Preprint
Machine Learning

DeepCDA: deep cross-domain compound–protein affinity prediction through LSTM and convolutional neural networks

Karim Abbasi(Laboratory of Systems Biology and Bioinformatics (LBB), Institute of Biochemistry and Biophysics, University of Tehran , Tehran 1417614411, Iran), Parvin Razzaghi(Department of Computer Science and Information Technology, Institute for Advanced Studies in Basic Sciences (IASBS) , Zanjan 4513766731, Iran), Antti Poso(School of Pharmacy, Faculty of Health Sciences, University of Eastern Finland , Kuopio 80100, Finland), Massoud Amanlou(Department of Medicinal Chemistry, Drug Design and Development Research Center, Tehran University of Medical Sciences , Tehran 1416753955, Iran), Jahan B Ghasemi(Chemistry Department, Faculty of Sciences, University of Tehran , Tehran 1417614418, Iran), Ali Masoudi-Nejad(Laboratory of Systems Biology and Bioinformatics (LBB), Institute of Biochemistry and Biophysics, University of Tehran , Tehran 1417614411, Iran)
May 27, 2020Bioinformatics255 citations
DOI

255

Citations

11

Influential Citations

Bioinformatics

Venue

2020

Year

Abstract

Abstract Motivation An essential part of drug discovery is the accurate prediction of the binding affinity of new compound–protein pairs. Most of the standard computational methods assume that compounds or proteins of the test data are observed during the training phase. However, in real-world situations, the test and training data are sampled from different domains with different distributions. To cope with this challenge, we propose a deep learning-based approach that consists of three steps. In the first step, the training encoder network learns a novel representation of compounds and proteins. To this end, we combine convolutional layers and long-short-term memory layers so that the occurrence patterns of local substructures through a protein and a compound sequence are learned. Also, to encode the interaction strength of the protein and compound substructures, we propose a two-sided attention mechanism. In the second phase, to deal with the different distributions of the training and test domains, a feature encoder network is learned for the test domain by utilizing an adversarial domain adaptation approach. In the third phase, the learned test encoder network is applied to new compound–protein pairs to predict their binding affinity. Results To evaluate the proposed approach, we applied it to KIBA, Davis and BindingDB datasets. The results show that the proposed method learns a more reliable model for the test domain in more challenging situations. Availability and implementation https://github.com/LBBSoft/DeepCDA.

Analysis

Why This Paper Matters

Accurate prediction of compound-protein binding affinity is critical for drug discovery, but most computational methods assume training and test data come from the same distribution. In practice, new compounds or proteins often differ from those seen during training, leading to poor generalization. DeepCDA directly addresses this real-world challenge by incorporating adversarial domain adaptation, making it highly relevant for practitioners who need robust models in dynamic drug discovery pipelines.

The paper combines multiple deep learning techniques—CNNs, LSTMs, attention, and domain adaptation—into a cohesive framework. This integration is significant because it leverages the strengths of each component: CNNs capture local substructures, LSTMs model sequential dependencies, attention highlights important interactions, and domain adaptation aligns feature distributions. The result is a method that not only predicts affinity but also adapts to new domains without requiring retraining from scratch.

Technical Contributions

  • Hybrid representation learning: CNNs extract local patterns from compound and protein sequences, while LSTMs capture long-range dependencies, enabling the model to learn occurrence patterns of substructures.
  • Two-sided attention mechanism: Unlike standard attention that focuses on one modality, this mechanism jointly attends to both protein and compound substructures, encoding their interaction strength more effectively.
  • Adversarial domain adaptation: A feature encoder for the test domain is learned by minimizing distribution divergence between training and test features, using a domain classifier that is trained adversarially. This allows the model to generalize to unseen data distributions.
  • Three-phase training: The approach separates representation learning, domain adaptation, and prediction into distinct phases, providing modularity and clarity in the training process.

Results

The method was evaluated on three benchmark datasets: KIBA, Davis, and BindingDB. DeepCDA consistently outperformed baseline methods, including standard deep learning models without domain adaptation, especially in cross-domain settings where training and test sets had different distributions. For example, on the KIBA dataset, the model achieved higher concordance index (CI) and lower mean squared error (MSE) compared to alternatives. The authors note that the improvement is more pronounced in challenging scenarios, demonstrating the value of domain adaptation.

Significance

DeepCDA contributes to the broader AI field by demonstrating how domain adaptation can be effectively integrated into deep learning for structured biological data. This approach is not limited to drug discovery; similar techniques could be applied to other domains where distribution shifts occur, such as materials science or personalized medicine. The code is publicly available, facilitating reproducibility and further research. By addressing a practical limitation of existing methods, DeepCDA moves AI closer to reliable deployment in real-world scientific discovery.

deep-learninglstmcnnattentiondomain-adaptationdrug-discoverybinding-affinity
Attention MechanismsNeural Networks

Related Papers