Uni-Fold: Training your own deep protein-folding models.

Related tags

Deep LearningUni-Fold
Overview

Uni-Fold: Training your own deep protein-folding models.

This package provides and implementation of a trainable, Transformer-based deep protein folding model. We modified the open-source code of DeepMind AlphaFold v2.0 and provided code to train the model from scratch. See the reference and the repository of DeepMind AlphaFold v2.0. To train your own Uni-Fold models, please follow the steps below:

1. Install the environment.

Run the following code to install the dependencies of Uni-Fold:

  conda create -n unifold python=3.8.10 -y
  conda activate unifold
  ./install_dependencies.sh

Uni-Fold has been tested for Python 3.8.10, CUDA 11.1 and OpenMPI 4.1.1. We recommend using Conda >= 4.10 when installing the environment: using Conda with lower level may lead to some conflicts between packages.

2. Prepare data before training.

Before you start to train your own folding models, you shall prepare the features and labels of the training proteins. Features of proteins mainly include the amino acid sequence, MSAs and templates of proteins. These messages should be contained in a pickle file /features.pkl for each training protein. Uni-Fold provides scripts to process input FASTA files, relying on several external databases and tools. Labels are CIF files containing the structures of the proteins.

2.1 Datasets and external tools.

Uni-Fold adopts the same data processing pipeline as AlphaFold2. We kept the scripts of downloading corresponding databases for searching sequence homologies and templates in the AlphaFold2 repo. Use the command

  bash scripts/download_all_data.sh /path/to/database/directory

to download all required databases of Uni-Fold.

If you successfully installed the Conda environment in Section 1, external tools of search homogenous sequences and templates should be installed properly. As an alternative, you can customize the parameters of feature preparation script to refer to your own databases and tools.

2.2 Run the preparation code.

An example command of running the feature preparation pipeline would be

  python generate_pkl_features.py \
    --fasta_dir ./example_data/fasta \
    --output_dir ./out \
    --data_dir /path/to/database/directory \
    --num_workers 1

This command automatically processes all FASTA files under fasta_dir, and dumps the results to output_dir. Note that each FASTA file should contain only one sequence. The default number of cpu used in hhblits and jackhmmer are 4 and 8. You can modify them in unifold/data/tools/hhblits.py and unifold/data/tools/jackhmmer.py, respectively.

2.3 Organize your training data.

Uni-Fold uses the class DataSystem to automatically sample and load the training entries. To make everything goes right, you shall pay attention to how the training data is organized. Two directories should be established, one with input features (features.pkl files, referred as features_dir) and the other with labels (*.cif files, referred as mmcif_dir). The feature directory should have its files named as _ _ /features.pkl , and the label directory should have its files named as .cif . Users shall make sure that all proteins used for training have their corresponding labels. See ./example_data/features and ./example_data/mmcif for instances of features_dir and mmcif_dir.

3. Train Uni-Fold.

3.1 Configuration.

Before you conduct any actual training processes, please make sure that you correctly configured the code. Modify the training configurations in unifold/train/train_config.py. We annotated the default configurations to reproduce AlphaFold in the script. Specifically, modify the data setups in unifold/train/train_config.py:

"data": {
  "train": {
    "features_dir": "where/training/protein/features/are/stored/",
    "mmcif_dir": "where/training/mmcif/files/are/stored/",
    "sample_weights": "which/specifies/proteins/for/training.json"
  },
  "eval": {
    "features_dir": "where/validation/protein/features/are/stored/",
    "mmcif_dir": "where/validation/mmcif/files/are/stored/",
    "sample_weights": "which/specifies/proteins/for/training.json"
  }
}

The specified data should be contained in two folders, namely a features_dir and a mmcif_dir. Organizations of the two directories are introduced in Section 2.3. Meanwhile, if you want to specify the subset of training data under the directories, or assign customized sample weights for each protein, write a json file and feed its path to sample_weights. This is optional, as you can leave it as None (and the program will attempt to use all entries under features_dir with uniform weights). The json file should be a dictionary contains the basename of directories of protein features ([pdb_id]_[model_id]_[chain_id]) and the sample weight of each protein in the training process (integer or float), such as:

{"1am9_1_C": 82, "1amp_1_A": 291, "1aoj_1_A": 60, "1aoz_1_A": 552}

or for uniform sampling, simply using a list of protein entries suffices:

["1am9_1_C", "1amp_1_A", "1aoj_1_A", "1aoz_1_A"]

Meanwhile, the configurations of models can be edited in unifold/model/config.py for users who want to customize their own folding models.

