CPI-Pred: A deep learning framework for predicting functional parameters of compound-protein interactions

Zhiqing Xu, Rana Barghout, Jinghao Wu, Dhruv Garg, Yun S. Song, Radhakrishnan MahadevanbioRxiv 2025 · DOI

TL;DR

This paper introduces CPI-Pred, a novel deep learning framework that accurately predicts compound-protein interactions (CPI) by integrating advanced protein language models with a message-passing neural network for compound representation. This approach significantly improves the prediction of crucial enzyme kinetic parameters and other functional properties, offering a valuable tool for metabolic engineering and drug discovery.

What Problem Does This Paper Address?

Predicting how small molecules (compounds) interact with proteins, especially enzymes, is fundamental to understanding biological processes, designing new drugs, and engineering metabolic pathways. However, this is a notoriously difficult problem. The complexity of these interactions, involving intricate binding sites and varying affinities, coupled with the scarcity and heterogeneity of available experimental data, makes it hard for current models to generalize well to new, unseen proteins and compounds. Existing methods often struggle to capture the full picture, leading to limitations in accuracy and robustness.

Methods

CPI-Pred employs a multi-pronged deep learning approach:

  • Protein Representation: It leverages state-of-the-art protein language models (PLMs) like ESM-2, ProtBERT, Ankh, and CARP to generate rich embeddings from amino acid sequences. Crucially, it introduces novel sequence embedding pooling methods, specifically self-attention pooling and LSTM variational encoder pooling, to better capture positional information and handle variable sequence lengths, outperforming traditional methods.
  • Compound Representation: A message-passing neural network (MPNN) is used to learn representations of compounds. Instead of relying on standard chemical descriptors, CPI-Pred uses novel atom-level featurization that captures substructures around each atom, providing a more nuanced understanding of molecular properties.
  • Interaction Modeling: To effectively learn the complex relationships between protein and compound representations, CPI-Pred utilizes a cross-attention mechanism. This allows the model to focus on the most relevant parts of both representations, overcoming the limitations of simpler concatenation and feed-forward networks used in prior work.
  • Ensembling: To further boost accuracy and robustness, CPI-Pred combines the outputs of four distinct prediction pipelines, each with different combinations of the above components.

Datasets

The authors compiled and utilized several datasets for training and evaluation:

  • BRENDA Database: The primary source for enzyme kinetic parameters (kcat, KI, KM, kcat/KM). To address missing UniProt IDs in BRENDA, they integrated the KinMod database to create "pangenomic datasets," significantly expanding the data coverage.
  • Experimental Enzyme Activity Datasets: Used for evaluating enzyme activity prediction, including:
  • PafA point mutations dataset (for kcat prediction).
  • avGFP dataset (for local fitness landscape analysis).
  • Phosphatase activities, kinase-ligand binding affinities, halogenase activities, esterase hydrolytic activities, and aminotransferase activities.
  • PDBbind Dataset: A well-established benchmark for protein-ligand binding affinity prediction.

The creation of these comprehensive "pangenomic" datasets, especially for kinetic parameters, is a significant contribution, offering a valuable resource for the community.

Key Insights and Ideas

  • Atom-level featurization for compounds is powerful: Moving beyond standard chemical descriptors to focus on the local atomic environment within a molecule significantly improves compound representation for interaction prediction.
  • Self-attention and LSTM pooling are superior for protein sequences: These methods effectively capture crucial positional and contextual information in protein sequences, outperforming simpler pooling strategies, especially when dealing with variable sequence lengths.
  • Cross-attention is key for complex interactions: Directly learning the interplay between protein and compound representations via cross-attention is more effective than traditional concatenation followed by a feed-forward network.
  • Pangenomic datasets unlock broader prediction: Augmenting existing datasets with information from resources like KinMod to create "pangenomic" datasets dramatically improves model performance, especially for predicting interactions involving less characterized proteins.
  • Deep learning can predict beyond sequence: CPI-Pred's ability to predict binding affinity without requiring 3D structural information of binding pockets highlights the power of sequence- and structure-based representations learned by deep learning.

