InterLabelGO+

InterLabelGO+ combines a deep learning model (InterLabelGO) and sequence homology search to perform predict a query protein's biological function, in the form of Gene Ontology (GO) terms. InterLabelGO uses the last three layers the ESM2 large language model to extract sequence features, which are then learned by a series of neural networks to predict GO terms under a new loss function that incorporates label imbalances and inter-label dependencies. These deep learning predicted terms are then combined with DIAMOND search results through a dynamic weighting scheme to derive the consensus prediction. InterLabelGO+ (as team Evans) was ranked among the top teams in the recent CAFA5 challenge.

System Requirements

InterLabelGO is developed under a Linux environment with the following software:

• Python 3.11.5
• CUDA 12.1
• cuDNN 8.9.6
• DIAMOND v2.1.8
• NVIDIA drivers v.535.129.03

Python packages are detailed separately in requirements.txt.

Set up InterLabelGO+

1. Downloading the xz file from https://seq2fun.dcmb.med.umich.edu/InterLabelGO/InterLabelGO+.tar.xz
2. Extract the file
3. Run setup_env.sh to create an InterLabelGO conda environment and download ESM2 models

Alternatively, you can create the conda environment manually:

conda create -n InterLabelGO python=3.11.5 -y
conda activate InterLabelGO
pip install -r requirements.txt

Then download ESM2 models:

this_file_path=$(dirname $(readlink -f $0))
mode_pt_url="https://dl.fbaipublicfiles.com/fair-esm/models/esm2_t36_3B_UR50D.pt"
regression_url="https://dl.fbaipublicfiles.com/fair-esm/regression/esm2_t36_3B_UR50D-contact-regression.pt"
save_path=$this_file_path/Data/esm_models
mkdir -p $save_path
wget -P $save_path $mode_pt_url
wget -P $save_path $regression_url

This will download the ESM2 models into Data/esm_models.

To update the current GOA database, run:

python update_data.py

This will download the latest GOA database and convert it into the required format.

Data Processing

Run the following command:

python prepare_data.py --make_db --ia

This will:

• Convert raw data (train_terms.tsv and train_seq.fasta) into required training data
• Create a DIAMOND database for the training data (--make_db)
• Create an Information Content file for the training data (--ia)
• Extract the ESM embeddings for the training data

To use stratified multi-label in k-fold (time-consuming), add the --stratifi argument.

All paths are specified in settings.py.

Model Usage

1. Prediction

python predict.py -w example -f example/seq.fasta --use_gpu

This will predict GO terms for the example sequence.

Additional options:

• Add --seqid to use sequence ID for combination of alignments and neural network method
• Add -c to continue from the last unfinished prediction (won't check for changes in the fasta file)

2. Retrain Models

python train_InterLabelGO.py

Training configuration is specified in settings.py.

Notes

1. The data processing code will overwrite the original data, and the training code will overwrite the original model.

Citation

Quancheng Liu, Chengxin Zhang, Lydia Freddolino (2024) InterLabelGO+: unraveling label correlations in protein function prediction Bioinformatics, 40(11): btae655.