3.2 Run the training code!

To train the model on a single node without MPI, run

python train.py

You can also train the model using MPI (or workload managers that supports MPI, such as PBS or Slurm) by running:

mpirun -n <numer_of_gpus> python train.py

In either way, make sure you properly configurate the option use_mpi in unifold/train/train_config.py.

4. Inference with trained models.

4.1 Inference from features.pkl.

We provide the run_from_pkl.py script to support inferencing protein structures from features.pkl inputs. A demo command would be

python run_from_pkl.py \
  --pickle_dir ./example_data/features \
  --model_names model_2 \
  --model_paths /path/to/model_2.npz \
  --output_dir ./out

or

python run_from_pkl.py \
  --pickle_paths ./example_data/features/1ak0_1_A/features.pkl \
  --model_names model_2 \
  --model_paths /path/to/model_2.npz \
  --output_dir ./out

The command will generate structures of input features from different input models (in PDB format), the running time of each component, and corresponding residue-wise confidence score (predicted LDDT, or pLDDT).

4.2 Inference from FASTA files.

Essentially, inferencing the structures from given FASTA files includes two steps, i.e. generating the pickled features and predicting structures from them. We provided a script, run_from_fasta.py, as a more friendly user interface. An example usage would be

python run_from_pkl.py \
  --fasta_paths ./example_data/fasta/1ak0_1_A.fasta \
  --model_names model_2 \
  --model_paths /path/to/model_2.npz \
  --data_dir /path/to/database/directory
  --output_dir ./out

4.3 Generate MSA with MMseqs2.

It may take hours and much memory to generate MSA for sequences,especially for long sequences. In this condition, MMseqs2 may be a more efficient way. It can be used in the following way after it is installed:

# download and build database
mkdir mmseqs_db && cd mmseqs_db
wget http://wwwuser.gwdg.de/~compbiol/colabfold/uniref30_2103.tar.gz
wget http://wwwuser.gwdg.de/~compbiol/colabfold/colabfold_envdb_202108.tar.gz
tar xzvf uniref30_2103.tar.gz
tar xzvf colabfold_envdb_202108.tar.gz
mmseqs tsv2exprofiledb uniref30_2103 uniref30_2103_db
mmseqs tsv2exprofiledb colabfold_envdb_202108 colabfold_envdb_202108_db
mmseqs createindex uniref30_2103_db tmp
mmseqs createindex colabfold_envdb_202108_db tmp
cd ..

# MSA search
./scripts/colabfold_search.sh mmseqs "query.fasta" "mmseqs_db/" "result/" "uniref30_2103_db" "" "colabfold_envdb_202108_db" "1" "0" "1"

5. Changes from AlphaFold to Uni-Fold.

  • We implemented classes and methods for training and inference pipelines by adding scripts under unifold/train and unifold/inference.
  • We added scripts for installing the environment, training and inferencing.
  • Files under unifold/common, unifold/data and unifold/relax are minimally altered for re-structuring the repository.
  • Files under unifold/model are moderately altered to allow mixed-precision training.
  • We removed unused scripts in training AlphaFold model.

6. License and disclaimer.

6.1 Uni-Fold code license.

Copyright 2021 Beijing DP Technology Co., Ltd.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0.

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

6.2 Use of third-party software.

Use of the third-party software, libraries or code may be governed by separate terms and conditions or license provisions. Your use of the third-party software, libraries or code is subject to any such terms and you should check that you can comply with any applicable restrictions or terms and conditions before use.

6.3 Contributing to Uni-Fold.

Uni-Fold is an ongoing project. Our target is to design better protein folding models and to apply them in real scenarios. We welcome the community to join us in developing the repository together, including but not limited to 1) reports and fixes of bugs,2) new features and 3) better interfaces. Please refer to CONTRIBUTING.md for more information.

Owner
DeepModeling
Define the future of scientific computing together
DeepModeling
i-RevNet Pytorch Code

i-RevNet: Deep Invertible Networks Pytorch implementation of i-RevNets. i-RevNets define a family of fully invertible deep networks, built from a succ

Jörn Jacobsen 378 Dec 06, 2022
A simple, fully convolutional model for real-time instance segmentation.

You Only Look At CoefficienTs ██╗ ██╗ ██████╗ ██╗ █████╗ ██████╗████████╗ ╚██╗ ██╔╝██╔═══██╗██║ ██╔══██╗██╔════╝╚══██╔══╝ ╚██

Daniel Bolya 4.6k Dec 30, 2022
[ICCV 2021] FaPN: Feature-aligned Pyramid Network for Dense Image Prediction

FaPN: Feature-aligned Pyramid Network for Dense Image Prediction [arXiv] [Project Page] @inproceedings{ huang2021fapn, title={{FaPN}: Feature-alig

Shihua Huang 23 Jul 22, 2022
Reporting and Visualization for Hazardous Events

Reporting and Visualization for Hazardous Events

Jv Kyle Eclarin 2 Oct 03, 2021
FOSS Digital Asset Distribution Platform built on Frappe.

