timed-design is a library to use protein sequence design models and analyse predictions. We feature retrained Keras models for novel models (TIMED and TIMED-rotamer) as well as re-implementations of well known models for which code or model are not intuitively available (ProDCoNN, DenseCPD, DenseNet).
- Install
- 1. Use Models
- 2. Sample Sequences Using Monte Carlo
- 3. Analyse Rotamer Predictions
- 4. Background
- 5. Cite This Work
- Clone and move inside the TIMED-Design directory:
git clone https://github.com/wells-wood-research/timed-design.git
cd timed-design
- Setting up conda:
conda create --name timed_design python=3.8
conda activate timed_design
- Install requirements for headless version (no UI):
pip install -r requirements.txt
- Install everything (TIMED-Design and UI):
Install TIMED-Design with pip:
pip install .
- For GPU support. Of course, NVIDIA drivers will never work out of the box. In our servers we found this workaround to work:
conda install cudatoolkit cudnn cupti
export PATH=/usr/local/cuda/bin:$PATH
pip install nvidia-cudnn-cu11==8.6.0.163
CUDNN_PATH=$(dirname $(python -c "import nvidia.cudnn;print(nvidia.cudnn.__file__)"))
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/:$CUDNN_PATH/lib
python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"
File: predict.py
Description:
Use any model to predict a 3D structure. This requires a backbone in a .pdb structure. The side-chains of the residues will be automatically removed by aposteriori, thus the prediction will be performed uniquely on the empty backbone. Your chosen model will attempt to predict which residues best fit the position and will return a .fasta file as well as a probability distribution in .csv format.
-
Make a folder with all the pdb files you want to predict
-
Create the dataset using aposteriori (For more info about other options, please see https://github.com/wells-wood-research/aposteriori/)
make-frame-dataset $YOUR_PDB_FOLDER -e $YOUR_PDB_EXTENSION --voxels-per-side 21 --frame-edge-length 21 -g True -p 6 -n dataset -v -r -cb True -ae $YOUR_CODEC --compression_gzip True -o . --voxelise_all_states True$YOUR_PDB_FOLDER is a path to your folders containing PDB files. These need to be all with the same format (e.g. .pdb)
$YOUR_PDB_EXTENSION is the file extension of your PDB files in $YOUR_PDB_FOLDER
$YOUR_CODEC This will normally be CNOCBCA except for TIMED_Charge CNOCBCAQ and TIMED_Polar CNOCBCAP
For a sample dataset use:
make-frame-dataset tests/testing_files -e .pdb1.gz --name data --voxels-per-side 21 --frame-edge-length 21 -p 8 -vrz -cb True -ae CNOCBCA -g True- Download your model of interest from:
https://github.com/wells-wood-research/timed-design/releases
For example, download TIMED.h5 . Make sure to place it in this directory.
- Finally run:
python3 predict.py --path_to_dataset $DATASET_PATH --path_to_model MODEL_PATH --path_to_output OUTPUT_DIRECTORY
eg.
python3 predict.py --path_to_dataset data.hdf5 --path_to_model TIMED.h5 --path_to_output .
The output files include:
-
{MODEL_NAME}.csv, Contains probability distribution over all amino acids at each position of the protein as predicted by the selected model. -
{MODEL_NAME}.fasta, Protein sequences generated by selecting the amino acid with the highest probability (ArgMax) at each position as indicated by the probability predictions in the{MODEL_NAME}.csvfile. -
{MODEL_NAME}.txt, A file listing the the protein/ chain ID and the number of its residues for amino acid sequences found in {MODEL_NAME}.fasta file. -
dataset.fasta, Fasta sequences of the input protein structures. -
datasetmap.txt, detailed mapping of the protein sequences presented indataset.fasta, with each line including the protein identifier, chain identifier, residue sequence number, and residue type. -
encoded_labels.csv, Sequences fromdataset.fastaencoded in labels.
