Germinal Multi fits one antibody sequence to many antigen variants, and its BA.2/KP.3 campaign yielded a dual-neutralizing antibody after one screening round
Synopsis
Antibodies often lose potency as their targets evolve, and earlier de novo design methods optimized a single paratope against one antigen structure at a time. Germinal Multi frames design of one antibody sequence against multiple antigen variants as multi-objective optimization; on cancer and viral antigens it produced antibodies that bound resistance-associated EGFR variants or neutralized both BA.2 and KP.3, with recognition extending to variants withheld from design. It also combined evolutionary forecasting with deep mutational scanning to redesign BD55-1205, restoring neutralization of the engineered escape variant XFG-Next while preserving activity against parental XFG.
Interpretation
The design target moves from a single antigen structure to a set of antigen variants, with one paratope required to fit all of them. Earlier de novo design methods designed against one antigen structure at a time and did not jointly optimize a single paratope against multiple variants of a shared epitope. For EGFR, co-optimization against wild-type plus a composite challenge antigen carrying six resistance-associated substitutions gave 60 designs, of which E2, E16, E18 and E26 exceeded the 0.5 nm binding-response threshold against wild-type and all six single-substitution variants.
The pipeline makes one shared antibody sequence satisfy structural confidence and sequence likelihood on several target complexes at once. Structure prediction and sequence redesign are no longer computed for one target but evaluated per target and then combined. Each optimization step runs AlphaFold-Multimer separately per target and merges the gradients, sequence design samples one sequence across multiple structures with AbMPNN, and a cycle is accepted only if every target passes the filters.
Evolutionary forecasting can name an escape direction that has not yet appeared and guide redesign of an existing antibody. Earlier antibody redesign acted on observed variation rather than on substitutions proposed by a forecasting model. antiGen-predicted A475V, E493R and Y501T overlapped escape sites in a KP.3.1.1 deep mutational scanning landscape; combined into XFG-Next they abolished measurable BD55-1205 neutralization, and the redesigned X51 reached IC50 values of 2.1 and 6.3 ng/mL against XFG and XFG-Next.
Perspective
The demonstrated scope is antigenic variants of the same protein: six resistance-associated substitutions in EGFR domain III, SARS-CoV-2 receptor-binding-domain variants BA.2 and KP.3, and the engineered XFG-Next; each campaign tested tens of designs experimentally.
Whether XFG-Next could emerge or spread in nature remains uncertain; X51 neutralized the A475V and E493R single-substitution pseudoviruses but showed no detectable neutralization of Y501T within the tested range; and how the choice of design targets shapes breadth against variants not included is still an open question.
