Combinatorial factorizable libraries outperform enumerated and random libraries for antibody discovery

The effective use of high-throughput functional assays to discover novel biological therapeutics requires a diverse library of desirable candidates. Here we design candidate antibody libraries using factorizable neural networks (FNNs) that permit the economic synthesis of factorizable libraries that are depleted of non-specific binders. We experimentally evaluated an FNN designed antibody CDR-H3 factorizable library that is designed to contain binders to peptide-MHC (pMHC) complexes with depletion of library sequences that exhibit non-specific pMHC binding. The factorizable library comprises 104 prefixes that were randomly combined with 104 suffixes. We find that it contains more viable candidates than designed libraries with the same DNA synthesis budget on our pMHC binding task. FNNs make it possible to efficiently computationally design combinatorial libraries, a task that is not possible with conventional methods.