Digistore FOSS Digital Assets Marketplace. Distribute digital assets, like a pro. Video Demo Here Features Create, attach and list digital assets (PDF

Mohammad Hussain Nagaria 30 Dec 08, 2022
RCD: Relation Map Driven Cognitive Diagnosis for Intelligent Education Systems

RCD: Relation Map Driven Cognitive Diagnosis for Intelligent Education Systems This is our implementation for the paper: Weibo Gao, Qi Liu*, Zhenya Hu

BigData Lab @USTC 中科大大数据实验室 10 Oct 16, 2022
Data visualization app for H&M competition in kaggle

handm_data_visualize_app Data visualization app by streamlit for H&M competition in kaggle. competition page: https://www.kaggle.com/competitions/h-an

Kyohei Uto 12 Apr 30, 2022
PyTorch implementation of deep GRAph Contrastive rEpresentation learning (GRACE).

GRACE The official PyTorch implementation of deep GRAph Contrastive rEpresentation learning (GRACE). For a thorough resource collection of self-superv

Big Data and Multi-modal Computing Group, CRIPAC 186 Dec 27, 2022
moving object detection for satellite videos.

DSFNet: Dynamic and Static Fusion Network for Moving Object Detection in Satellite Videos Algorithm Introduction DSFNet: Dynamic and Static Fusion Net

xiaochao 39 Dec 16, 2022
Where2Act: From Pixels to Actions for Articulated 3D Objects

Where2Act: From Pixels to Actions for Articulated 3D Objects The Proposed Where2Act Task. Given as input an articulated 3D object, we learn to propose

Kaichun Mo 69 Nov 28, 2022
Notes, programming assignments and quizzes from all courses within the Coursera Deep Learning specialization offered by deeplearning.ai

Coursera-deep-learning-specialization - Notes, programming assignments and quizzes from all courses within the Coursera Deep Learning specialization offered by deeplearning.ai: (i) Neural Networks an

Aman Chadha 1.7k Jan 08, 2023
E2VID_ROS - E2VID_ROS: E2VID to a real-time system

E2VID_ROS Introduce We extend E2VID to a real-time system. Because Python ROS ca

Robin Shaun 7 Apr 17, 2022
UniMoCo: Unsupervised, Semi-Supervised and Full-Supervised Visual Representation Learning

UniMoCo: Unsupervised, Semi-Supervised and Full-Supervised Visual Representation Learning This is the official PyTorch implementation for UniMoCo pape

dddzg 49 Jan 02, 2023
Learning recognition/segmentation models without end-to-end training. 40%-60% less GPU memory footprint. Same training time. Better performance.

InfoPro-Pytorch The Information Propagation algorithm for training deep networks with local supervision. (ICLR 2021) Revisiting Locally Supervised Lea

78 Dec 27, 2022
OneShot Learning-based hotword detection.

EfficientWord-Net Hotword detection based on one-shot learning Home assistants require special phrases called hotwords to get activated (eg:"ok google

ANT-BRaiN 102 Dec 25, 2022
[CVPR 2021] Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers

[CVPR 2021] Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers

Fudan Zhang Vision Group 897 Jan 05, 2023
Non-stationary GP package written from scratch in PyTorch

NSGP-Torch Examples gpytorch model with skgpytorch # Import packages import torch from regdata import NonStat2D from gpytorch.kernels import RBFKernel

Zeel B Patel 1 Mar 06, 2022
MoCoPnet - Deformable 3D Convolution for Video Super-Resolution

Deformable 3D Convolution for Video Super-Resolution Pytorch implementation of l

Xinyi Ying 28 Dec 15, 2022
Shallow Convolutional Neural Networks for Human Activity Recognition using Wearable Sensors

-IEEE-TIM-2021-1-Shallow-CNN-for-HAR [IEEE TIM 2021-1] Shallow Convolutional Neural Networks for Human Activity Recognition using Wearable Sensors All

Wenbo Huang 1 May 17, 2022
A non-linear, non-parametric Machine Learning method capable of modeling complex datasets

Fast Symbolic Regression Symbolic Regression is a non-linear, non-parametric Machine Learning method capable of modeling complex data sets. fastsr aim

VAMSHI CHOWDARY 3 Jun 22, 2022