Capsids are protein boxes that viruses use to carry their genetic material. In gene therapy, these boxes serve as couriers: you put a desired gene inside, and the outside retains the ability to enter cells. The problem is that natural boxes are few, while the tasks are many: one must target only the liver, another the brain, and a third must avoid triggering an immune response.

Stochastic modeling helps search for new variants, but the space of possible sequences is enormous — larger than the number of atoms in the universe. Exhaustive search is impossible. So the authors took a model trained on millions of proteins and fine-tuned it on validated capsids. The model learned to distinguish viable sequences from junk. But simply generating similar sequences is boring and inefficient: you only get variations of what's already known.

To break free from the familiar, they added deep learning in medical imaging — specifically, reinforcement learning, a method where the program receives rewards for two qualities at once: viability and novelty. The farther a new sequence is from all known ones, the higher its novelty score. This way, the program ventures into uncharted regions of protein space while retaining a chance of functionality.

The results were compared: simple fine-tuning yields many 'correct' sequences that are similar to known ones. With reinforcement learning, the program finds variants that differ greatly from anything seen in nature and, according to predictions, remain viable. The final selection of candidates considers not only novelty and viability but also physical properties of the protein — stability and solubility.

This approach is not just a random mutation generator. It's a way to purposefully explore the vast space of possible proteins and find unexpected solutions. In the future, it could accelerate the creation of personalized viral vectors for gene therapy.