🧬 biology

Robust enzyme kinetics prediction through pairwise relative learning

The paper introduces DeltaKcat, a siamese neural network framework that reformulates enzyme kinetics prediction as a pairwise relative learning problem to overcome the limitations of absolute regression, thereby achieving superior accuracy, robustness, and generalization in identifying highly active enzymes across noisy biological datasets.

Original authors: Feiran Li, Xiongwen Li, Zhengkai Li, Jiawei Zou, Wenjie Chen, Shujia Liu, Ke Wu, Jiahao Luo, Yu Chen

Published 2026-10-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Feiran Li, Xiongwen Li, Zhengkai Li, Jiawei Zou, Wenjie Chen, Shujia Liu, Ke Wu, Jiahao Luo, Yu Chen

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Life depends on tiny biological machines called enzymes. These proteins act as catalysts, speeding up the chemical reactions that keep cells alive, from digesting food to repairing DNA. To understand how well an enzyme works, scientists measure its speed and efficiency, values known as kinetic parameters. These numbers tell researchers how fast an enzyme can turn a raw material into a product and how tightly it holds onto that material. While scientists have cataloged millions of protein sequences, fewer than one in a thousand have these crucial speed measurements. This gap leaves a vast landscape of biological potential uncharted. To fill it, researchers have turned to computers, hoping to predict these speeds based on the enzyme's genetic code. However, the data available to train these computer programs is messy. Measurements for the same enzyme often vary wildly depending on which laboratory took them or which specific conditions were used, creating a noisy signal that confuses standard prediction models.

A team of researchers at Tsinghua University and Boston University has developed a new approach to solve this problem, shifting the focus from guessing exact numbers to understanding relative differences. They created a system called DeltaKcat, which treats enzyme prediction not as a test of memory, but as a lesson in comparison. Instead of asking a computer to predict the exact speed of a single enzyme in isolation, the system looks at pairs of enzymes working under similar conditions. It learns to predict how much faster or slower one enzyme is compared to another within the same experimental context. By focusing on the difference between two things rather than the absolute value of one, the model effectively filters out the background noise caused by inconsistent laboratory methods. This allows the system to learn the true relationship between an enzyme's structure and its function, ignoring the specific quirks of any single study.

The researchers built this system using a massive collection of data gathered from public databases and scientific literature, organizing over 135,000 speed measurements and hundreds of thousands of binding measurements. They paired these entries so that the computer could compare enzymes that shared a substrate or substrates that shared an enzyme, ensuring the comparison happened within a consistent experimental setting. The model uses advanced language tools to read the genetic sequences of proteins and the chemical structures of the molecules they act upon, translating these complex patterns into a format the computer can understand. It then calculates the difference in performance between the paired items. This method proved remarkably effective. When tested on new enzymes it had never seen before, the system could accurately predict how their speeds would change relative to known examples, even when the genetic similarity was low. It also excelled at identifying the most efficient enzymes, a task where older models often failed because they tended to guess average values rather than extreme ones.

To prove the system works in the real world, the team tested it on a specific group of enzymes called adenylate kinases, which are vital for energy management in cells. This entire group was completely absent from the training data, meaning the computer had to rely entirely on its learned ability to compare and contrast. Without any prior exposure to these specific enzymes, the model successfully identified the most active variants, outperforming all existing methods. When the researchers provided just a small amount of new data to fine-tune the system, its accuracy improved significantly, demonstrating that the approach can adapt quickly to new biological families. The findings suggest that by changing how we ask the question—moving from absolute prediction to relative comparison—we can build more reliable tools for discovering and engineering enzymes. This progress could accelerate the design of new biocatalysts for medicine and industry, turning the vast, unmeasured library of life into a usable resource.

Technical Summary: Robust Enzyme Kinetics Prediction through Pairwise Relative Learning

Problem Statement

Enzyme kinetics prediction is a critical bottleneck in biocatalyst design and metabolic engineering. While over 230 million protein sequences exist in UniProt, fewer than 0.1% have experimentally measured kinetic parameters (e.g., $k_{cat}$ , $K_m$ , $k_{cat}/K_m$ ) in databases like BRENDA and SABIO-RK. Existing computational methods, which have evolved from traditional machine learning to deep learning and large pretrained models, face two fundamental limitations:

Data Sparsity and Heterogeneity: The enzyme-substrate space is sparsely sampled, and kinetic measurements exhibit substantial cross-study variability. Reported values for the same enzyme-substrate system can differ by orders of magnitude due to varying experimental conditions, introducing systematic noise.
Bias in Absolute Regression: Current approaches typically rely on absolute regression across heterogeneous datasets. This leads to "regression-to-the-mean" bias, where models capture dataset-specific biases rather than true biophysical relationships, resulting in poor identification of highly active enzymes and limited generalization to unseen sequences or substrates.

Methodology: DeltaKcat

To address these challenges, the authors propose DeltaKcat, a framework that reformulates enzyme kinetics prediction as a pairwise relative learning problem rather than an absolute value estimation task.

