DeepAVC

DeepAVC is a framework for finding small molecules with broad-spectrum antiviral activity, built by Boming Kang, Qinghua Cui and colleagues at Peking University. Antiviral drug discovery splits into two traditions: phenotype-based screening, which asks whether a compound suppresses viral replication in cells without specifying a mechanism, and target-based screening, which asks whether a compound binds a particular viral or host protein. Existing computational tools have concentrated almost entirely on the second, leaving the phenotypic signal — the property clinicians actually care about — unmodelled.

DeepAVC covers both. DeepPAVC takes a compound alone and returns an antiviral activity score; DeepTAVC takes a compound and a protein sequence and returns an interaction score. Both sit on frozen pretrained backbones rather than learning molecular representations from the small antiviral datasets available: compounds are encoded by KPGT, a pretrained molecular graph transformer, and proteins by ESM-2. A third component, CADTI, generalizes the target-based head to compound-protein interaction prediction outside the antiviral setting.

Posted as a preprint in February 2025 and published in Communications Biology in 2026, the work is notable less for architectural novelty than for closing the loop: predictions on unseen compounds were carried through to wet-lab testing, and one of the nominated molecules held up.

#Key Features

• Two complementary prediction modes: A phenotype head scores intrinsic antiviral activity from structure alone, while a target head scores compound-protein binding; the paper shows the two are complementary and can be used together to triage candidates.
• Frozen pretrained encoders: Compound and protein representations come from models pretrained on far larger corpora than any antiviral dataset, so the trained heads operate on general-purpose chemical and protein features.
• Cross-modal attention: The target-based model applies self-attention within each modality and cross-attention between them, capturing interaction patterns rather than concatenating independent embeddings.
• Atom- and residue-level interpretability: Attention weights identify the functional groups on a compound and the residues on a target that drive a predicted interaction, giving medicinal chemists a starting point for optimization.
• Experimentally validated output: MNS and NVP-BVU972 were selected from top predictions among compounds unseen during training and tested experimentally, with MNS confirmed as a broad-spectrum antiviral in both cell-based and animal experiments.

#Technical Details

DeepPAVC feeds SMILES-derived molecular graphs through the pretrained KPGT encoder and a prediction head that outputs an antiviral activity score, using only compound information as input. DeepTAVC extends this to two modalities: the same molecular encoder for the compound and ESM-2 for the protein target, followed by self-attention within each stream and cross-attention between them before a scoring head. CADTI shares the dual-encoder design but is trained for general compound-protein interaction rather than antiviral activity specifically. The published evaluation reports that DeepAVC outperforms existing baselines on antiviral compound prediction, and that MNS shows greater efficacy than the RSV fusion inhibitor Sisunatovir. Pretrained weights for all three models, together with the KPGT and ESM-2 feature extractors, are distributed from the authors' institutional web server rather than a standard model hub; the code repository carries an MIT license and provides training and inference notebooks for each component.

#Applications

The framework is aimed at the early triage stage of antiviral drug discovery, where a large compound library must be narrowed before any cell-based assay is run. The phenotype head is useful when no specific target has been chosen — for a newly emerging virus, or when host-directed activity is the goal — while the target head supports campaigns organized around a known viral protein or host factor. Running both and looking for agreement gives a stricter filter than either alone. Beyond virology, the CADTI component addresses generic compound-protein interaction prediction, and the hosted web server allows groups without deep learning infrastructure to submit compounds and retrieve scores.

#Impact

DeepAVC's contribution is the demonstration that phenotype-level antiviral activity, not just target binding, is learnable from pretrained chemical representations, and that the resulting predictions survive contact with the bench. Progressing a nominated compound through in vitro and in vivo confirmation is a stronger validation than the retrospective benchmarks that dominate this literature. The practical caveats are real: pretrained weights depend on the longevity of a single institutional server rather than an archival host, the code repository has seen limited external adoption, and the interpretability claims rest on attention weights, which indicate model attribution rather than a physical binding mode. Publication in Communications Biology and the release of both code and hosted inference make the framework straightforward to evaluate against alternatives.