The high-throughput protein drug design language model GPDL, based on generative artificial intelligence, uses the ESM2 protein language model and ESMFold structure prediction as its core architecture. On 24 standard design tasks, compared to RFdiffusion, it significantly improves the rationality of backbone design: success rate increases by 8 percentage points, and diversity increases by 33%. Combined with the GPD sequence generation algorithm (which improves diversity by 2.2 times and speed by 1.6 times compared to ProteinMPNN), it achieves a complete design loop from "novel backbone" to "highly active sequence".

Traditional target discovery based on structured proteins has reached a saturation point. Our developed IDPFold, based on a diffusion model, is the first to achieve high-precision dynamic conformational ensemble prediction of natural random proteins (IDPs), overcoming the limitations of static structure prediction methods such as AlphaFold. Random proteins account for more than 30% of the human proteome and are closely related to complex diseases such as cancer and neurodegenerative diseases. The dynamic conformations generated by IDPFold can be combined with molecular screening to open up new drug design spaces for "undruggable" targets that are inaccessible by traditional methods.