Host-aware Estimation of antigen immunogenicity and Ranking through Deep-learning.
A pan-species tool to predict immunogenic proteins across a wider animal world.
Pick the path that fits your workflow — from checking a handful of sequences to scoring whole proteomes.
Paste your protein sequences, choose a host, and read the immunogenicity scores in minutes — straight in your browser.
No-code interfaceRun HERD in the cloud on a ready-made notebook — score a proteome or a shortlist without touching your own hardware.
Low-codeInstall HERD on your machine or cluster to score thousands of proteins and wire it into your own vaccinology pipeline.
CodeHERD models protein immunogenicity as a host-dependent prediction problem rather than assigning a single host-agnostic score to every antigen.
Read the full tutorial ›Protein-level immunogenicity datasets are resolved by host species. Core datasets combine manually curated experimental evidence with B-cell records mined from IEDB, while negatives are sampled from the same source taxa and matched by sequence length after homologous positives are removed.
Each protein is encoded with the 650M-parameter ESM-2 protein language model. Final-layer residue embeddings are mean-pooled into a 1,280-dimensional representation used as input for downstream prediction.
The integrated model uses a shared neural-network trunk followed by a dedicated sigmoid output head for each host. Joint training lets hosts share cross-species immunogenicity signal while retaining host-specific decision functions.
Performance is estimated with homology-aware nested cross-validation so homologous proteins cannot leak across folds. Leave-one-pathogen-out evaluation additionally tests transfer to antigen repertoires from previously unseen pathogen genera.
Manuscript in preparation · citation details will be updated after publication.
For questions about HERD — the models, the data, or running the tool — reach out by email.
