Forecasting a virus's next move: antiGen shortlists mutations from the evolutionary tree, and Germinal Multi redesigns antibodies against them
Synopsis
Monoclonal antibodies were an important early tool against COVID-19, but they lost effect as SARS-CoV-2 mutated, and screening for antibodies that hold across variants struggles to keep up. antiGen learns from mutations already made along a virus's phylogenetic tree rather than from sequences alone, using roughly half a million parameters to forecast plausible next mutations; Germinal Multi tests one antibody design against several versions of a target at once on a computer. Combined, they redesigned BD55-1205, which neutralized XFG but not XFG-Next, into X51, which blocked infection by both in laboratory tests.
Interpretation
Viral mutation forecasting shifts from sequences alone to evolutionary history: antiGen learns directly on phylogenetic trees how mutations followed one another, so it can point to mutations not yet observed. Simply counting how often mutations had already occurred was highly predictive of what came next once a pandemic had run a few months, but a count can only point to mutations already seen. Trained on data through 2023, about 7.8 million sequences spanning 240,000 mutations, antiGen predicted future mutations more accurately overall than the other methods compared, with its largest lead on mutations never in the training data and accessible through a single nucleotide change.
Antibody design moves from one version of a target to requiring, on a computer, that a single design bind several versions before it is kept. Germinal had designed antibodies against a single target, while an antibody tuned to one version can fail on the next. Designed from scratch against the site sipavibart binds on both BA.2 and KP.3, the best of 57 tested designs, R50, blocked both variants from infecting cells and also bound variants left out of the design.
Forecasting and design can be chained: narrow the virus's plausible changes first, then redesign an existing antibody against them. BD55-1205 stops XFG from infecting cells, but combining three individually tolerated substitutions, A475V, E493R and Y501T, into one spike produced XFG-Next, completely insensitive to its neutralization. Of 60 redesigned antibodies, the lead X51 blocked infection by both XFG and XFG-Next, cutting infection in half at 2.1 ng/mL and 6.3 ng/mL respectively; another design, X58, bound both variants but blocked only XFG.
Perspective
The demonstrated scope is computational forecasting plus pseudovirus experiments: XFG-Next was built only in replication-incompetent pseudoviruses under approved institutional biosafety procedures at BSL-2 containment, and it was not tested for transmission or disease; the antibody results are cell-infection experiments, not clinical efficacy.
antiGen cannot yet say which circulating variant will outcompete the others, because there are far fewer cases where one variant conclusively beat another than there are spike mutations to learn from, and the answer may need a different kind of model; early in a pandemic with scarce data, a simple count of past mutation frequency can beat antiGen at picking the very top predictions, so a hybrid leaning on raw frequency was built; and what antiGen has learned when it predicts never-observed mutations remains to be understood.
