A hybrid machine learning framework leveraging biophysicochemical insights for scalable discovery of protein-ligand interactions

Achal Ajeet Rayakar, Chaitanya Kumar Jaladanki, Qunxiang Ong, Bin Tan, Junjun Chen, Ah‐Hwee Tan, Laurence A. Gan Lim, Navita Kohaal, Yu Chen, Jun Wang, W. Y. Han, Pengcheng He, Hwee Kuan Lee, Hao FanbioRxiv 2025 · DOI

TL;DR

This paper introduces COMRADE, a novel hybrid machine learning framework that significantly accelerates drug discovery by efficiently screening vast chemical libraries for potential drug-protein interactions. By combining diverse pre-trained representations with a unique contrastive learning approach, COMRADE achieves superior accuracy and speed compared to existing methods, leading to the identification of novel drug candidates, including a new scaffold for the understudied BCKDK target.

What Problem Does This Paper Address?

The process of discovering new drugs is notoriously slow, expensive, and has a low success rate. A critical bottleneck is identifying promising small molecules (ligands) that can interact with specific protein targets in the body to modulate their function. Traditional methods like high-throughput screening and molecular docking are either too slow or too costly for the massive chemical libraries available today. While deep learning has shown promise, current end-to-end models often struggle because they are trained on relatively small, task-specific datasets, failing to leverage the rich information contained in larger, domain-specific "foundation models" (like protein language models). This paper aims to bridge this gap by developing a more efficient and accurate in silico method for predicting compound-protein interactions (CPIs) that can handle massive datasets and incorporate diverse biological insights.

Methods

The core of the COMRADE framework is a module called CE-Screen. CE-Screen takes a compound-protein pair and processes it through several stages:

  1. Feature Generation: It leverages seven distinct, high-dimensional representations. These include:
  2. Protein Language Model (pLM) embeddings: Derived from models like DLM-LSTM and ESM-2, these capture protein sequence information.
  3. Molecular Transformer embeddings: From models like MolFormer-XL, these capture molecular structure and context.
  4. Molecular descriptors: Traditional cheminformatics features (e.g., from Mordred).
  5. An original physics-based Interaction Potential Encoding (IPE): This novel feature uses Merck Molecular Force Field (MMFF) atom types to capture target-compound compatibility without requiring 3D structures or docking.
  6. Contrastive Compression: A residual neural network is employed to compress these diverse, high-dimensional features into a single, compact, and discriminative 750-dimensional representation. This is achieved using a modified supervised contrastive loss function that not only separates interacting (ligand) and non-interacting (decoy) pairs but also encourages feature decorrelation. This compression is key to making the representation more robust and less prone to overfitting.
  7. Classification: The compressed representation is then fed into a lightweight ensemble classifier, specifically an Extra Trees model. This ensemble of decision trees makes the final prediction of whether a compound interacts with a protein.

This hybrid approach decouples representation learning from classification, allowing for more robust and generalizable predictions.

Datasets

The study utilizes several established datasets for training and evaluation:

  • D-DUD-E (Directory of Useful Decoys, Enhanced): Used for training and testing.
  • BindingDB: Used for training and testing.
  • ChEMBL: Used for training and testing.

The authors also created specific subsets of BindingDB and ChEMBL for testing, designed to be distinct from each other and D-DUD-E to assess generalization across diverse compound-protein pairs and novel targets.

For prospective testing, a large dataset of approximately 10.8 million compounds from the ZINC database was screened against five targets.

