Summary
A bioRxiv preprint describes a transformer model trained on 8.8 million measured SARS-CoV-2 RBD–ACE2 interactions. It predicts receptor-binding compatibility across animal species and pairs predictions with uncertainty estimates to help prioritise surveillance.
A bioRxiv preprint describes a platform that combines laboratory screening with deep learning to identify animal species whose ACE2 proteins may bind SARS-CoV-2. The researchers used measurements from 15 ACE2 orthologs—corresponding proteins in different species—to train a transformer model that predicts receptor-binding compatibility from sequence alone. The work is aimed at helping prioritise animals for surveillance of reverse zoonosis, the transmission of a pathogen from humans to animals.
The study focuses on the virus’s receptor-binding domain (RBD), the part of its spike protein that binds ACE2, a cell-surface receptor used by SARS-CoV-2 to enter cells. The researchers describe their work as a preprint; the abstract reports a laboratory screening and computational modelling study.
How the screening and model work
The team displayed a library of SARS-CoV-2 RBD variants on yeast and screened it against ACE2 proteins from 15 diverse species. They then deep-sequenced populations that bound to ACE2 and those that did not. The resulting dataset contains 8.8 million unique RBD:ACE2 interaction measurements, according to the preprint.
A transformer-based model was trained on these measurements to predict whether an RBD and an ACE2 sequence are compatible. The researchers report that it can also make “zero-shot” predictions for ACE2 orthologs not included in training. In this context, zero-shot means applying the model to a receptor sequence it was not trained on, rather than measuring that specific pairing in the screening experiment.
The platform pairs predictions with a distance-based uncertainty estimate. That gives researchers a way to consider both predicted compatibility and confidence when sorting through candidate species. The authors say this could help triage hundreds of animals for SARS-CoV-2 surveillance.
What the predictions can guide
Reverse zoonotic infections matter because an animal population could potentially become a reservoir in which a virus continues to circulate and evolve, creating a route for later transmission back to people. Prioritising species for surveillance could help direct attention towards animals whose ACE2 sequences are predicted to be compatible with the virus.
The readout here is receptor binding: it provides a molecular signal for choosing candidates to investigate. Binding compatibility alone does not determine whether a species can be infected or sustain transmission; those outcomes also involve biological processes beyond the RBD–ACE2 interaction. The predictions are therefore best understood as a way to guide follow-up surveillance, not as confirmation that a particular animal species is a reservoir.
The study demonstrates the approach for SARS-CoV-2 RBD and ACE2. The authors present it as a broader strategy for anticipating viral host range, but applying it to other viruses would require suitable experimental interaction data and validation for the relevant virus–host system.