With ML-assisted drug discovery in, say, kinase inhibitors we search the molecule variation space for test candidates. What stops us from applying this to the protein scale by folding binding sites to discover variant candidates that would be more virulent? I'd like to understand this better.