We recommend selecting `openfold/resources` as the params directory as this is the default directory used by the `run_pretrained_openfold.py` to locate parameters.
We recommend selecting `openfold/resources` as the params directory as this is the default directory used by the `run_pretrained_openfold.py` to locate parameters.
If you choose to use a different directory, you may make a symlink to the `openfold/resources` directory, or specify an alternate parameter path with the command line argument `--jax_path` for AlphaFold parameters or `--openfold_checkpoint_path` for OpenFold parameters.
If you choose to use a different directory, you may make a symlink to the `openfold/resources` directory, or specify an alternate parameter path with the command line argument `--jax_param_path` for AlphaFold parameters or `--openfold_checkpoint_path` for OpenFold parameters.
### Model Inference
### Model Inference
...
@@ -138,6 +138,7 @@ Some commonly used command line flags are here. A full list of flags can be view
...
@@ -138,6 +138,7 @@ Some commonly used command line flags are here. A full list of flags can be view
- `--data_random_seed`: Specifies a random seed to use.
- `--data_random_seed`: Specifies a random seed to use.
- `--save_outputs`: Saves a copy of all outputs from the model, e.g. the output of the msa track, ptm heads.
- `--save_outputs`: Saves a copy of all outputs from the model, e.g. the output of the msa track, ptm heads.
- `--experiment_config_json`: Specify configuration settings using a json file. For example, passing a json with `{globals.relax.max_iterations = 10}` specifies 10 as the maximum number of relaxation iterations. See for [`openfold/config.py`](https://github.com/aqlaboratory/openfold/blob/main/openfold/config.py#L283) the full dictionary of configuration settings. Any parameters that are not manually set in these configuration settings will refer to the defaults specified by your `config_preset`.
- `--experiment_config_json`: Specify configuration settings using a json file. For example, passing a json with `{globals.relax.max_iterations = 10}` specifies 10 as the maximum number of relaxation iterations. See for [`openfold/config.py`](https://github.com/aqlaboratory/openfold/blob/main/openfold/config.py#L283) the full dictionary of configuration settings. Any parameters that are not manually set in these configuration settings will refer to the defaults specified by your `config_preset`.
- `--use_custom_template`: Uses all .cif files in `template_mmcif_dir` as template input. Make sure the chains of interest have the identifier _A_ and have the same length as the input sequence. The same templates will be read for all sequences that are passed for inference.
### Advanced Options for Increasing Efficiency
### Advanced Options for Increasing Efficiency
...
@@ -159,12 +160,12 @@ Note that chunking (as defined in section 1.11.8 of the AlphaFold 2 supplement)
...
@@ -159,12 +160,12 @@ Note that chunking (as defined in section 1.11.8 of the AlphaFold 2 supplement)
#### Long sequence inference
#### Long sequence inference
To minimize memory usage during inference on long sequences, consider the following changes:
To minimize memory usage during inference on long sequences, consider the following changes:
- As noted in the AlphaFold-Multimer paper, the AlphaFold/OpenFold template stack is a major memory bottleneck for inference on long sequences. OpenFold supports two mutually exclusive inference modes to address this issue. One,`average_templates`in the`template`section of the config, is similar to the solution offered by AlphaFold-Multimer, which is simply to average individual template representations. Our version is modified slightly to accommodate weights trained using the standard template algorithm. Using said weights, we notice no significant difference in performance between our averaged template embeddings and the standard ones. The second,`offload_templates`, temporarily offloads individual template embeddings into CPU memory. The former is an approximation while the latter is slightly slower; both are memory-efficient and allow the model to utilize arbitrarily many templates across sequence lengths. Both are disabled by default, and it is up to the user to determine which best suits their needs, if either.
- As noted in the AlphaFold-Multimer paper, the AlphaFold/OpenFold template stack is a major memory bottleneck for inference on long sequences. OpenFold supports two mutually exclusive inference modes to address this issue. One,`average_templates`in the`template`section of the config, is similar to the solution offered by AlphaFold-Multimer, which is simply to average individual template representations. Our version is modified slightly to accommodate weights trained using the standard template algorithm. Using said weights, we notice no significant difference in performance between our averaged template embeddings and the standard ones. The second,`offload_templates`, temporarily offloads individual template embeddings into CPU memory. The former is an approximation while the latter is slightly slower; both are memory-efficient and allow the model to utilize arbitrarily many templates across sequence lengths. Both are disabled by default, and it is up to the user to determine which best suits their needs, if either.
- Inference-time low-memory attention (LMA) can be enabled in the model config. This setting trades off speed for vastly improved memory usage. By default, LMA is run with query and key chunk sizes of 1024 and 4096, respectively. These represent a favorable tradeoff in most memory-constrained cases. Powerusers can choose to tweak these settings in`openfold/model/primitives.py`. For more information on the LMA algorithm, see the aforementioned Staats & Rabe preprint.
- Inference-time low-memory attention (LMA) can be enabled in the model config. This setting trades off speed for vastly improved memory usage. By default, LMA is run with query and key chunk sizes of 1024 and 4096, respectively. These represent a favorable tradeoff in most memory-constrained cases. Powerusers can choose to tweak these settings in`openfold/model/primitives.py`. For more information on the LMA algorithm, see the aforementioned Staats & Rabe preprint.
