We designed a library of 10,000 binders using state of the art diffusion, hallucination and pLM models (BindCraft, Latent X, La Proteina, ESM3, APM and DPLM2) and evaluated the designs using AlphaFold3 and Boltz2. Due to the discrepancy between the prediction methods (that is also reflected in the medium score of our collection even though we obtain high AlphaFold3 scores), we focused on AlphaFold3 for our ranking. From this library, we selected the 600 highest scoring candidates and screened them experimentally in the wet lab within two weeks using a novel ultra fast surface display method (FLASH- Display). This method is highly scalable and can accommodate substantially larger binder inputs. From this experimentally characterized collection we identified the five strongest binders and, for comparison, selected the five top ranked in silico designs based on the ipSAE_min metric that did not show measurable binding in our assay. This integrated workflow demonstrates how pre-selection of candidate binders based on in silico scores can be combined with rapid wet lab testing to identify binders within a short time frame. The approach we used here provides a generalizable platform for validating next generation ML based binder design methodologies in a minimal time frame.

The project was implemented in collaboration with Amir M. Karimi, Vincent Pauline, Yichen Zhou and Leon Hetzel.