Source-linked AI summary
DeepConv-DTI: Prediction of drug-target interactions via deep learning with convolution on protein sequences
Ingoo Lee, Jongsoo Keum, Hojung Nam
TL;DR
DTI prediction is needed because experimental identification is costly, while conventional protein descriptors may not provide sufficiently accurate predictions. The paper applies CNNs to raw protein sequences to capture local residue patterns, and reports better performance than prior descriptor-based and deep-learning approaches, while pooled convolution results can detect protein binding sites.
Problem
DTI identification is costly, and conventional protein descriptors can provide insufficient predictive performance while losing local sequence information.
Method
The model applies convolutional neural networks to raw protein sequences to capture local residue patterns, then combines learned protein features with drug features for DTI prediction.
Results
The model exhibited better performance than previous protein descriptor-based models and a previous model using CTD and SW scores.
Takeaways & Limitations
Pooled convolution results can identify local residue patterns and detect protein binding sites for DTIs.
Takeaways & Limitations
Feature-based models do not show sufficiently good predictive performance, motivating the raw-sequence approach.
Abstract
from arXiv · showhide
Identification of drug-target interactions (DTIs) plays a key role in drug discovery. The high cost and labor-intensive nature of in vitro and in vivo experiments have highlighted the importance of in silico-based DTI prediction approaches. In several computational models, conventional protein descriptors are shown to be not informative enough to predict accurate DTIs. Thus, in this study, we employ a convolutional neural network (CNN) on raw protein sequences to capture local residue patterns participating in DTIs. With CNN on protein sequences, our model performs better than previous protein descriptor-based models. In addition, our model performs better than the previous deep learning model for massive prediction of DTIs. By examining the pooled convolution results, we found that our model can detect binding sites of proteins for DTIs. In conclusion, our prediction model for detecting local residue patterns of target proteins successfully enriches the protein features of a raw protein sequence, yielding better prediction results than previous approaches.
Introduction
DTI prediction methods face costly experiments and limitations in conventional protein descriptors, which can lose local sequence information. DeepConv-DTI applies convolution to raw protein sequences for diverse protein classes and lengths, improving performance over earlier approaches.
- Chemical and biological experiments for identifying DTIs can require 2~3 years and substantial costs.
- Similarity-based methods work well within specific protein classes but not across other classes.
- Feature-based models using protein descriptors and drug fingerprints showed worse performance than previous conventional QSAR models.
- Transforming amino acid sequences into conventional features can lose local residue patterns and relationships that are difficult to recover with traditional machine learning.
- DeepConv-DTI uses convolution filters across entire raw protein sequences to capture local residue patterns for massive-scale DTI prediction across protein classes and lengths.The model pools maximum convolution results, combines learned protein features with drug fingerprint features, and predicts DTI probabilities through fully connected layers.
- The model exhibited better performance than a previous model using CTD and SW scores.
Results and Discussion
Figure 2 compares optimized protein-descriptor models using AUPR and AUC for convolution, CTD, and similarity descriptors.
- AUPR and AUC are shown for convolution, CTD, and similarity descriptors in panels A and B, respectively.
Performances of the validation dataset and selected hyperparameters
The authors tuned model hyperparameters and evaluated the optimized convolution model on validation and independent PubChem datasets. The model achieved an AUPR of 0.832 on an unseen validation dataset, although the validation set was too small for broad performance assessment.
- 0.832 AUPR was obtained on an unseen validation dataset after hyperparameter selection.The search evaluated learning rate, window configuration, hidden layers, and concatenating layers.
- The validation dataset was too small to evaluate general performance reliably against the similarity method.The model's AUPR was lower than the similarity descriptor's AUPR on this validation set.
- The convolution model showed better performance on independent PubChem datasets than models using other protein descriptors.
Comparison of performance with other protein descriptors
Across independent datasets, the convolution model outperformed conventional protein descriptors on key classification measures, while descriptor informativeness varied by dataset and protein representation.
- The convolution model outperformed other protein descriptors across all independent datasets.The comparison included CTD and normalized SW score descriptors.
- The model performed equally well on PubChem and KinaseSARfari using an equal-error-rate threshold, suggesting general application power.
- The convolution model achieved the highest accuracy and F1 score on PubChem and its subsets.
- CTD produced the lowest score for every dataset and metric, indicating lower informativeness and enrichment than the other descriptors.
- Similarity descriptors performed similarly to the proposed model on KinaseSARfari, potentially reflecting informative local residue patterns at the domain level.The passage contrasts domain-level information with representation of the whole protein complex.
Performance comparison with a previous model
The authors compared their convolution model with Wen et al.'s DBN-based model under aligned evaluation conditions. Their model performed better, while pooled convolution outputs also supported binding-region detection analyses.
