A Gene Variant-Based and Proteomic-Informed Contrastive Cross-Modal Ensemble Deep Attention Approach for Antithrombotic Candidate Selection and Rationalisation

Background: Genetic variations such as single nucleotide polymorphisms (SNPs) as part of pharmacogenomics play an important role in the metabolism of drug and hence their active concentrations in blood plasma. Objectives: The aim of this study is to select candidate compounds from a TCM dataset that may be repurposed for arterial and venous thromboses management. This shall be achieved through development and evaluation of an ensemble deep learning model that taking into account the genetic variations in protein sequences. Methods: BIOSNAP dataset was supplemented with 321,657 drug–target pairs consisting of SNP variants of wild-type proteins. The application dataset consisted of a TCM dataset containing 35,553 ingredients. The control group was set as the pathogenic group, whilst the treatment group was set as the non-pathogenic group. Contrastive and non-contrastive deep cross-modal attention ensemble modelling was developed, evaluated and applied. Results: Contrastive regularisation effect improved the performance of the Contrastive Learning (CL) Ensemble over the Non-CL Ensemble model as well as the Dong et al. (May 2025) CL model in the test set (AUPR 0.919 vs. 0.894 vs. 0.813). Safflower yellow A, Paeoniflorin and Notoginsenoside R6 were associated with existing TCM and highly ranked for interaction with Factor Xa genetic variants. Highly interacting protein targets were identified. Conclusions: Ensemble modelling with contrastive learning resulted in performance improvements and can be useful for selecting TCM compounds for antithrombotic management. This is a step towards personalised drug selection and can simultaneously facilitate interpretation of the biological rationales for risk vs benefit evaluations during decision making.