Generative protein design
Turning a functional specification into a molecule a model can actually produce.
The problem. Protein generative models can produce plausible sequences, but a scientist rarely wants “a plausible protein” — they want one that catalyses a particular reaction, binds a particular target, or carries a particular motif in a particular place. The gap is a language: how do you state a functional requirement in terms a model can condition on?
What I build. Generative models that accept combinations of functional annotations as conditions rather than a single label. (Yin et al., 2025) introduces combinatorial functional protein generation with diffusion language models, conditioning jointly on enzyme-commission numbers, GO terms and motifs so that multiple functional constraints can be imposed at once. (Feng et al., 2026) makes the same idea modular at the segment level, so that individual regions of a protein can be re-engineered while the rest is held fixed. (Guo et al., 2025) extends design beyond the twenty canonical residues to peptides containing non-standard amino acids, which is where much of the therapeutic space actually lives.
Where it is going. Function is currently specified by whatever annotation vocabularies happen to exist. The next step is learning the specification language itself, so that a design goal can be stated the way a biologist would state it and still be executable by a model.