Neopep: Reinforcement Learning From Machine-Learned Force Field Energy for Non-Canonical Peptide Structure Prediction

Neopep, a sequence-conditioned diffusion model for ncAA-containing peptides, fine-tuned with MLFF-derived energy, produces structures with lower average RMSD to ground truth than without finetuning and other state-of-the-art models.

• Why it matters: Sequence-to-structure prediction models struggle with non-canonical amino acids because fewer than 0.02% of residues in PDB deposits correspond to ncAAs, creating a training-data bottleneck.
• What they did: Neopep is pretrained on PDB structures and then fine-tuned with reinforcement learning, where a frozen MLFF scores proposed structures and its energy signal is backpropagated into the model.
• The result: Trained on a fixed set of 189 ncAAs, Neopep generates ncAA peptide structures that achieve lower average RMSD versus ground truth, compared with no fine-tuning and other state-of-the-art structure prediction models.