequence, combined with pretrained molecular representations for the substrate.
- Regressor selection: 16 machine learning models plus 2 deep learning architectures (CNN, RNN) were benchmarked; ensemble tree methods substantially outperformed both simple linear models and the deep learning baselines on this relatively small (~10k), high-dimensional (2048-d) dataset — Extra Trees (R²=0.65) was the best-performing model, versus Linear Regression (R²=0.38), CNN (R²=0.10), and RNN (R²=0.19).
- EF-UniKP: a two-layer framework that additionally models pH and temperature as inputs, trained on newly constructed pH and temperature datasets.
- Reweighting: four reweighting strategies were explored to reduce prediction error specifically for high-value (high-activity) enzymes, which are disproportionately important for enzyme engineering applications but underrepresented in typical training data.
Performance¶
- kcat prediction: Extra Trees regressor achieves R²=0.65 on the primary kcat benchmark.
- Km prediction: R²=0.530 on an 11,722-entry natural enzyme-substrate combination benchmark (dataset originally curated by Kroll et al.'s Km predictor) — described as matching state-of-the-art performance.
- EF-UniKP vs. base UniKP: on independent pH/temperature test sets, EF-UniKP outperformed both UniKP and a "Revised UniKP" baseline (R²=0.44 vs. 0.38 on pH data); restricting to strictly unseen substrate/enzyme combinations (62 pH samples, 61 temperature samples), EF-UniKP still led with R²=0.45 (pH) and R²=0.31 (temperature).
- Independent benchmark: CatPred reports that UniKP is outperformed by CatPred especially on out-of-distribution samples (e.g., kcat R²=0.365 vs. 0.260 for CatPred vs. UniKP on enzymes ≤40% sequence-similar to any training enzyme), while both clearly outperform DLKcat and a naive baseline.
Availability¶
Code and usage examples: github.com/Luo-SynBioLab/UniKP. Data: Zenodo 10.5281/zenodo.10115498.
See Also¶
- Structural Km Predictor — source of UniKP's Km benchmark dataset.
- CatPred — later model reporting improved out-of-distribution accuracy relative to UniKP.
- CataPro — retrains and directly compares against UniKP under unbiased, sequence-clustered evaluation.
- Enzyme Kinetic Parameter Prediction — field overview.