Abstract
Recent breakthroughs in protein structure prediction, such as AlphaFold, have catalyzed significant interest in deep learning for Drug-Target Affinity (DTA) tasks. However, many existing approaches focus on single molecular properties or homogeneous information, neglecting the rich heterogeneous data inherent in molecular structures and interactions. We propose MolHF (Molecular Heterogeneous attributes Fusion), an end-to-end framework designed to address this gap. By utilizing a Molecular Heterogeneous Information Learning (MHIL) module and a Molecular Heterogeneous Attention Fusion (MHAF) module, MolHF extracts and integrates diverse structural information. Experimental results on benchmark datasets demonstrate that MolHF outperforms state-of-the-art baselines across multiple metrics, highlighting the value of heterogeneous information in affinity prediction.
Introduction
Predicting Drug-Target Affinity (DTA)—the strength of the binding between a drug molecule and a target protein—is a critical component of modern drug discovery. Traditional experimental methods are slow and costly, making computational (in silico) prediction essential for rapid screening.
Existing computational models often suffer from "information tunnel vision," representing molecules through a single lens, such as SMILES strings, 2D graphs, or hand-crafted fingerprints. This approach ignores the multi-faceted nature of molecular behavior, which is governed by local chemical environments, global topology, and sequential connectivity. MolHF addresses this by fusing these diverse "views" to create a more holistic molecular representation.
Methodology
MolHF introduces two primary modules to capture and fuse heterogeneous molecular data:
1. Molecular Heterogeneous Information Learning (MHIL)
This module extracts three distinct feature types from drug molecules:
- Atom Association: Utilizes Graph Neural Networks (GCN, GAT, or GIN) to capture local sub-graph patterns and the biochemical properties of neighboring atoms.
- Molecular Topology: Employs a "global storage node" strategy to represent the overall 3D structure and shape of the molecule.
- Atoms-Bonds Arrangement: Combines Convolutional Neural Networks (CNNs) and Bidirectional Gated Recurrent Units (BiGRU) to process the sequential "language" of the molecule via SMILES strings.
2. Molecular Heterogeneous Attention Fusion (MHAF)
Rather than simple concatenation, MolHF employs an attention mechanism to weigh these features. This allows the model to dynamically prioritize the most relevant structural information for a specific drug-target pair while filtering out redundant data.
On the protein side, the framework uses a CNN + BiLSTM architecture to capture both local amino acid patterns and long-range dependencies. These drug and protein representations are then integrated into a Fully Connected Network (FCN) to predict the final affinity score.
Key Results and Discussion
MolHF was evaluated on the Davis and KIBA benchmark datasets using four key metrics: Mean Squared Error (MSE), Concordance Index (CI), $r_m^2$, and Pearson Correlation (PPC).
- Performance Gains: On the Davis dataset, MolHF achieved an MSE of 0.208, significantly outperforming GraphDTA (0.229) and DeepDTA (0.261). On the KIBA dataset, it reached a CI of 0.904, demonstrating superior reliability in ranking drug-target pairs.
- Impact of Attention: Ablation studies confirmed that the MHAF module is crucial for performance. By intelligently weighting features, the model avoids the "information redundancy" that can occur when simply mashing different data types together.
- Protein-Drug Synergy: A notable finding was that the quality of the protein embedding directly influences the model's ability to select relevant drug features. A robust protein representation enables the attention mechanism to more effectively identify which molecular attributes of the drug are critical for binding.
Conclusion
MolHF demonstrates that the diverse gains brought by heterogeneous information significantly contribute to the accuracy of drug-target affinity prediction. By fusing local, global, and sequential molecular attributes through an attentive mechanism, the framework provides a more comprehensive understanding of molecular interactions. Future work will explore more advanced graph architectures and the direct interaction between specific chemical functional groups and amino acid fragments.
Summary updated