In order to use a rotamer model, use the flag --predict_rotamers True:
python3 predict.py --path_to_dataset dataset.hdf5 --path_to_model timed_rot.h5 --predict_rotamers True
To run the UI you must have streamlit installed. Then run:
streamlit run ui.py -- --path_to_models /models --path_to_pdb /pdb --path_to_data /data --workers 8
If you prefer to use Docker, you can use the Dockerfile provided.
docker build -t timed_design .
docker run -v /local/path/to/models/:/scratch/timed_dataset/models -v /local/path/to/biounit:/scratch/datasets/biounit -p 8501:8501 timed_design
Enter the directory if you are not:
cd timed-design
Install with pip:
pip install .
File: sample.py
Description:
Uses Monte Carlo sampling to sample sequences from a probability distribution. A temperature factor can be applied to affect the distributions. It will return a .fasta file and/or a .json file with the sequences and a .csv file with basic sequence metrics such as isoelectric point, molecular weight and charge. Further metrics can be calculated using NetSolP-1.0 (see scripts/run_netsolp.sh).
Run the prediction step using predict.py as instructed in section 1.2.
python sample.py --path_to_pred_matrix /PATH/TO/{MODEL_NAME}.csv --path_to_datasetmap /PATH/TO/{MODEL_NAME}.txt --seed 42 --workers 35 --sample_n 200 --temperature 0.1
Replace the path placeholders with the actual paths to your files generated in the prediction step.
eg.
python sample.py --path_to_pred_matrix TIMED.csv --path_to_datasetmap TIMED.txt --seed 42 --workers 35 --sample_n 200 --temperature 0.1
This will sample 200 sequences using the provided probability distribution and temperature factor of 0.1. Output sampled sequences can be increased or decreased by adjusting the number of output structures ''--sample_n 200''.
Sampling temperature controls the randomness of predictions made by the model. By adjusting the temperature, you can either steer the model towards generating more predictable sequences (exploitation) or more diverse sequences (exploration). This parameter can range from 0 to any positive number. However, we recommend sampling at temperatures between 0.1 and 1 because sequences predicted for temperature > 1 exhibit high variations in their Root-Mean-Square-Deviation (RMSD).
-
Temperature = 0: This is equivalent to taking the most probable next token every time, also known as the argmax. At this temperature, the model's output is entirely deterministic and focused on the most likely sequence.
-
Exploitation (0<T<1): Setting the temperature to a low value, such as 0.2, results in sequences that are very similar and close to the argmax, making the generated sequences less diverse but more predictable and focused.
-
Exploration with High Temperature (T≈1): Setting the temperature equal or close to 1, keeps the distribution closer to its original form. This allows for more randomness, resulting in a greater variety of sequences. Hence, T=1 enhances the diversity of the generated sequences (exploration) while retaining the original probability distribution.
-
Enhanced Exploration with Higher Temperatures (T>1): At temperatures greater than 1, the logits are reduced, making the probability distribution flatter and more uniform (higher entropy of the data). This increases the likelihood of selecting less probable classes, leading to highly diverse sequences.
File: analyse_rotamers.py
Description:
---Under construction---
Proteins are macro-molecules present in all living organisms (and viruses). They are involved in important chemical reactions in metabolism, DNA replication and also have structural properties.
The 3D shape of the protein defines its function. Proteins are made of twenty subunits called amino acids (or residues), all of which have different atomic structures and chemical properties. Different combinations of amino acids produce different 3D shapes.
The Protein Folding Problem aims at identifying the 3D shape of a protein based solely on the sequence of amino acid. This problem is being tackled by models like AlphaFold2 by Google Deep Mind.
The other side of this problem is the Inverse Folding Problem (Protein Sequence Design), that is, given a desired 3D structure with a useful function, identify the residue sequence that will reliably fold into this structure. This problem is arguably harder, as multiple sequence of amino acids can fit the same 3D shape.
Nature evolves by natural selection (depth first search), thus, the protein sequences sampled in nature account for a tiny fraction of all physically possible proteins. Even for a relatively small protein with around 200 amino acids, there are around 10^260 possible sequences, which is significantly more than that have been sampled in every cell of every organism since proteins arose. There are therefore many 3D shapes of proteins that are physically possible, potentially useful, which have not been explored by nature. Protein designers aim at unlocking the potential of this pool of unobserved proteins, known as the dark matter of protein folding space.
