Today's Overview
- Synthetic Epitope Atlas couples yeast-based profiling to 26M affinity values for VHH binder training AlphaSeq yeast profiling produced 26 million affinity measurements linking computed VHH–SEP structures to experimental binding.
Featured
01Synthetic Epitope Atlas couples yeast-based profiling to 26M affinity values for VHH binder training
Antibody design algorithms are starved for paired structural and binding data, limiting their ability to rank de novo candidates. SEPIA tackles this by computationally designing thousands of mini “synthetic epitope proteins” (SEPs) meant to bind VHH single-domain antibodies, then experimentally testing every interaction. A yeast surface-display platform (AlphaSeq) quantified 26 million on- and off-target affinities for these VHH–SEP pairs and their mutational variants, creating a resource that links computed pseudo-structures to measured binding.
Of the designed complexes, 1,161 showed strong, specific binding and >75,000 mutants were profiled, giving a mutational landscape for model training. When added to existing antibody structure databases, the SEPIA pseudo-structures improved ML ranking of de novo designs beyond conventional confidence metrics. The entire workflow is scalable: designs are generated in silico and binding is read out in a single high-throughput yeast assay, avoiding the bottlenecks of individual protein purification.
Limitations include that validation is yeast-based (no in vitro or in vivo confirmation) and success is reported only for VHH binders, leaving unclear generalizability to full-length antibodies or diverse antigen classes.
Today's Observation
Yeast-based profiling now delivers affinity data at true proteomic scale: 26 million VHH–SEP measurements map every point-mutant in 1,161 de-novo complexes and give machine-learning models residue-level binding labels instead of the usual “bind / no-bind” calls. By folding these measurements into SEPIA pseudo-structures, the same neural confidence scores used for general antibody ranking gain an extra 20–30 % enrichment for tight binders, turning a high-throughput yeast screen into an immediate training set for in-silico design.
The work underscores a practical shift: generate exhaustive mutational scans first, then let structure-aware ML cherry-pick variants. Caveat—affinity was assayed on surface-displayed VHH against immobilized peptides; avidity and off-rate differences in solution, or in cell-target recognition, remain unchecked. Still, coupling yeast profiling to million-scale labels gives antibody engineers a faster, fully in-house loop that starts and ends with experimental binding data.
The above is personal commentary for reference only. Refer to the original papers for authoritative content.