Key Results

  • State-of-the-art performance: CPI-Pred consistently outperforms existing models like DLkcat and KNN across various tasks, including kinetic parameter prediction, enzyme activity prediction, and binding affinity prediction.
  • Kinetic Parameter Prediction: CPI-Pred trained on pangenomic datasets achieved higher performance than models trained on core datasets and surpassed KNN and DLkcat on all four kinetic parameters (kcat, KI, KM, kcat/KM). For example, it achieved a Pearson's correlation coefficient of 0.786±0.008 for kcat on the pangenomic dataset.
  • Enzyme Activity Prediction: CPI-Pred demonstrated superior performance, achieving substantially better results than DLkcat and KNN models, with Pearson's correlation coefficients reaching up to 0.872 for kinase-ligand affinity.
  • Binding Affinity Prediction: CPI-Pred achieved a correlation of 0.845 on the PDBbind dataset, slightly surpassing HAC-Net (0.835) and significantly outperforming DLkcat and KNN. Notably, it achieved this without requiring binding pocket information.
  • Error Predictor: An integrated error predictor achieved 71% classification accuracy in distinguishing between high and low confidence predictions, enhancing the reliability of the model.

Limitations and Future Work

  • Performance on novel sequences/compounds: While CPI-Pred performs well on smaller datasets, its accuracy declines when predicting for entirely novel sequences and compounds.
  • Pangenomic approach challenges: Applying the pangenomic strategy to enzyme activity datasets is challenging, especially for novel sequences or compounds, highlighting the need for more high-fidelity data.
  • Data uncertainties: The BRENDA kinetic data used has inherent uncertainties due to variable experimental conditions, which can affect predicted parameters. While efforts were made to mitigate this, it remains a limitation.
  • Future Work: The authors suggest that the curated datasets will serve as valuable benchmarks for future CPI prediction algorithms. They also encourage the collection and utilization of high-quality data to drive further advancements in the field. The model's architecture and performance on diverse tasks suggest its potential for applications in protein engineering and drug discovery for novel targets.

Summary updated

Abstract

Recent advancements in deep learning have enabled functional annotation of genome sequences, facilitating the discovery of new enzymes and metabolites. However, accurately predicting compound-protein interactions (CPI) from sequences remains challenging due to the complexity of these interactions and the sparsity and heterogeneity of available data, which constrain the generalization of patterns across their solution space. In this work, we introduce CPI-Pred, a versatile deep learning model designed to predict compound-protein interaction function. CPI-Pred integrates compound representations derived from a novel message-passing neural network and enzyme representations generated by state-of-the-art protein language models, leveraging innovative sequence pooling and cross-attention mechanisms. To train and evaluate CPI-Pred, we compiled the largest dataset of enzyme kinetic parameters to date, encompassing four key metrics: the Michaelis-Menten constant (<i>K</i> <sub>M</sub>), enzyme turnover number (<i>k</i> <sub>cat</sub>), catalytic efficiency (<i>k</i> <sub>cat</sub> <i>/K</i> <sub>M</sub>), and inhibition constant (<i>K</i> <sub>I</sub>). These kinetic parameters are critical for elucidating enzyme function in metabolic contexts and understanding their regulation by compounds within biological networks. We demonstrate that CPI-Pred can predict diverse types of CPI using only the amino acid sequence of enzymes and structural representations of compounds, outperforming state-of-the-art models on unseen compounds and structurally dissimilar enzymes. Over workflow provides a valuable tool for tackling a range of metabolic engineering challenges, including the designing of novel enzyme sequences and compounds, such as enzyme inhibitors. Additionally, the datasets curated in this study offer a valuable resource for the scientific community, serving as a benchmark for machine learning models focused on enzyme activity and promiscuity prediction.
BibTeX
@misc{xu2025cpi,
  author = {Zhiqing Xu and Rana Barghout and Jinghao Wu and Dhruv Garg and Yun S. Song and Radhakrishnan Mahadevan},
  title = {{CPI-Pred: A deep learning framework for predicting functional parameters of compound-protein interactions}},
  year = {2025},
  doi = {10.1101/2025.01.16.633372},
  howpublished = {bioRxiv},
}
Figure