amdenterpriseai/aim-instinct-model-cytokineking-freebindcraft repository overview

⁠FreeBindCraft Protein Binder Design

FreeBindCraft designs miniprotein binders against target proteins using AlphaFold2-based hallucination (ColabDesign) combined with ProteinMPNN sequence design. In --no-pyrosetta mode, OpenMM performs structure relaxation, the sc-rs binary computes shape complementarity, and FreeSASA/Biopython compute solvent-accessible surface area. Given a target PDB and optional hotspot residues, the pipeline iteratively hallucinates binder backbones, sequences them with MPNN, predicts the complex with AlphaFold2 multimer, filters by pLDDT/pTM/i_pTM/SC/dSASA thresholds, and writes accepted and ranked PDB files, the per-design score CSVs and a summary.json (trajectories, accepted designs, timings) to the output directory.

⁠Release notes

FreeBindCraft is a fork of BindCraft that replaces PyRosetta (commercial license) with OpenMM for relaxation, sc-rs for shape complementarity, and FreeSASA for SASA computation. The image runs it in --no-pyrosetta mode, which makes it fully open source. It uses JAX 0.8.2 with ROCm 7.2.2 wheels. The AlphaFold2 parameters are not in the image; mount them and set AIM_CACHE_PATH. This is a long-running job container (EnterpriseAI-specs ADR-0008), not a service: running the image with no arguments starts the design run, reads target.pdb from AIM_INPUT_PATH, writes results to AIM_OUTPUT_PATH and exits. A run can take minutes to many hours; bound it with the wall-clock and trajectory settings described in the image README.