Generating mutants of monotone affinity towards stronger protein complexes through adversarial learning

Despite breakthroughs achieved in protein sequence-to-structure and function-to-sequence predictions, the affinity-to-mutation prediction problem remains unsolved. Such a problem is of exponential complexity deemed to find a mutated protein or protein complex having a guaranteed binding-affinity change. Here we introduce an adversarial learning-based mutation method that creates optimal amino acid substitutions and changes the mutant’s affinity change significantly in a preset direction. The key aspect in our method is the adversarial training process that dynamically labels the real side of the protein data and generates fake pseudo-data accordingly to construct a deep learning architecture for guiding the mutation. The method is sufficiently flexible to generate both single- and multipointed mutations at the adversarial learning step to mimic the natural circumstances of protein evolution. Compared with random mutants, our mutated sequences have in silico exhibited more than one order of change in magnitude of binding free energy change towards stronger complexes in the case study of Novavax–angiotensin-converting enzyme-related carboxypeptidase vaccine construct optimization. We also applied the method iteratively each time, using the output as the input sequence of the next iteration, to generate paths and a landscape of mutants with affinity-increasing monotonicity to understand SARS-CoV-2 Omicron’s spike evolution. With these steps taken for effective generation of protein mutants of monotone affinity, our method will provide potential benefits to many other applications including protein bioengineering, drug design, antibody reformulation and therapeutic protein medication.