DynamicDTA: Drug-Target Binding Affinity Prediction Using Dynamic Descriptors and Graph Representation

TL;DR

DynamicDTA tackles the challenge of static-protein biases in DTA prediction by integrating MD-derived protein dynamics with graph-based drug representations. The framework combines ligand graphs, dilated-protein sequence encodings, and dynamic descriptors through cross-attention and a Tensor Fusion Network, yielding improved binding affinity predictions. Across three curated datasets and an external Kiba$^*$ validation, DynamicDTA achieves notable gains in $e_{ m RMSE}$ and $R$, and provides interpretable insights via attention visualizations and HIV-1 case studies. This work highlights the practical impact of dynamic protein information for drug discovery and points to future avenues to reduce MD-data requirements, such as diffusion-model-based MD data generation.

Abstract

Predicting drug-target binding affinity (DTA) is essential for identifying potential therapeutic candidates in drug discovery. However, most existing models rely heavily on static protein structures, often overlooking the dynamic nature of proteins, which is crucial for capturing conformational flexibility that will be beneficial for protein binding interactions. We introduce DynamicDTA, an innovative deep learning framework that incorporates static and dynamic protein features to enhance DTA prediction. The proposed DynamicDTA takes three types of inputs, including drug sequence, protein sequence, and dynamic descriptors. A molecular graph representation of the drug sequence is generated and subsequently processed through graph convolutional network, while the protein sequence is encoded using dilated convolutions. Dynamic descriptors, such as root mean square fluctuation, are processed through a multi-layer perceptron. These embedding features are fused with static protein features using cross-attention, and a tensor fusion network integrates all three modalities for DTA prediction. Extensive experiments on three datasets demonstrate that DynamicDTA achieves by at least 3.4% improvement in RMSE score with comparison to seven state-of-the-art baseline methods. Additionally, predicting novel drugs for Human Immunodeficiency Virus Type 1 and visualizing the docking complexes further demonstrates the reliability and biological relevance of DynamicDTA.

Figures (6)

• Figure 1: The architecture of DynamicDTA. The framework integrates drug molecular graph, target sequences and dynamic descriptors, and a cross-attention mechanism to extract meaningful representations. TFN is employed to effectively fuse the extracted features for accurate drug–target binding affinity prediction.
• Figure 2: The performance of DEAttentionDTA (a) and DynamicDTA (b) on three datasets for the prediction of DTA.
• Figure 3: Affinity distribution comparison across datasets: (a) ${\mathrm{IC}_{50}}^*$ dataset. (b) ${K_{\mathrm{d}}}^*$ dataset. (c) ${K_{\mathrm{i}}}^*$ dataset.
• Figure 4: Parameter sensitivity analysis in the ${K_{\mathrm{i}}}^*$ dataset: (a) Dropout rate $P$. (b) Dilation rate $D$. (c) Attention heads $H$. (d) GCN hidden layers $L$.
• Figure 5: Visualization of drug-target interactions in the 2FOS complex, which consists of two distinct ligand-binding regions. The model's attention weights for the top 20 residues are highlighted, where correctly identified binding residues are shown in red, and misidentified residues are shown in cyan. The two smaller panels on the right provide zoomed-in views of the two distinct ligand-binding regions from the left panel. These views are rotated to optimal angles to offer a clearer perspective of the binding interactions.