Peptide drug design has long faced a structural challenge: how to transfer the binding intelligence of small molecules to peptide sequences? Traditional peptide generation models rely on sequence language models or structural predictions, lacking explicit learning of target-ligand interaction patterns.

PharmaPepGen is the industry's first pharmacophore-guided graph diffusion peptide generation model. The input is not a sequence, but rather the 3D features of small molecule pharmacophores—pharmacophore type and spatial coordinates. The output is a peptide sequence that matches the target's physicochemical properties and binding pattern. A graph convolutional network learns the interaction pattern between the target pocket and the small molecule ligand, and the diffusion model generates the peptide within that pattern.

The model was trained on approximately 360 crystal complex structures. On the test set, 10 sequences were generated for each target, and the average ipTM of the high-affinity sequences reached 0.79, close to the level of the natural ligand (0.83). This means that the generated peptides not only have reasonable sequences, but also have reliable binding postures to the targets.

In industrial applications, PharmaPepGen has supported the design of ultra-long-acting GLP-1 receptor agonists. The team generated 10,000 novel sequences, screened 60 candidate peptides, and achieved a 52% success rate in in vitro experiments. The candidate peptide D13 has a half-life three times that of smegglutinin and exhibits superior blood glucose-lowering effects in diabetic mouse models.

PharmaPepGen integrates "small molecule binding logic" with "peptide drugability" to achieve cross-modal targeted design, enabling targeted peptide discovery, cyclic peptide design, and oral peptide lead compound generation.