TIMED (Three-dimensional Inference Method for Efficient Design) is a Convolutional Neural Network (CNN) model to tackle the Inverse Folding Problem, developed by us. We also feature our implementation of models described in the literature but unavailable for public use. These models take a 3D shape (empty backbone) as input and predict the identity of the subunits at each position of the backbone, as shown below:
The input of the model is a cube of gridded, voxelised space (a "Frame") around each amino acid position of the backbone. The alpha Carbon is centered in the frame, and the frame is rotated so that the Alpha Carbon to Carbon bond lies along the x-axis. Each atom (C, N, O, alpha-C, beta-C) is one-hot-encoded in a different channel, thus producing a 4D array. The beta-Carbon position is hard-coded to the average position of all beta-Carbon in the protein 1QYS after the aforementioned rotations.
For a 100-amino-acid protein we therefore generate 100 frames of equal width, height and length and feed them to our models. To produce these frames we use a library we developed called aposteriori.
The output of our models is a probability distribution over all amino acids at each position. For instance, at each position the models output a probability over every residue being at that position. For a 100-amino-acid protein will have an output of shape (100, 20) as there are twenty amino acids:
The TIMED architecture features a Convolutional Block composed of a 3D Convolution Operation, followed by ELU Activation and Batch Normalisation. We feature several convolution blocks which end with Spatial Dropout and a Global Average Pooling layer rather than a Dense layer. The output of the softmax layer is either 20 (for the 20 residues) or 338 (for rotamers). The architecture is illustrated below
Amino acids are made of chemical bonds that can rotate in space, thus resulting in a different spacial configuration of the residue. These are called Rotational Isomers (Rotamers), and longer residues like Lysine (K), tend to have more rotamers.
Rather than predicting 20 amino acid, we built models to predict both the amino acid AND the rotamer configuration in space. Therefore, for one frame our model identifies both, the amino acid identity and its predicted conformation in space:
The rotamer models therefore predict 338 classes rather than 20. Rotamer models tend to significantly outperform conventional models, even with the same network structure, while also providing increased granularity of predictions which can then be further refined using molecular dynamics simulations.
We are the first in the field to validate rotamer models and provide them and various implementation of other models, free for use.
As the output of our models is a probability distribution, an alternative to picking the amino acid with the highest probability (argmax), is to sample from the probability distribution using methods like Monte Carlo. For one protein shape we can therefore generate several sequences that can then be screened for specific properties like charge, isoelectric point, solubility, and expressivity.
We provide a simple CLI interface to generate several of these design which output to a .fasta file.
@article{timed,
author = {Castorina, Leonardo V and Ünal, Suleyman Mert and Subr, Kartic and Wood, Christopher W},
title = "{TIMED-Design: Flexible and Accessible Protein Sequence Design with Convolutional Neural Networks}",
journal = {Protein Engineering, Design and Selection},
pages = {gzae002},
year = {2024},
month = {01},
abstract = "{Sequence design is a crucial step in the process of designing or engineering proteins. Traditionally, physics-based methods have been used to solve for optimal sequences, with the main disadvantages being that they are computationally intensive for the end user. Deep learning based methods offer an attractive alternative, outperforming physics-based methods at a significantly lower computational cost.In this paper, we explore the application of Convolutional Neural Networks (CNNs) for sequence design. We describe the development and benchmarking of a range of networks, as well as reimplementations of previously described CNNs. We demonstrate the flexibility of representing proteins in a three-dimensional voxel grid by encoding additional design constraints into the input data. Finally, we describe TIMED-Design, a web application and command line tool for exploring and applying the models described in this paper.The User Interface (UI) will be available at the URL: https://pragmaticproteindesign.bio.ed.ac.uk/timed. The source code for TIMED-Design is available at https://github.com/wells-wood-research/[email protected] data are available at Journal Name online.}",
issn = {1741-0126},
doi = {10.1093/protein/gzae002},
url = {https://doi.org/10.1093/protein/gzae002},
eprint = {https://academic.oup.com/peds/advance-article-pdf/doi/10.1093/protein/gzae002/56453873/gzae002.pdf},
}