InterpretContactMap

Deep learning methods for interpreting protein contact maps

Contribute

This work has been developed and used by the Bioinformatics, Data Mining and Machine Learning Laboratory (BDM) . Help from every community member is very valuable to make the tool better for everyone. Checkout the Lab Page.

Required software:

PSI-BLAST 2.2.26 (For generating PSSM sequence profile)

CCMpred (For generating pseudo-likelihood maximization)

Python 3.6

Required Python modules:

numpy 1.18.1
pandas 1.1.2
tensorflow-gpu(Recommended if a CUDA-compatibale GPU is available) /tensorflow 1.15.2 
keras 2.1.6

We also provided the Dockerfile to help users setup the required environment when Docker is available on the host system. Please refer to the Docker Tutorials (https://www.docker.com/101-tutorial) for how to build your own image under different systems.

Feature geraration

Two types of features are required: PSSM sequence profile which can be generated from PSI-BLAST, and PLM (pseudo-likelihood maximization) which can be generated from CCMpred from multiple sequence alignments (MSAs) produced by DeepMSA. The details of how to acquire both features are described in the DeepDist paper, which is also developed by the BDM lab.

The sequence databases used in the DeepMSA homologous sequences search include Uniclust30 (2017-10), Uniref90 (2018-04) and Metaclust50 (2018-01), our in-house customized database which combines Uniref100 (2018-04) and metagenomics sequence databases (2018-04), and NR90 database (2016). Sample features can be found under the example folder, and users can build both features from their own customized sequence databases.

Prediction

Predict from given PLM and PSSM data (predict.py):

-h, --help show this help message and exit
-m, --model_type Type of model, can be one of "sequence", "regional", or "combine"
-l, --plm_data Path to PLM data. Should be a numpy array flatten from (441,L,L), where L is the length of the input sequence. It should be saved as .npy format (https://numpy.org/doc/stable/reference/generated/numpy.save.html).
-s, --pssm_data Path to PSSM data. Should be a text file start with " # PSSM" as the first line, and the following contents should be 20 lines each contains L values.
-o, --out_file Path to output contact map. An L by L numeric matrix saved as TSV format.
-w, --weights Should attention weights be extracted. Sequence attention weights would have shape (heads,L,L). Regional attention weights would have shape (heads,n²,L,L), where n is the side length of the scope of attention mechanism. Both weights are saved as .npy format. Detailed description of how attention weights are computed can be found at https://www.biorxiv.org/content/10.1101/2020.09.04.283937v1.

Example:

python predict.py -m sequence -l example/plm/T0970.plm -s example/other/T0970.pssm.npy -o ./outmap.tsv

Training

Train models for protein contact prediction with given PLM, PSSM and labels for training and validation datasets (train.py):

-h, --help show this help message and exit.
-plm, --plm_data_path Path to PLM data files.
-pssm, --pssm_data_path Path to PSSM data files.
-l, --label_path Path to label data files. Should be text files with 0/1 indicate the contacts.
-s, --sample_list_file Config file indicating the sample names for training and validation.
-m, --model_type Type of model, can be "sequence" or "regional".
-o, --output_dir Path where the trained models and history are saved.
-pa, --patience Stop the training early for no improvements in validation after x epochs, default is 5.
-e, --epochs Number of epochs, default is 60.