MoFeDTA: A Dual-Channel Drug-Target Affinity Prediction Network Integrating Morgan Fingerprints and Transformer-TextCNN Protein Modeling

Estimating the binding affinity between drug molecules and target proteins is a pivotal task in early-stage drug discovery and high-throughput screening. To improve the modeling of molecular structures and protein sequences, this paper presents MoFeDTA, a dual-pathway architecture that integrates Morgan fingerprint-based structural representation with Transformer and TextCNN-driven sequence encoding. In the molecular representation stream, a Transformer encoder is employed to capture global dependencies from the binary fingerprint vectors. In parallel, the protein stream utilizes a multi-kernel TextCNN network to extract localized features from amino acid sequences in FASTA format. The outputs from both channels are fused and passed into a regression layer to predict the interaction affinity. Experiments on the Davis and KIBA datasets reveal that MoFeDTA achieves stable and competitive performance across metrics such as CI, MSE, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$r_{m}^{2}$</tex>. Notably, the model attains an MSE of 0.149 and an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$r_{m}^{2}$</tex> of 0.777 on the KIBA benchmark, demonstrating strong predictive capability. Ablation studies further validate the individual effectiveness of each module, highlighting their complementary roles in feature learning. Collectively, MoFeDTA provides a robust and extensible framework for accurate affinity prediction in computational drug development.