Prediction workflows
Pairwise Interaction Prediction
- Human and viral protein sequences
Enter the human and viral protein sequences in their respective fields.
- ESM-2 + AAindex
Represent each protein using ESM-2 features, the saved projection, and AAindex physicochemical features.
- Predict the interaction
Calculate the interaction probability and assign a label using the model threshold.
Interaction Partner Sequence Prediction
- Input protein sequence
Enter one sequence and specify whether it is a human or viral protein.
- Predict a partner embedding
Use the input protein's ESM-2 and AAindex features to predict an embedding for a potential interaction partner.
- Retrieve natural partner sequences
Rank natural proteins by cosine similarity to the predicted partner embedding. Human inputs return viral candidates; viral inputs return human candidates.
The two tasks can be used independently. Pairwise prediction requires both protein sequences; partner prediction requires one.
Protein representations
ESM-2 produces sequence representations using attention mean pooling that includes special tokens. A saved projection reduces the representation from 320 to 256 dimensions. The input limit is 1,024 tokens in total, including BOS/EOS (normally up to 1,022 residues).
Physicochemical features
AAindex describes physicochemical properties of the first 100 residues. Under the encoder rules, ambiguous or non-standard residues receive zero AAindex vectors.
Classification
The classifier uses sequence representations and physicochemical features from both proteins. It returns the interaction probability, predicted label, classification threshold, and model version.
Natural sequence retrieval
The model uses ESM-2 and AAindex features from the input protein to predict a partner embedding. It then retrieves natural sequences by cosine similarity to this embedding. Prediction and retrieval form a single task; no amino-acid sequence decoding is required.
Limitations
Truncation excludes parts of the sequence from feature extraction. The fixed natural libraries cover the complete specified NCBI RefSeq and UniProtKB sources; they are finite snapshots and may overlap with proteins used for training or evaluation. Similarity is not a calibrated interaction probability. Predictions require experimental validation.