Source-linked AI summary
DeepDTA: Deep Drug-Target Binding Affinity Prediction
Hakime Öztürk, Elif Ozkirimli, Arzucan Özgür
TL;DR
Predicting continuous drug-target binding affinity remains challenging because earlier approaches largely treated interactions as binary classification. DeepDTA uses CNNs to learn representations from protein sequences and compound SMILES, achieving better-than-baseline performance, including a CI of 0.863 on KIBA.
Problem
Drug-target prediction has largely used binary classification, overlooking binding-affinity values and relying on difficult-to-obtain true negatives or synthetic negatives.
Method
DeepDTA uses separate CNN blocks to learn representations from raw protein sequences and compound SMILES, then combines them for affinity prediction.
Results
DeepDTA’s combined CNN model achieved the best KIBA CI of 0.863 and significantly outperformed both baselines, while improving over baselines on Davis and KIBA.
Takeaways & Limitations
Raw sequence representations can match or exceed baseline performance for drug-target affinity prediction without relying on multiple external feature-extraction tools.
Takeaways & Limitations
A CNN-based protein representation with 2D compound similarity performed poorly, suggesting CNNs may not capture amino-acid ordered relationships effectively.
Abstract
from arXiv · showhide
The identification of novel drug-target (DT) interactions is a substantial part of the drug discovery process. Most of the computational methods that have been proposed to predict DT interactions have focused on binary classification, where the goal is to determine whether a DT pair interacts or not. However, protein-ligand interactions assume a continuum of binding strength values, also called binding affinity and predicting this value still remains a challenge. The increase in the affinity data available in DT knowledge-bases allows the use of advanced learning techniques such as deep learning architectures in the prediction of binding affinities. In this study, we propose a deep-learning based model that uses only sequence information of both targets and drugs to predict DT interaction binding affinities. The few studies that focus on DT binding affinity prediction use either 3D structures of protein-ligand complexes or 2D features of compounds. One novel approach used in this work is the modeling of protein sequences and compound 1D representations with convolutional neural networks (CNNs). The results show that the proposed deep learning based model that uses the 1D representations of targets and drugs is an effective approach for drug target binding affinity prediction. The model in which high-level representations of a drug and a target are constructed via CNNs achieved the best Concordance Index (CI) performance in one of our larger benchmark data sets, outperforming the KronRLS algorithm and SimBoost, a state-of-the-art method for DT binding affinity prediction.
Introduction
The introduction motivates binding-affinity prediction as a richer alternative to binary drug-target interaction classification and proposes learning from protein and ligand sequences with deep learning. The approach uses CNN blocks to derive representations from raw protein sequences and compound SMILES rather than external features or 3D complex structures.
- Motivation: Binding affinity captures the strength of a drug-target interaction, unlike binary classification, which determines only whether the pair interacts.Affinity is commonly expressed using Kd, Ki, or IC50; low IC50 and Ki values indicate stronger binding.
- Motivation: Binary drug-target prediction is hindered by the scarcity of true-negative samples and the effects of synthetic negative-sample generation on algorithm performance.Unknown-binding pairs are often treated as non-binding samples.
- Related work: Protein-ligand affinity prediction is used in scoring after virtual screening and docking to estimate the putative strengths of proposed ligands.Random Forest methods have been used as alternatives to scoring functions that depend on multiple parameters.
- Related work: Prior affinity work included SimBoost, which used gradient boosting, similarity-based information, network-derived features, and 2D compound representations.These studies relied on traditional machine learning algorithms and engineered features.
- Contribution: The proposed approach predicts protein-ligand binding affinities from protein sequences and compound SMILES using CNN blocks to learn representations from raw inputs.It avoids external features and 3D structures of binding complexes.
Materials and Methods
DeepDTA was evaluated on the Davis and KIBA benchmark datasets as a regression model using sequence-based representations of compounds and proteins. Inputs were integer-encoded, length-standardized sequences processed by separate CNN blocks for SMILES and protein sequences.
