Low-Rank Approximation Based Multimodal DTI Prediction Model with Hypercomplex Representations

Predicting drug-target interactions (DTI) plays a crucial role in accelerating drug discovery, repurposing existing drugs, and identifying potential adverse effects, thereby reducing both time and cost. A key factor in DTI prediction is the diverse encoding of drugs and target proteins. However, existing multimodal DTI models fail to utilize the benefits brought by hypercomplex embedding representations, or they only focus on exploiting the advantages of low parameter complexity, while ignoring such representations could also bring gains including high integration precision and robustness to interferences. To fully leverage such merits, a low-rank approximation based multimodal DTI prediction model integrating hypercomplex principal component analysis is proposed to enhance noise robustness, boosting the performance in the multimodal data fusion scenario. Furthermore, with hypercomplex tensor singular value decomposition (SVD) and improved computational efficiency, our approach also supports environmentally sustainable practices in large-scale biomedical computation. Experiments on public benchmark datasets demonstrate that our approach outperforms existing SOTA methods, e.g., 2.6 % improvement at MSE on KIBA over the baseline model.