Arc Institute redesigned an antibody to target a coronavirus variant with three mutations predicted by a neural network, and it blocks infection in pseudovirus experiments
Arc Institute redesigned an antibody to target a coronavirus variant with three mutations predicted by a neural network, and it blocks infection in pseudovirus experiments
On November 17, Brian Hie’s group at Stanford University and Arc Institute described two preprints: one on antiGen, a neural network that predicts amino acid substitutions in the coronavirus spike protein, and another on Germinal Multi, a system that designs an antibody to target several protein variants. In experiments with pseudoviruses (particles that infect cells but cannot replicate), the redesigned antibody blocks infection by XFG and by XFG-Next, a version of XFG with three predicted substitutions added.
The coronavirus attaches to cells through its spike protein. A therapeutic antibody recognizes a specific site on that protein, which the virus can change. AstraZeneca’s antibody sipavibart received a positive opinion from the European Medicines Agency in December 2024. That same month, the agency’s task force reported that a single amino acid substitution prevented sipavibart from blocking infection by KP.3, then the predominant variant in Europe.
Hie’s group found that just a few months into the pandemic, counting how often each mutation had occurred already provided good predictions of subsequent mutations, though only those seen before. To predict new mutations as well, they trained antiGen on a phylogenetic tree built from millions of spike protein sequences. Each branch records which substitution arose and in which ancestral version of the protein. According to the authors, a prediction “can narrow thousands of possibilities to a short list of plausible ones, giving antibody developers enough to prepare for.”
The model was trained on data collected before January 1, 2023, months before BA.2.86 emerged, and tested on the sequence of mutations from BA.2.86 to KP.3, against which sipavibart lost activity. It ranked six of the seven substitutions in the top 10% of possible substitutions. The highest-ranked substitutions were introduced individually into an XFG pseudovirus to test whether it remained infectious. Six of the ten most likely substitutions yielded infectivity equal to or greater than that of XFG, as did four of the ten most likely substitutions that had never been observed in five years of the pandemic.
The Germinal system designed an antibody to target a specified site on a single protein variant, while Germinal Multi retains an antibody design only if it is predicted to bind all specified variants. The authors tested 57 antibody designs targeting the site recognized by sipavibart. These were designed for both BA.2, against which sipavibart had worked, and KP.3, against which it had stopped working. The lead design blocked infection by both. For comparison, sipavibart with five substitutions selected using other methods failed to block infection by either KP.3 or BA.2.
To test whether an antibody could be prepared in advance, the authors took BD55-1205, an antibody active against many variants, including the then-predominant XFG, and constructed a variant that escaped it. Previously published measurements on a closely related viral variant showed that substitutions at six positions in the spike protein weakened this antibody’s activity. antiGen predicted substitutions at three of those positions, each of which left the pseudovirus infectious. The authors combined all three substitutions in XFG-Next. It remained infectious, although less so than XFG, and BD55-1205 failed to block infection by XFG-Next.
Germinal Multi redesigned BD55-1205 using its sequence as the starting point. X51 emerged as the lead among 60 candidates: it blocks infection by both XFG and XFG-Next. Another candidate bound both variants but blocked infection only by XFG. The authors write: “Binding alone did not tell us which designs would work.”
- https://arcinstitute.org/news/antigen-and-germinal-multi
- https://www.biorxiv.org/content/10.64898/2026.08.27.741102v2
- https://www.biorxiv.org/content/10.64898/2026.11.10.765734v2
- https://www.ema.europa.eu/en/documents/other/etf-statement-loss-activity-anti-spike-protein-monoclonal-antibodies-due-emerging-sars-cov-2-variants-december-2024-update_en.pdf
- https://t.me/UkhvatNews/1961