Shanghai Artificial Intelligence Lab, in collaboration with Fudan University, Ruijin Hospital, School of Medicine, Shanghai Jiao Tong University, and the Shanghai Institute for Viral Research, has launched a new AI predictive model called ViraHInter. The release of this model marks a new stage in the development of antiviral drugs, as it can predict how viruses "hijack" human proteins without the need for wet experiments.

image.png

ViraHInter is an AI model that integrates dual-modal information from protein sequences and structures. Traditional methods for predicting protein interactions usually analyze only amino acid sequences or three-dimensional structures, but the breakthrough of ViraHInter lies in its ability to accurately capture both types of information simultaneously. This model generates full-atom three-dimensional structures of virus-host protein complexes, detailing the relationships between each atom, laying the foundation for drug design. At the same time, it uses a protein language model to identify conserved patterns that remain even as viruses rapidly mutate, thereby improving the accuracy of predictions.

In a series of benchmark tests, ViraHInter has shown remarkable performance. Its prediction accuracy for virus-human protein interactions reached 0.50, which is 4.5 times higher than that of AlphaFold3, far exceeding other predictive methods. The research team also used this model to analyze three influenza subtypes, successfully identifying 33 shared host factors, demonstrating the great potential of ViraHInter in the development of antiviral drugs.

More importantly, when facing newly emerging viruses, ViraHInter has shown strong adaptability. In tests where sequence homology was strictly controlled, the model still performed excellently, proving its broad application potential in dealing with new pathogens. This discovery will provide new targets and directions for the development of drugs against influenza and coronaviruses.