Adaptyv Bio Round 2 | EGFR Binder Design
Methodology
I employed a two-phase approach to design these binders. Initially, I utilized BindCraft, the previous winning method, to generate ~100 potential binders. Subsequently, I applied the recently published structural evolution method (“Unsupervised evolution of protein and antibody complexes with a structure-informed language model”, Science 2024), to optimize the sequence. The objective was to minimize the energy of the Rosetta FastRelax+REF2015 scoring function. I conducted several rounds of sequence evolution, selecting the best (lowest energy) mutation in each iteration.
I developed this approach after analyzing the Pearson correlation of various metrics on the initial round data. Notably, the FastRelax+REF2015 Rosetta score exhibited a significantly higher Pearson correlation of 0.670 compared to other methods like PRODIGY, which had a Pearson correlation of approximately 0.2. For instance, a Cetuximab+EGFR Extracellular Domain structure predicted using AF2Multimer yielded an energy of -2381.98, substantially lower than any binder generated in the first round of the challenge. The closest binder was the winning submission by Martin Pacessa, which had an energy of -2145.768.
To further reduce the Rosetta energy, I incorporated sequence evolution, as it proved challenging to achieve significant improvements using BindCraft due to limited computational resources. While the PLL of these sequences is underwhelming, I do not think that this will be representative of binding strength for these binders. I performed expression prediction using SoluProt and found that my proposed binders are predicted to express well.
Results
Will update when Adaptyv posts results.
UPDATE Dec 7, 2024: All of my binders were expressed successfully, but none of them bound EGFR. I suspect this is because of the poor ESM2LL, iPTM, and pLDTT scores. Perhaps the FastRelax+REF2015 Rosetta score of a complex is not useful alone, and should be used in conjuction with traditional metrics.