A Task-Adaptive Multimodal Pretrained Framework for Antibiotic Virtual Screening with Joint Activity and Cytotoxicity Prediction

Abstract

Antimicrobial resistance (AMR) poses a growing threat to global public health, particularly for priority pathogens such as Staphylococcus aureus (SA) and Neisseria gonorrheae (NG), for which therapeutic options are increasingly limited. Although deep learning has accelerated antibiotic discovery, existing approaches are often constrained by single-modality molecular representations and a lack of joint modeling for antibacterial activity and host cell toxicity. Here, we present MAPViS, a task-adaptive multimodal pretrained framework for jointly predicting antibacterial activity against SA and NG together with cytotoxicity risk in three human cell lines (HepG2, HSkMC, and IMR-90). MAPViS integrates molecular graph representations, molecular fingerprints, and descriptor/3D geometric features through a gated fusion mechanism and combines large-scale self-supervised pretraining with supervised task-specific fine-tuning. The graph encoder is pretrained on approximately 10 million molecules using hierarchical pseudolabel classification and masked graph contrastive learning to improve representation quality and transferability. On benchmark datasets derived from experimentally screened compounds, MAPViS achieves competitive performance across antibacterial activity prediction tasks and shows advantages in early recognition metrics relevant to virtual screening, including AUPRC, Precision@30, and EF@1.0%. External validation further shows strong top-ranked enrichment on the SA task, where all Top-10 candidates are experimentally active, whereas performance on NG is more modest, indicating task-dependent generalization. Overall, MAPViS provides a scalable multimodal framework for joint activity–toxicity modeling and offers a practical computational tool for antibiotic virtual screening and candidate prioritization.