This repository contains the full source code and a detailed example of a case-of-use of DeeProtGO, a deep learning model for automatic function prediction by integrating protein knowledge from multiple databases.
The following sections contain the steps and the libraries used in each of them. All datasets and libraries are open sourced and free to use. If you use any of the following methods, please cite them properly:
Gabriela A Merino, Rabie Saidi, Diego H Milone, Georgina Stegmayer, Maria J Martin, Hierarchical deep learning for predicting GO annotations by integrating protein knowledge, Bioinformatics, Volume 38, Issue 19, 1 October 2022, Pages 4488–4496, https://doi.org/10.1093/bioinformatics/btac536
This workflow describes the steps required to predict GO terms for No-knowledge (NK) proteins, that is, without experimental annotations in any of the GO sub-ontologies at a reference time, but that have accumulated at least one GO term with an experimental evidence code during a growth period.
Data used for developing and evaluating the proposed models were obtained from different protein-knowledge databases, considering the proteins provided in the CAFA3 challenge. The CAFA3 dates were considered:
Since DeeProtGO was developed to learn new annotations gained during a time gap, it is trained here with proteins that gained GO terms during the growth period defined between t-1 and t0, and evaluated on NK proteins of the CAFA3 benchmark dataset (at t1).
All the data required for training and testing DeeProtGO to predict BP terms for NK proteins is already provided in the data folder.
Note: If you want to replicate the procedure, you can follow the instructions in the data preparation notebook. Doing so may take several hours.
The DeeProtGO training notebook is provided with all the steps required to train DeeProtGO for predicting BP terms for NK proteins.
Note: All the results generated in this section are provided in the examples/train_NK_EUKA_BP directory.
The DeeProtGO testing notebook is provided with all the steps required to evaluate DeeProtGO when predicting BP terms for NK proteins of eukarya organisms in the CAFA3 benchmark dataset.
The following scheme illustrates the organization of this repository and contains a detailed description of each folder and file.
Protein function prediction
│
├── data -> All data for preparing files for DeeProtGO training and testing.
│ │
│ ├── extras -> Files useful for datasets preparation.
│ │ │
│ │ └── gene_ontology_edit.obo -> GO file with the ontology used for CAFA3.
│ │
│ ├── intermediate -> Intermediate files generated during training and benchmark datasets preparation.
│ │ │
│ │ ├── propAnnot_Bench_Euka_BP.tab -> Propagated GO annotations gained between t0 and t1 for benchmark
│ │ │ proteins.
│ │ └── propAnnot_Train_Euka_BP.tab -> Propagated GO annotations gained between t-1 and t0 for training
│ │ proteins.
│ ├── processed -> Datasets used for training and testing DeeProtGO when predicting annotations for NK proteins.
│ │ │
│ │ ├── Benchmark -> Data for NK proteins from eukarya organisms in CAFA3 benchmark for BP.
│ │ │ │
│ │ │ ├── Emb_BP_Euka.h5 -> SeqVec embeddings.
│ │ │ │
│ │ │ ├── LevSim_BP_Euka.h5 -> Proteins similarity based on edit distance.
│ │ │ │
│ │ │ ├── NegEntries_Euka_BP.tab -> UniProt entry names of negative proteins.
│ │ │ │
│ │ │ ├── netOut_BP_Euka.h5 -> One-hot encoding matrix representing GO BP terms of benchmark proteins.
│ │ │ │
│ │ │ ├── PosEntries_Euka_BP.tab -> UniProt entry names of positive proteins.
│ │ │ │
│ │ │ └── Taxon_BP_Euka.h5 -> One-hot encoding matrix representing proteins taxon.
│ │ │
│ │ └── Training -> Data used for training DeeProtGO to predict GO BP terms of NK proteins from eukarya organisms.
│ │ │
│ │ ├── dataPreparation.sh -> Bash script for downloading proteins similarity file and
│ │ │ uncompressing *tar.gz files
│ │ ├── Emb_BP_Euka.h5.tar.gz -> SeqVec embeddings.
│ │ │
│ │ ├── GOTermsPropRel_Euka_BP_train.tab -> One-hot encoding of relationships between GO terms, used for scores
│ │ │ propagation.
│ │ ├── NegEntries_Euka_BP.tab -> UniProt entries of negative proteins.
│ │ │
│ │ ├── netOut_BP_Euka.h5.tar.gz -> One-hot encoding matrix representing GO BP terms of training
│ │ │ proteins.
│ │ ├── PosEntries_Euka_BP.tab -> UniProt entries of positive proteins.
│ │ │
│ │ └── Taxon_BP_Euka.h5 -> One-hot encoding matrix representing proteins taxon.
│ │
│ └── raw -> Metadata of training and benchmark proteins
│ │
│ ├── benchmarkNKEukaBPInfo.tab -> Data of benchmark proteins
│ │
│ └── trainingNKEukaBPInfo.tab -> Data of training proteins
|
│
├── examples -> Files obtained when running the DeeProtGO example provided here
│ │
│ └── train_NK_EUKA_BP -> Files obtained during training of DeeProtGO for predicting BP terms of NK proteins
│ │
│ ├── modelPreparation.sh -> Bash script for downlading a pre-trained DeeProtGO model.
│ │
│ ├── test_13052021.log -> DeeProtGO performance on test partition of the training dataset.
│ │
│ ├── train_13052021.log -> DeeProtGO loss on train partition of training the dataset.
│ │
│ └── valid_13052021.log -> DeeProtGO performance on validation partition of the training dataset.
|
│
├── scripts -> Python scripts for data preparation, model fitting and evaluation
│ │
│ ├── dataPreparation.ipynb -> Python notebook for preparing the data for training and testing DeeProtGO
│ │
│ ├── DeeProtGOTesting.ipynb -> Python notebook for training DeeProtGO
│ │
│ └── DeeProtGOTraining.ipynb -> Python notebook for testing DeeProtGO
│
├── src -> Source files of DeeProtGO
│
├── dataloader.py -> Class and methods for loading the data
│
├── DNN.py -> Deep neural network behind DeeProtGO
│
├── DNNModel.py -> DeeProtGO model and its functions
│
├── earlyStop.py -> Method for early stopping
│
├── logger.py -> Class and methods for saving performance measures and model loss during training and testing.
│
└── sampler.py -> Class and methods for generating random batchs for being used during model training.