Forecasting SARS-CoV-2 spike protein evolution from small data by deep learning and regression.

Area of Science:

• Virology
• Computational Biology
• Genomics

Background:

• SARS-CoV-2 variants have driven global outbreaks, complicating pandemic control.
• Predictive models for viral evolution require substantial data, a limitation for emerging viruses.

Purpose of the Study:

• To develop a model for predicting viral protein evolution using sparse data.
• To address computational challenges in modeling discrete mutations by encoding sequences into continuous numerical representations.

Main Methods:

• Developed a viral protein evolution prediction model (VPRE) combining variational autoencoders (VAE) and Gaussian process (GP) regression.
• Encoded discrete amino acid sequences into continuous numbers using a VAE.
• Modeled evolutionary trajectories using GP regression on the continuous numerical data.

Main Results:

• VPRE successfully predicted SARS-CoV-2 spike protein evolution up to 5 months using 104 sequences.
• Predictions included novel variants, with the most frequent prediction differing by a single amino acid from known dominant sequences.
• Predicted novel variants in the spike receptor binding domain (RBD) showed strong binding to human ACE2 in silico.

Conclusions:

• Combining deep learning and regression enables effective viral protein evolution modeling with sparse datasets.
• The VPRE model demonstrates utility in anticipating future viral variants and their potential impact.
• This approach supports the development of more effective medical interventions against rapidly evolving viruses.