Antiviral drug discovery targeting viral proteases often begins with structure-based virtual screening, but docking scores alone are unreliable predictors of true biochemical inhibition because they oversimplify complex binding, solvation, and conformational effects. Existing machine learning models typically treat molecular activity as a direct regression or classification target, which can overlook the relational structure among compounds, particularly when docked molecules share interaction patterns yet differ in measured bioactivity. To address these limitations, this article proposes a contrastive molecular learning framework for antiviral hit prioritization, designed to learn an embedding space in which active antiviral compounds cluster near structurally and interactionally similar inhibitors. The model integrates molecular graph encoders, docking-derived interaction fingerprints, protease pocket features, docking scores, and bioactivity labels, using a contrastive objective to bring together compounds with similar activity and interaction profiles while separating inactive or dissimilar molecules. Conceptually, this approach generates a ranked virtual library in which likely antiviral hits are distinguished from docking false positives, with latent space visualization and substructural attribution providing qualitative insight into compound prioritization. By combining binding plausibility with activity-consistent molecular representations, this contrastive framework could enhance the practical utility of virtual screening against viral proteases and reduce unnecessary biochemical testing.