- Datasets: The model was evaluated on the Davis and KIBA benchmark datasets for binding-affinity prediction.Davis contains 442 proteins and 68 ligands with Kd values; KIBA combines Ki, Kd, and IC50 bioactivities into KIBA scores.
- Affinity preprocessing: Davis Kd values were transformed into log-space pKd values, while KIBA scores were pre-processed by sign inversion and shifting.The Davis distribution includes 20,931 of 30,056 observations at pKd 5, representing true negatives with very weak or unobserved binding.
- Input representation: Compound SMILES and protein sequences were integer-encoded and padded or truncated to dataset-specific maximum lengths.Maximum lengths were 85 SMILES and 1200 proteins for Davis, and 100 SMILES and 1000 proteins for KIBA, covering at least 90% of compounds and 80% of proteins.
- Model architecture: DeepDTA treated interaction prediction as regression and used separate CNN blocks to learn representations from SMILES and protein sequences.Each block used three consecutive 1D-convolutional layers with increasing filters, followed by max pooling; ReLU activations and mean squared error were used.
Experiments and Results
DeepDTA was evaluated on benchmark Davis and KIBA datasets using Concordance Index and MSE under matched cross-validation settings against KronRLS and SimBoost. Learning compound and protein representations jointly with CNNs achieved the strongest reported KIBA performance and statistically significant improvements over both baselines.
- Evaluation and baselines: DeepDTA was compared with KronRLS and SimBoost using Concordance Index (CI) to evaluate continuous binding-affinity predictions.MSE was also used as an evaluation metric.
- Experimental setup: The model was evaluated on benchmark datasets using six-way partitioning, five-fold cross-validation for hyper-parameters, and identical train-test folds for both baselines.The independent test set was one of six equal data partitions.
- Ablation results: Using predefined similarity features with fully connected networks reduced CI to 0.79 on Davis and 0.71 on KIBA, while MSE increased to more than 0.5.These results motivated CNN-based representation learning for drugs and proteins.
- Ablation results: The compound-CNN and protein-Smith-Waterman combination outperformed both baselines significantly, with p-value of 0.0001 for both on KIBA and around 0.03 for both on Davis.This suggested CNNs captured more information than Pubchem Sim for compound representation.
- Final model results: The combined CNN model matched baselines at CI score of 0.878 on Davis and achieved the best CI score (0.863) on KIBA, significantly surpassing both baselines.The p-value was 0.0001 for both comparisons, and its MSE was notably lower than baseline MSE on both datasets.
Conclusion
DeepDTA predicts drug-target binding affinity from raw protein and drug sequences by learning representations with CNNs and fully connected layers. Its performance exceeded baseline methods, especially on the larger KIBA dataset, while future work targets improved protein-sequence representations and broader prediction of novel targets and drugs.
- Method: CNNs learn protein and drug representations from raw sequences, which are combined with fully connected layers for affinity prediction.Experiments compare DeepDTA with KronRLS and SimBoost on the Davis kinase-drug and KIBA datasets.
- Results: CNN-based raw-sequence representations significantly improved performance over KronRLS and SimBoost on both KIBA and Davis datasets.Predefined features used with DeepDTA were insufficient to describe protein-ligand interactions.
- Limitations and future work: The protein-CNN and 2D-compound-similarity model performed poorly, motivating LSTM-based protein-sequence modeling and improved protein representations as future work.The authors also plan to use a ligand-based protein representation based on SMILES sequences of interacting ligands and extend prediction to novel targets and drugs.
- Results: 0.863 was the KIBA CI score after improvement from 0.836, compared with Davis improvement from 0.872 to 0.878.The deep-learning advantage was statistically significant and notably larger on KIBA, which is four times larger than Davis.
- Contribution: DeepDTA predicts drug-target binding affinity using only character-level sequence representations of proteins and drugs.The model uses raw sequence information rather than multiple tools and algorithms for feature extraction.