Two task workflows
Interaction Prediction
- A human–virus pair
Keep the human and viral sequences on their assigned sides.
- ESM-2 + AAindex
Encode each sequence with the fixed projection and physicochemical features.
- Classify the pair
Report probability, prediction and the configured threshold.
Natural Sequence Retrieval
- One query sequence
Choose whether the query is a human or viral protein.
- Fixed ESM-2 encoding
Encode the query directly, using the same representation as the reference index.
- Search the opposite library
Rank existing natural sequences by cosine similarity: human → viral, viral → human.
To estimate a candidate pair's interaction probability, load it into Interaction Prediction and submit a separate classification task.
Protein representations
The fixed server encoder uses ESM-2, attention mean pooling including special tokens, and a saved projection from 320 to 256 dimensions. ESM input is limited to 512 tokens including special tokens.
Physicochemical features
AAindex features represent the first 100 residues. Ambiguous or non-standard residues have zero AAindex vectors under the fixed encoder rules.
Classification
Human and viral representations and physicochemical features enter the fixed classifier. Results report probability, the applied threshold, the predicted label, and model version.
Natural sequence retrieval
Encode one query with the same versioned server encoder and compare it with the opposite natural library using cosine similarity. This is retrieval, not the paper's full generative task.
Limitations
Truncation omits sequence information. Training-derived libraries have limited coverage and possible overlap with evaluated proteins. Similarity is not calibrated interaction probability. Predictions require experimental validation.