Related Experiment Videos
Optimizing drug discovery through long short-term memory recurrent neural networks (LSTM RNNs): A hybrid multi-modal
Siddharth Goswami1, Sachin Sharma2
1Department of Biotechnology, Graphic Era Deemed to be University, Dehradun, Uttarakhand 248002, India.
Computational Biology and Chemistry
|June 17, 2026
Summary
This study introduces a hybrid CNN-LSTM-GNN model for efficient drug-target interaction prediction, achieving good performance on modest hardware. Sequence modeling proved most crucial for prediction accuracy.
Area of Science:
- Computational chemistry
- Bioinformatics
- Machine learning
Background:
- Drug discovery is resource-intensive, necessitating efficient machine learning (ML) for large biochemical datasets.
- Scaling ML frameworks on modest hardware is crucial for broader accessibility.
Purpose of the Study:
- To develop and evaluate a hybrid Convolutional Neural Network-Long Short-Term Memory-Graph Neural Network (CNN-LSTM-GNN) model, termed HMLCG, for drug-target interaction (DTI) prediction.
- To assess the model's performance and computational efficiency using BindingDB and PubChem BioAssay datasets.
Main Methods:
- A hybrid CNN-LSTM-GNN (HMLCG) architecture was implemented with approximately 1.0 million parameters.
- The model was trained on ligand-target pairs from BindingDB, utilizing techniques like mixed-precision, label smoothing, and OneCycleLR scheduling.
- Ablation studies were conducted to evaluate the contribution of individual components (LSTM, CNN, GNN).
Main Results:
- The HMLCG model achieved an ROC-AUC of 0.7917, accuracy of 75.11%, and F1-score of 0.4885 in under 9 minutes on CPU.
- Ablation studies revealed significant performance degradation upon removing the LSTM pathway (ROC-AUC to 0.4955), highlighting its importance.
- LSTM-only configurations yielded the best AUC values, suggesting sequence modeling as the primary driver of DTI prediction accuracy in this framework.
Conclusions:
- The HMLCG model demonstrates computational efficiency without significant compromise in predictive power for DTI prediction.
- While CNN and GNN components offer complementary structural insights, sequence modeling (LSTM) is the dominant factor in this architecture.
- The model serves as a valuable prioritization tool for high-throughput screening, particularly for imbalanced datasets, but should not be used as a standalone decision system.