LassoDiff Inference

Project Goals

• All-atom lasso peptide modeling under a backbone + isobond flow matching framework
• Pocket ligand-conditioned all-atom lasso peptide prediction
• De novo all-atom design of lasso peptides

Current Progress

• Integrated and organized core inference flow matching modules
• Added loss computation and feature processing for backbone and isobond constraints
• Completed adaptation paths and attribution statements for Protenix and ml-simplefold

Training (bash)

python lassodiff/train_toy.py \
  --n 128 \
  --L 40 \
  --K 3 \
  --epochs 3 \
  --batch_size 4 \
  --lr 1e-4

torchrun --nproc_per_node=2 lassodiff/train_toy.py \
  --structure_dir /path/to/structures \
  --epochs 3 \
  --batch_size 4 \
  --lr 1e-4 \
  --export_val_pdb \
  --export_val_dir val

Next Steps

• Complete the feature pipeline for pocket ligand-conditioned inputs
• Evaluate stability and usability of all-atom prediction and de novo design workflows
• Build a minimal reproducible inference pipeline and evaluation metrics

Acknowledgements

• Thanks to Protenix for the model architecture and engineering implementation
• Thanks to ml-simplefold for key modules and engineering references