- Disable`tune_chunk_size`for long sequences. Past a certain point, it only wastes time.
- Disable`tune_chunk_size`for long sequences. Past a certain point, it only wastes time.
- As a last resort, consider enabling`offload_inference`. This enables more extensive CPU offloading at various bottlenecks throughout the model.
- As a last resort, consider enabling`offload_inference`. This enables more extensive CPU offloading at various bottlenecks throughout the model.
- Disable FlashAttention, which seems unstable on long sequences.
- Disable FlashAttention, which seems unstable on long sequences.
Using the most conservative settings, we were able to run inference on a 4600-residue complex with a single A100. Compared to AlphaFold's own memory offloading mode, ours is considerably faster; the same complex takes the more efficent AlphaFold-Multimer more than double the time. Use the`long_sequence_inference`config option to enable all of these interventions at once. The`run_pretrained_openfold.py`script can enable this config option with the`--long_sequence_inference`command line option
Using the most conservative settings, we were able to run inference on a 4600-residue complex with a single A100. Compared to AlphaFold's own memory offloading mode, ours is considerably faster; the same complex takes the more efficent AlphaFold-Multimer more than double the time. Use the`long_sequence_inference`config option to enable all of these interventions at once. The`run_pretrained_openfold.py`script can enable this config option with the`--long_sequence_inference`command line option
Input FASTA files containing multiple sequences are treated as complexes. In this case, the inference script runs AlphaFold-Gap, a hack proposed [here](https://twitter.com/minkbaek/status/1417538291709071362?lang=en), using the specified stock AlphaFold/OpenFold parameters (NOT AlphaFold-Multimer).
Input FASTA files containing multiple sequences are treated as complexes. In this case, the inference script runs AlphaFold-Gap, a hack proposed [here](https://twitter.com/minkbaek/status/1417538291709071362?lang=en), using the specified stock AlphaFold/OpenFold parameters (NOT AlphaFold-Multimer).
In this guide, we will OpenFold and its dependencies.
**Pre-requisites**
This package is currently supported for CUDA 11 and Pytorch 1.12. All dependencies are listed in the [`environment.yml`](https://github.com/aqlaboratory/openfold/blob/main/environment.yml). To install OpenFold for CUDA 12, please refer to the [Environment specific modifications](#Environment-specific-modifications) section.
At this time, only Linux systems are supported.
## Instructions
:::
### Installation:
1. Clone the repository, e.g. `git clone https://github.com/aqlaboratory/openfold.git`
1. From the `openfold` repo:
- Create a [Mamba]("https://github.com/conda-forge/miniforge/releases/latest/download/) environment, e.g.
You may optionally set this as a conda environment variable according to the [conda docs](https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#saving-environment-variables) to activate each time the environment is used.
1. Download parameters. We recommend using a destination as `openfold/resources` as our unittests will look for the weights there.
To test your installation, you can run OpenFold unit tests. Make sure that the OpenFold and AlphaFold parameters have been downloaded, and that they are located (or symlinked) in the directory `openfold/resources`
Run with the following script:
> scripts/run_unit_tests.sh
The script is a thin wrapper around Python's `unittest` suite, and recognizes `unittest` arguments. E.g., to run a specific test verbosely:
> scripts/run_unit_tests.sh -v tests.test_model
**Alphafold Comparison tests:**
Certain tests perform equivalence comparisons with the AlphaFold implementation. Instructions to run this level of tests requires an environment with both AlphaFold 2.0.1 and OpenFold installed, and is not covered in this guide. These tests are skipped by default if no installation of AlphaFold is found.
## Environment specific modifications
### CUDA 12
To use OpenFold on CUDA 12 environment rather than a CUDA 11 environment.
In step 1, use the branch [`pl_upgrades`](https://github.com/aqlaboratory/openfold/tree/pl_upgrades) rather than the main branch, i.e. replace the command in step 1 with `git clone -b pl_upgrades https://github.com/aqlaboratory/openfold.git`
and follow the rest of the steps of [Installation Guide](#Installation)
### MPI
To use OpenFold with MPI support, you will need to add the package [`mpi4py`](https://pypi.org/project/mpi4py/). This can be done with pip in your OpenFold environment, e.g. `$ pip install mpi4py`.
### Install OpenFold parameters without aws
If you don't have access to `aws` on your system, you can use a different download source:
- Google Drive: `scripts/download_openfold_params_gdrive.sh`
### Docker setup
A [`Dockerfile`] is provided to build an OpenFold Docker image. Additional notes for setting up a docker container for OpenFold and running inference can be found [here](original_readme.md#building-and-using-the-docker-container).
:alt: Comparison of OpenFold and AlphaFold2 predictions to the experimental structure of PDB 7KDX, chain B._
:alt: Comparison of OpenFold and AlphaFold2 predictions to the experimental structure of PDB 7KDX, chain B._
```
```
Welcome to the Documentation for OpenFold, the fully open source, trainable, PyTorch-based reproduction of DeepMind's
Welcome to the Documentation for [OpenFold](https://github.com/aqlaboratory/openfold), the fully open source, trainable, PyTorch-based reproduction of DeepMind's
