Prediktering av protein-DNA-bindningsaffinitet med AlphaFold-inbäddningar och graf-attentionsnätverk

Predicting the binding affinity between proteins and DNA is a fundamental challenge in molecular biology, crucial for understanding gene regulation and the molecular basis of diseases. Conventional methods often struggle with the complexity of these interactions and the limited availability of structural data. This thesis explores a novel deep learning framework for predicting protein-DNA binding affinity by leveraging structural embeddings generated by AlphaFold 3, a state-of-the-art protein structure prediction model. The proposed approach represents protein-DNA complexes as fully connected graphs, where amino acid and nucleotide residues are nodes and their potential interactions are edges. AlphaFold’s single representation serves as node features, and the pair representation provides edge features, eliminating the need for handcrafted features. The study contains a comparative analysis of different Graph Neural Network (GNN) architectures, including Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and Graph Transformers. Models were evaluated on three distinct datasets: two high-throughput HiTS-FLIP datasets (KLF1 and LacI) and the more diverse PDBbind dataset. The results demonstrate that attention-based architectures, particularly GAT, outperform the GCN baseline on the HiTS-FLIP datasets, achieving a Pearson Correlation Coefficient (PCC) of up to 0.94 after hyperparameter optimization on the KLF1 dataset. However, this performance advantage was less pronounced on the heterogeneous PDBbind dataset. Our analysis revealed that model performance is more sensitive to node features than to edge features, suggesting that current GNN architectures may not fully utilize the rich interaction information in AlphaFold’s pair embeddings. While the models show strong predictive power on specific datasets, experiments testing generalization across datasets indicate that further work is needed to overcome domain shift and improve applicability to unseen protein-DNA complexes.