Directed protein evolution for generating antibodies with improved binding affinity or stability requires exploration of a vast space of possible mutations. Experimental high-throughput antibody engineering methods screen thousands to millions of variants using techniques like phage display or cell surface display. This imposes a heavy experimental burden. Computational methods provide a structure-guided rationale for selecting mutations, typically within the complementarity-determining regions, but still require experimental testing of many mutants.

Better prior information on evolutionary plausibility helps improve protein engineering efficiency.

A team of researchers led by Peter Kim at Stanford University has performed guided protein evolution using protein language models that were trained on millions of natural protein sequences. The models thereby learn amino acid patterns that are likely to be seen in nature. “Because the models are trained on millions of protein sequences produced by natural evolution, they are also helpful in suggesting mutations that are likely to have a functional impact when conducting directed evolution in the laboratory,” says Brian Hie, the lead author of the paper. “Unlike other methods for machine-learning-guided directed evolution, our method also requires no initial task-specific training data and recommends mutations directly from the wild-type sequence alone”, Hie adds.

Using these models, the team evolved seven human immunoglobulin G antibodies that bind to antigens from coronavirus, ebolavirus and influenza A virus. Screening 20 or fewer variants of each antibody across only two rounds of laboratory evolution allowed then to improve the affinity of all antibodies — an impressive feat. “We were able to improve the neutralization potency of an FDA-approved antibody against an Ebola pseudovirus and showed that for weak binders we can improve the affinity up to two orders of magnitude using very low-throughput experimentation,” says Kim.

While the results show that general language models outperform antibody-specific language models, because the training is done on general sequences, there is no guarantee that the language-model-recommended mutations will improve binding affinity in every instance. “Another open question is whether these models could be applicable to evolving more unnatural, de novo designed proteins”, says Hie. The researchers plan to leverage data beyond protein sequence information including protein structure information and binding affinity data to further improve the outcome. We look forward to developments in the area.

Original reference: Nat. Biotechnol. https://doi.org/10.1038/s41587-023-01763-2 (2023)