- The comparison used the previous model's optimized shape hyperparameters and descriptors while evaluating both models on PubChem.
- Fine-tuning used a learning rate of 0.01 because the previous model could not learn with a rate of 0.1.
- The convolution model performed better than the previous DBN-based model on the PubChem comparison.The previous model was built by Wen et al. and used deep belief networks.
- Figure 6 visualizes annotated binding sites and pooled convolution boxes for P53_HUMAN and ESR1_HUMAN.
Detection of the binding site by CNN on the sequence
The model’s pooled convolution results highlight protein regions associated with annotated binding sites, supporting detection of local residue patterns relevant to DTIs. Validation used sc-PDB binding-site annotations and randomization-based statistical testing.
- The pooled results cannot be measured exactly for their effects on DTI prediction scores and also cover more complicated sites.
- The model captures local residue patterns that participate in DTIs, as shown by examination and validation of intermediate convolution results.
- The analysis assumes pooled maximum convolution results cover important protein regions, including binding sites.Each window contains 128 pooled convolution results, producing bias in which regions are covered.
- Convolution results showed high values when binding sites were included, and pooled results exactly covered annotated p53 binding sites with high filter ranks.
- For estrogen receptor ESR1_HUMAN, four binding sites were covered by high-ranked convolution results across filters in windows.
t-SNE visualization of proteins
The authors visualized intermediate protein features with t-SNE to examine captured local residue patterns. The resulting representation roughly discriminated proteins by class, although class discrimination was not the model’s intended task.
- The visualization examined protein features from the fully connected layer after global max-pooling of convolution results.
- t-SNE reduced the high-dimensional protein features to low-dimensional representations for visualization.
- The intermediate representation could roughly discriminate protein classes, despite protein-class identification not being the model’s intended purpose.
Materials and Methods
The study assembled training, validation, and independent experimental-assay datasets from multiple databases, retaining Vertebrata proteins and constructing negative DTI samples through specified sampling procedures.
- Known DTIs were obtained from DrugBank, KEGG, and IUPHAR, with duplicate interactions removed across databases.
- The final training collection contained 11,950 compounds, 3,675 proteins, and 32,568 DTIs after restricting proteins to Vertebrata.
- The collected DTIs were treated as positive training samples, while negative DTIs were not defined in the source data.
- Ten negative DTI sets were constructed to reduce bias from random negative-interaction generation.
- Independent test datasets used experimental assays from PubChem BioAssay and ChEMBL KinaseSARfari.PubChem positives used Active assays with Kd < 10μm, while negative samples were selected from Inactive assays under the described sampling rules.
- The PubChem dataset contained 18,228 positive and negative samples involving 21,907 drugs and 698 proteins, with subsets for new compounds, proteins, and DTIs.
Drug feature representation
The model represents proteins with raw sequences but represents drugs using Morgan/Circular fingerprints derived from SMILES strings. Each drug becomes a 2,048-dimensional binary vector encoding molecular substructures.
- The protein input is the raw protein sequence, whereas the drug input is not the raw SMILES string.
- Morgan/Circular fingerprints analyze molecules as graphs and retrieve substructures from subgraphs of the whole molecular graph.
- RDKit generated the Morgan/Circular fingerprint with radius 2 from each raw SMILES string.
- Each drug is represented as a 2,048-dimensional binary vector whose indices indicate the presence of specific substructures.
Selection of hyperparameters
Hyperparameters were tuned with external validation rather than a training subset, using staged selection and grid search to optimize AUPR. The model uses fully connected layers for nonlocal numerical protein descriptors.
- External validation: External validation was used to select hyperparameters instead of relying on a training-data subset.The negative datasets were randomly sampled, motivating the external validation step.
- Selection procedure: Learning rate was selected first, followed by regularization parameters such as dropout ratio.The stated rationale is that a high learning rate cannot learn a pattern.
- Selection procedure: Grid search optimized the remaining hyperparameters, with the best settings identified by AUPR.AUPR was treated as an appropriate metric for positive-classification accuracy.
- Descriptor comparison: For numerical protein descriptors without locality, the model used fully connected layers rather than convolutional processing.These descriptors were represented as numerical vectors that do not have locality.
- Threshold selection: When measuring AUPR, the optimal classification threshold was determined using EER.EER minimizes a precision–recall cost expression controlled by γ, which was set to 2 in this model.
Evaluation of performances
Performance was measured on an independent test dataset after fixing the classification threshold. The evaluation reported sensitivity, specificity, precision, accuracy, and F1.
- Evaluation setup: The model was evaluated on an independent test dataset after the classification threshold was fixed.This separates threshold selection from final performance measurement.
- Evaluation metrics: Reported performance metrics were sensitivity, specificity, precision, accuracy, and the F1 measure.The metrics are abbreviated Sen, Spe, Pre, Acc, and F1, respectively.