Assigning function to protein-protein interactions: a weakly supervised BioBERT-based approach using PubMed abstracts

Assigning function to protein-protein interactions: a weakly supervised BioBERT-based approach using PubMed abstracts

This is the source code and data as is at the time when the paper is submitted

1. The typed PPI dataset created by processing Intact
2. Untyped version of PPI
3. AIMED - Part rep processed dataset
4. AIMED - All rep processed dataset

Aimed dataset 10 Fold F1-score: 60.46

10 Fold F1-score on AIMed dataset is 60.46

This experiment used the following parameters . The dataset is attached

export PYTHONPATH=./src
/usr/bin/python -m main_train_k_fold --batchsize 32 --cnn_kernel_size 3 --cnn_num_layers 2 --cnn_output 64 --dataset PpiAimedDatasetFactory --dropout_rate_cnn 0.5 --earlystoppingpatience 50 --embeddim 200 --embeddingfile PubMed-shuffle-win-2.bin.txt --epochs 1000 --fc_drop_out_rate 0.5 --fc_layer_size 256 --input_drop_out_rate 0.2 --learningrate 0.001 --log-level INFO --network RelationExtractorSimpleResnetCnnPosNetworkFactory --pool_stride 2 --pooling_kernel_size 3 --train_val_vocab_merge 1 --trainfile AIMedFull.json --weight_decay 1e-05

Data Extraction

Extracts imex dataset and converts them into data frame

Basic working version

0.5.0

Add pipeline