Official repository for Full-atom MPNN (FAMPNN), a sequence design method that explicitly models both sequence identity and sidechain conformation. For more details, see our preprint: Sidechain conditioning and modeling for full-atom protein sequence design with FAMPNN
This repository contains code for full-atom sequence design, sidechain packing, and mutation scoring.
To run the scripts in this repository, we recommend using a conda environment. First, clone this repository, navigate to the root directory, and run:
conda create -n fampnn python=3.10
conda activate fampnn
Next, to install PyTorch with GPU support, follow instructions from https://pytorch.org/get-started/locally/ for your system and CUDA version. We used PyTorch 2.4.1 with CUDA 12.1 for our experiments. Finally, run:
conda env update --file environment.yaml
pip install -e .
We provide the following weights under weights/ for FAMPNN:
| Model | Weights | Description |
|---|---|---|
| FAMPNN (0.0Å) | fampnn_0_0.pt |
Trained on the full PDB dataset with 0.0Å noise. |
| FAMPNN (0.3Å) | fampnn_0_3.pt |
Trained on the full PDB dataset with 0.3Å noise. Recommended for sequence design. |
| FAMPNN (0.3Å, CATH) | fampnn_0_3_cath.pt |
Trained on the CATH dataset with 0.3Å noise. Recommended for mutation scoring. |
Example scripts are provided under examples/scripts/ for sequence design, sidechain packing, and mutation scoring.
To design a sequence for a given list of backbones, use fampnn/inference/seq_design.py.
This script takes in a pdb_dir with PDB files (and optionally a pdb_key_list to subset to certain PDBs) and will design num_seqs_per_pdb sequences for each PDB. The number of sampling steps can be adjusted with timestep_schedule.num_steps (defaults to 100).
An example of sequence design on a small set of RFdiffusion-generated de novo backbones is provided in examples/scripts/seq_design_unconditional.sh.
FAMPNN can also condition on arbitrary mixtures of sequence-only and sequence-and-sidechain context. To condition on specific amino acid types and sidechains, supply a CSV to fixed_pos_csv. A row of the CSV might look like:
3e1r,"A165-190,A200-210","A170-180"
where the first column specifies the PDB file name (without extension), the second column specifies the sequence positions in the PDB to condition on, and the third column specifies the sidechains from the input PDB to condition on. Fixed sidechain positions must be a subset of fixed sequence positions, since it does not make sense to condition on a sidechain without conditioning on its sequence identity. If a particular column is empty, the script will assume no fixed positions for that context. If fixed_pos_csv is not provided, the script will perform unconditional sequence design using backbone-only context.
For example, the row above specifies that sequence design should condition on the sequence from 3e1r.pdb at positions 165-190 and 200-210 in chain A, and the sidechains at positions 170-180. An example of sequence design with conditional context is provided in examples/scripts/seq_design_conditional.sh.
To pack sidechains onto a given backbone, use fampnn/inference/pack.py.
This script takes in a pdb_dir with PDB files (and optionally a pdb_key_list to subset to certain PDBs) and will pack sidechains onto the given backbones. By default, if fixed_pos_csv is not provided, the script will perform sidechain packing assuming full sequence context but no sidechain context. If fixed sequence positions are specified, the model will only pack those provided positions. An example of sidechain packing with fixed positions is provided in examples/scripts/pack.sh. Packed structures also feature a per-atom confidence score stored in the B-factor column of a PDB, which can be visualized in a PyMOL session.
We provide two scripts for scoring mutations for a given PDB.
To score all possible single mutations for a given PDB, use fampnn/inference/score_all_muts.py. This script will iterate over each position in the PDB, and score the fitness of all possible mutations normalized by the likelihood of the wild-type amino acid. An example of how to run this script is provided in examples/scripts/score_all_muts.sh. The script will output a CSV in the directory specified by the out_dir argument with the score for each possible mutation.
To score specific mutations for a given PDB, use fampnn/inference/score_multi.py. This script requires a 1-column CSV file with the mutations to score, specified by the mutations_path argument. A mutation is specified as follows:
<Wild-type 1-letter amino acid code><0-index'd position in the PDB file><Mutant 1-letter amino acid code>
To evaluate multiple simultaneous mutations (eg. double and triple mutations), simply concatenate the single-mutations with colons :. For example, simultaneous mutations of Asparagine (N) in the first and second positons of a PDB to Proline (P) and Arginine (R) respectively, would be N0P:N1R
An example for running this script is provided in examples/scripts/score_multi.sh, and an example of an input csv for this script is provided in fampnn/examples/scoring/mutations.csv.