Core Framework

Pairwise Relative Learning: Instead of predicting absolute kinetic values independently, DeltaKcat learns the kinetic differences between paired enzyme-substrate complexes. Pairs are constructed from the same study to ensure consistent experimental contexts, thereby mitigating cross-study systematic bias.
Siamese Neural Network Architecture:
- Input: A pair of enzyme-substrate complexes.
- Encoding:
  - Protein sequences are encoded using ProtT5 (yielding residue-level embeddings).
  - Substrates are encoded using MolT5 (continuous vectors) and MACCS fingerprints.
- Interaction Modeling: A gated cross-attention module integrates enzyme and substrate features. The substrate representation queries the protein sequence representation to capture specific enzyme-substrate interactions, with a gating unit refining attention allocation.
- Prediction Head: The interaction representations of the two complexes are combined with their explicit difference vector and fed into a Multilayer Perceptron (MLP) to predict relative kinetic differences ( $\Delta k_{cat}$ , $\Delta K_m$ , $\Delta k_{cat}/K_m$ ).

Data Construction

The model was trained on a curated dataset of over 135,000 $k_{cat}$ , 208,000 $K_m$ , and 110,000 $k_{cat}/K_m$ measurements, integrated from BRENDA, SABIO-RK, and literature-mined resources (EnzyExtraDB, Enzyme Co-Scientist).

Sequence-Diversity Task: Pairs share the same substrate but have different enzyme sequences.
Substrate-Specificity Task: Pairs share the same enzyme but have different substrates.
Cold-Start Split: Evaluation splits were designed to hold out entire sequences or substrates to test generalization to unseen entities.

Absolute Parameter Inference

To enable practical application, the authors introduce a reference-based reconstruction strategy. Predicted relative differences are anchored to experimentally characterized values of a reference complex from the training set to recover absolute kinetic parameters.

Key Results

1. Relative Kinetics Prediction

DeltaKcat demonstrated strong performance in predicting relative kinetic differences:

Achieved Pearson Correlation Coefficients (PCC) up to 0.80 for relative $k_{cat}$ , $K_m$ , and $k_{cat}/K_m$ .
Showed task-dependent sensitivity: $\Delta k_{cat}$ was best predicted in sequence-diversity tasks, while $\Delta K_m$ excelled in substrate-specificity tasks.
Maintained robustness in cold-start scenarios (unseen sequences/substrates), achieving a PCC of 0.74 for $\Delta K_m$ in the unseen-substrate setting.
Accurately captured the direction of kinetic changes with 0.71 accuracy.

2. Absolute Kinetics Prediction

Using the reference-based inference:

Cold-Start Performance: DeltaKcat significantly outperformed baselines (DLKcat, CataPro, UniKP) in cold-start settings. For $k_{cat}$ , it achieved an $R^2$ of 0.71 and PCC of 0.85, compared to baseline $R^2$ values of 0.14–0.48.
Publication-Level Validation: When evaluated within individual publications (simulating consistent experimental protocols), DeltaKcat achieved a median PCC of 0.78 for $k_{cat}$ , vastly outperforming baselines (PCC 0.11–0.36).
Identification of Highly Active Enzymes: Unlike absolute regression models that bias toward the mean, DeltaKcat successfully identified highly active enzymes. It outperformed all baselines in the top 10–25% of active enzymes, with the largest gains in the top 10%.
Generalization: The model maintained strong performance (PCC > 0.76) even in low sequence similarity regimes (0–40% identity).

3. External Validation (Adenylate Kinase)

The framework was tested on an external Adenylate Kinase (ADK) dataset completely excluded from training:

Zero-Shot: DeltaKcat achieved a PCC of 0.40 and Spearman Correlation Coefficient (SCC) of 0.33, outperforming all baselines. It successfully identified highly active enzymes within its top 5 candidates, whereas baselines failed to retrieve any in their top 10.
Fine-Tuning: With as few as 60 ADK-specific training samples, the model's performance improved to a PCC of 0.69 (a 72% gain over zero-shot), surpassing previously reported ADK-specific models.

Significance and Claims

The paper claims that DeltaKcat establishes a general framework for enzyme kinetics prediction based on pairwise relative functional learning. Its primary contributions are:

Robustness to Noise: By learning invariant functional relationships within consistent experimental contexts, the model mitigates inter-study variability and measurement inconsistency inherent in heterogeneous biological datasets.
Improved Generalization: The approach shifts the learning objective from dataset-dependent absolute values to functional relationships, enabling robust extrapolation to unseen enzymes, substrates, and low-similarity regimes.
Practical Utility: The framework effectively prioritizes highly active enzymes, addressing a critical need in enzyme discovery and engineering where identifying outliers is more valuable than average predictions.
Broad Applicability: The authors posit that this relational learning approach is broadly useful for learning from noisy biological data beyond enzyme kinetics, potentially aiding in enzyme engineering, metabolic design, and functional annotation.

The authors acknowledge limitations, including the current lack of explicit environmental factors (pH, temperature) and structural information in the model, and the continued scarcity of high-quality kinetic measurements for rare reaction types. However, they assert that the proposed framework provides a principled approach to overcoming the biases of absolute regression in noisy biological data.