ADP-3D (Atomic Denoising Priors for 3D Reconstruction) is a method for leveraging a pretrained protein diffusion model as a prior to solve 3D reconstruction problems.
The current implementation uses Chroma as the pretrained model.
Start by cloning the repo and, from the main directory, run:
conda create adp-3d-env python=3.9
conda activate adp-3d-env
pip install -r requirements.txt
conda install -c conda-forge ffmpeg
Follow this link to request your Chroma API key. Then, run:
python register.py --key YOUR_API_KEY
We provide a structure completion example with PDB:8ok3. To launch your experiment with, for example, a sub-sampling factor of 4, run:
python structure_completion.py --outdir /path/to/output/directory --cif data/cifs/8ok3.cif --fix-every 4
Once your experiment is finished, you can find the loss curves, RMSD curves and the final structure in the output directory.
We also provide our implementation of the DPS method, which we found to perform worse than our DiffPIR-based implementation.
We provide an example for the "distances-to-structure" task with PDB:8ok3. To launch your experiment with, for example, 500 known pairwise distances, run:
python distances_to_structure.py --outdir /path/to/output/directory --cif data/cifs/8ok3.cif --n-distances 500 --lr-distance 0.4
We provide an example with PDB:7PZT. We generated the density map in ChimeraX from the deposited structure with:
molmap #1 2.0 gridSpacing 0.5 edgePadding 50
An incomplete model was obtained with ModelAngelo, using the default parameters.
To launch your experiment, run:
python model_refinement.py --outdir /path/to/output/directory --mrc data/mrcs/7pzt_2.0A.mrc --ma-cif data/cifs/7pzt_MA_2.0A.cif --cif data/cifs/7pzt.cif --unpad-len 85
We use the unpad_len parameter to remove empty voxels from the edges of the density map and speed up computation. Do not forget to change this parameter when using a different input density map.
In the output directory, you will find a plot showing the completeness of the model vs. RMSD:
Failed to resolve 'chroma-weights.generatebiomedicines.com'
You may not have access to the internet. If you are able to instantiate a Chroma model in a different configuration (e.g., on the login node of a cluster), you will see that the weights of the model (backbone model and design model) are stored in a temporary location (/tmp/... ending with .pt). To bypass the need for connecting to the internet, you can copy/paste these .pt files to a stable location and point to them with the following additional arguments:
--weights-backbone /path/to/weights/of/backbone/model --weights-design /path/to/weights/of/design/model