Structure-based computational design approaches rely on large-scale models of sequence-structure relationships to optimize function (ProteinMPNN, RFdiffusion). In functional environments, proteins operate within complex interaction networks that simultaneously require multiple molecular properties, making function optimization a multi-objective problem. The physiological complexity of environments that condition protein interactions can be difficult to model in-silico.
Rather than relying on the function-structure-sequence relationship, we investigate a direct sequence-function approach that leverages screening experiments. Long before computational design emerged, biologists successfully optimized protein function by mimicking natural selection in the laboratory. In phage display, for example, DNA libraries encoding protein variants are expressed on phage surfaces, subjected to selection pressure against immobilized targets, and iteratively enriched based on binding affinity. This process generates sequencing data that captures the relationship between sequence variants and their functional performance.
We demonstrate how computational models can be trained on experimental data to generalize local screening information and explore the sequence spaces in-silico. The key advantage of this sequence-based approach lies in its adaptability to multiple molecular properties through complex experimental setups that screen for diverse phenotypes. These models learn multiple selection "modes" from the data, enabling integration of screening information, management of property trade-offs, and filtering of spurious experimental effects.
These models could be used to computationally generate protein variants or to inform structure-based approaches for protein design. ​