Key Insights and Ideas

  • Leveraging Foundation Models is Crucial: The paper strongly advocates for incorporating pre-trained representations from large-scale foundation models (like pLMs and molecular transformers) rather than relying solely on end-to-end learning from limited CPI datasets.
  • Contrastive Learning for Robust Representations: The use of contrastive learning with a decorrelation penalty is a key innovation. It effectively compresses high-dimensional, heterogeneous features into a compact representation that is both discriminative and less redundant, improving downstream classifier performance.
  • Hybrid Approach Outperforms End-to-End: Decoupling representation learning (via contrastive compression) from classification (via Extra Trees) proves more effective than end-to-end deep learning models, especially in imbalanced datasets and for generalization to novel targets.
  • Physics-Informed Features Matter: The inclusion of an original physics-based interaction potential encoding (IPE) as a feature set demonstrates the value of integrating biophysical principles alongside learned representations.
  • Speed and Accuracy are Achievable Together: COMRADE achieves a significant speedup (~100x faster than docking) while maintaining or improving accuracy, making it practical for screening massive chemical libraries.

Key Results

  • Performance Improvement: CE-Screen (the core of COMRADE) outperforms state-of-the-art end-to-end deep learning models by up to 111.11% in terms of AUPRC and 23.39% in terms of logAUC on retrospective benchmarks.
  • Speed: CE-Screen is approximately 100 times faster than standard molecular docking.
  • Large-Scale Screening: COMRADE successfully screened a ~10.8 million-compound library against five targets.
  • Novel Hit Identification: The framework identified at least one previously untested ligand for each of the five targets. Notably, it discovered a new scaffold for BCKDK, an understudied but high-value target.
  • Robustness: CE-Screen shows superior performance with matched decoys compared to random decoys, indicating better generalizability and robustness against overfitting.

Limitations and Future Work

  • Decoy Limitations: The paper acknowledges that all decoy sets are approximations of true non-binders and have inherent limitations. While matched and random decoys offer different perspectives, they don't perfectly represent the complexity of real-world screening libraries.
  • Hyperparameter Tuning: While the framework's hyperparameters are tuned on a held-out validation set, the authors note that the specific form of some functions and hyperparameters cannot be directly optimized via a learning algorithm.
  • Future Work: The authors propose COMRADE as a modular and scalable framework. They suggest that its architecture can readily incorporate emerging biophysicochemical representations and alternative classifiers, enabling continuous improvement. The decoupled compression-and-classification paradigm is also presented as a blueprint for accelerating other molecular prediction tasks beyond CPI screening. The successful identification of novel scaffolds for targets like BCKDK opens avenues for further optimization and development.

Summary updated

Abstract

Abstract Improving in silico compound-protein interaction (CPI) predictability is critical for productive drug discovery. Current deep learning approaches largely rely on end-to-end models trained on limited labeled CPI datasets, overlooking the representational power of large-scale biochemical foundation models. We present COMRADE (Contrastive Multirepresentation Accelerated Docking Engine), a hybrid virtual screening framework that accelerates docking by triaging compounds using CE-Screen (Contrastive Embedding-Screen). CE-Screen leverages seven high-dimensional pretrained representations – including those from protein language models and molecular transformers, along with an original physics-based interaction potential encoding – for rapid first-pass screening ∼100× faster than docking. Its contrastive compression neural network maps these inputs onto a single compact, discriminative representation optimized for CPI prediction via a lightweight ensemble classifier. CE-Screen outperforms state-of-the-art end-to-end models by up to 111.11% on retrospective benchmarks and is successfully used to triage ∼10.8 million compounds against five targets, yielding novel hits for each one – including a new scaffold for the branched-chain ketoacid dehydrogenase kinase (BCKDK), an understudied yet high-value target in metabolic disease and oncology.
BibTeX
@misc{rayakar2025hybrid,
  author = {Achal Ajeet Rayakar and Chaitanya Kumar Jaladanki and Qunxiang Ong and Bin Tan and Junjun Chen and Ah‐Hwee Tan and Laurence A. Gan Lim and Navita Kohaal and Yu Chen and Jun Wang and W. Y. Han and Pengcheng He and Hwee Kuan Lee and Hao Fan},
  title = {{A hybrid machine learning framework leveraging biophysicochemical insights for scalable discovery of protein-ligand interactions}},
  year = {2025},
  doi = {10.1101/2025.03.16.643501},
  howpublished = {bioRxiv},
}
Figure