Explainable Deep Drug-Target Representations for Binding Affinity Prediction

We explore the reliability of Convolutional Neural Networks (CNNs) in the identification of important regions for binding, and the significance of the deep representations by providing explanations to the model’s decisions based on the identification of the input regions that contributed the most to the prediction. Furthermore, we implement an end-to-end deep learning architecture to predict binding affinity, where CNNs are exploited in their capacity to automatically surmise and extract discriminating deep representations from 1D sequential and structural data.

End-to-End Deep Learning Architecture: Convolutional Neural Networks + Feed-Forward Fully Connected Neural Network

Chemogenomc Representative K-Fold

Regression Discriminative Localization Map

3D Docking Visualization

• Potential Binding Sites (≤ 5 Å) : Green

• L-Grad-RAM Hits : Blue

• Matched Binding - L-Grad-RAM Hits : Red

ABL1(E255K)-phosphorylated - SKI-606

DDR1 - Foretinib

Binding Affinity Prediction Model

• Two Parallel Convolution Neural Networks + Fully Connected Neural Network

Gradient-Weighted Regression Activation Mapping (Grad-RAM)

• Global Max Pooling + Guided Gradients
• Global Max Pooling + Non Guided Gradients
• Global Average Pooling + Guided Gradients
• Global Average Pooling + Non Guided Gradients

Davis Kinase Binding Affinity

Dataset

• davis_original_dataset: original dataset
• davis_dataset_processed: dataset processed : prot sequences + rdkit SMILES strings + pkd values
• deep_features_dataset: CNN deep representations: protein + SMILES deep representations

Clusters

• test_cluster: independent test set indices
• train_cluster_X: train indices

Similarity

• protein_sw_score: protein Smith-Waterman similarity scores
• protein_sw_score_norm: protein Smith-Waterman similarity normalized scores
• smiles_ecfp6_tanimoto_sim: SMILES Morgan radius 3 similarity scores

Binding

• davis_scpdb_binding: davis-scpdb matching pairs binding information

PSSM

• pssm_X: davis-scpdb matching pairs PSSM

sc-PDB Pairs

Binding

• scpdb_binding: scpdb pairs binding information

PSSM

• pssm_X: scpdb pairs PSSM

Dictionaries

• davis_prot_dictionary: AA char-integer dictionary
• davis_smiles_dictionary: SMILES char-integer dictionary

State-of-the-Art Baselines Data

Davis Kinase Binding Affinity Dataset + Clusters in the SOTA method format

Docking

• abl1_pymol.pse: ABL1(E255K)-phosphorylated - SKI-606 PyMol Session
• ddr1_pymol.pse: DDR1 - Foretinib PyMol Session

Requirements:

• Python 3.7.9
• Tensorflow 2.4.1
• Numpy
• Pandas
• Scikit-learn
• Itertools
• Matplotlib
• Seaborn
• Glob
• Json

Usage:

Binding Affinity Prediction

Training

python cnn_fcnn_model.py --option Training --num_cnn_layers_prot 3 --prot_filters 64 64 128 --prot_filters_w 4 4 5 --num_cnn_layers_smiles 3 --smiles_filters 64 64 128 --smiles_filters_w 4 4 5 --num_fcnn_layers 3 --fcnn_units 1024 512 1024 --drop_rate 0.5 0.1 --lr_rate 0.0001 

Validation

python cnn_fcnn_model.py --option Validation --num_cnn_layers_prot 3 --prot_filters 64 64 128 --prot_filters_w 4 4 5 --num_cnn_layers_smiles 3 --smiles_filters 64 64 128 --smiles_filters_w 4 4 5 --num_fcnn_layers 3 --fcnn_units 1024 512 1024 --drop_rate 0.5 0.1 --lr_rate 0.0001 

Evaluation

python cnn_fcnn_model.py --option Evaluation

Gradient-weighted Regression Activation Mapping (L-Grad-RAM)

Example

• Protein Sequence : MLEICLKLVG...
• SMILES String : Cc1cn(...
• Window Length : 0 1 2 ...
• Feature Importance Threshold : 0.3 0.4 0.5 ...
• Binding Sites Positions : 5 10 50 ...

python gradram_testing.py --protein_sequence MLEICLKLVG... --smiles_string Cc1cn(... --window 0 1 2 ... --thresholds 0.3 0.4 0.5 ... --sites 5 10 50 ...