PaperlayerView exact revision on arXiv

Source-linked AI summary

WideDTA: prediction of drug-target binding affinity

Hakime Öztürk, Elif Ozkirimli, Arzucan Özgür

arXiv:1902.04166v1q-bio.QMcs.LGstat.ML

TL;DR

Drug-target binding-affinity prediction is challenging because drug discovery searches a large chemical space. WideDTA addresses this with a word-based deep-learning model using textual protein and ligand representations, and it outperformed DeepDTA on KIBA while added domain, motif, and MCS information generally contributed no further benefit.

  • Problem

    Drug discovery searches a large chemical space, and developing potential drugs for new targets is expensive and time-consuming.

  • Method

    WideDTA combines word-based protein sequences, PROSITE domains and motifs, ligand SMILES words, and ligand maximum common substructure words.

  • Results

    CI 0.875 versus DeepDTA CI 0.863 on KIBA, with statistical significance (p-value 0.001).

  • Takeaways & Limitations

    Word-based representations can outperform character-based DeepDTA, while protein domains and motifs alone perform as well as the complete protein sequence.

  • Takeaways & Limitations

    The study focuses on two kinase benchmark datasets, limiting how broadly domain and motif findings may generalize to more diverse datasets.

Abstract

from arXiv · show

Motivation: Prediction of the interaction affinity between proteins and compounds is a major challenge in the drug discovery process. WideDTA is a deep-learning based prediction model that employs chemical and biological textual sequence information to predict binding affinity. Results: WideDTA uses four text-based information sources, namely the protein sequence, ligand SMILES, protein domains and motifs, and maximum common substructure words to predict binding affinity. WideDTA outperformed one of the state of the art deep learning methods for drug-target binding affinity prediction, DeepDTA on the KIBA dataset with a statistical significance. This indicates that the word-based sequence representation adapted by WideDTA is a promising alternative to the character-based sequence representation approach in deep learning models for binding affinity prediction, such as the one used in DeepDTA. In addition, the results showed that, given the protein sequence and ligand SMILES, the inclusion of protein domain and motif information as well as ligand maximum common substructure words do not provide additional useful information for the deep learning model. Interestingly, however, using only domain and motif information to represent proteins achieved similar performance to using the full protein sequence, suggesting that important binding relevant information is contained within the protein motifs and domains.

1 Introduction

WideDTA addresses drug-target binding-affinity prediction using text-only protein and ligand information, extending sequence representations with protein motifs, domains, and ligand MCS words. On KIBA, it outperformed DeepDTA, while the added information sources did not improve the sequence-and-SMILES model.

  • Motivation: WideDTA predicts protein-ligand binding affinity from text-only information without relying on 3D complex structures or engineered compound representations.It learns features from protein sequences and ligand SMILES while integrating additional text-based information.
  • Method: The model represents proteins with sequence, domain, and motif information, and ligands with SMILES and maximum common substructure words.These four sources form the WideDTA input representation.
  • Findings: The added protein domain, motif, and ligand MCS information did not provide additional useful information when sequence and SMILES were already available.The study evaluates these additions as extensions to the core sequence-and-SMILES representation.
  • Results: WideDTA achieved CI 0.875 versus DeepDTA CI 0.863 on KIBA, with statistical significance (p-value 0.001).The comparison is reported for the CI metric.
  • Results: Sequence-and-SMILES words alone achieved CI 0.874 and still outperformed character-based DeepDTA significantly (p-value 0.0005).This supports word-based representation as an alternative to character-based modeling for this task.

2.1 Dataset

The study evaluates WideDTA on the Davis and KIBA kinase benchmark datasets, which provide complementary binding-affinity measurements and interaction scales.

  • Datasets: The evaluation uses Davis selectivity-assay data for kinase proteins and KIBA as benchmark datasets.Both datasets are used to assess the proposed model.
  • Davis: Davis contains about 30K interactions, with 69% having pKd=5, indicating weak or no interaction.Davis values are reported as dissociation constants.
  • KIBA: KIBA has about three-times more interactions than Davis, with scores computed from heterogeneous IC50, Ki, and Kd information.The reported KIBA values combine multiple affinity-related measurements.

2.2 Representation Modules

WideDTA converts protein and ligand information into overlapping word-based representations, including sequence-derived words, PROSITE domain and motif words, chemical words, and ligand MCS words.

  • Protein sequence: Protein sequences are collected from UniProt and converted into consecutive overlapping 3-residue words.The representation uses amino-acid subsequences rather than individual sequence characters.
  • Protein domains and motifs: PROSITE motifs and domain profiles are converted into overlapping 3-residue subsequences called PDM.PDM denotes the protein domain and motif information module.
  • Chemical words: Ligand SMILES strings are divided into consecutive overlapping 8-character chemical words.The chosen length followed experiments with k values from 7 to 12 that found no statistically significant performance difference.
  • SMILES processing: DeepSMILES syntax is used for ligand strings and corresponding MCS representations before extracting the ligand features.MCS patterns are converted into DeepSMILES so they can be extracted from the transformed representation.
  • Maximum common substructures: Ligand maximum common substructures are extracted as a separate input module called LMCS.The study uses approximately 100K MCS-derived words to represent chemical-space patterns.

2.3 Representation of text-based modules

WideDTA uses separate convolutional modules to process each text-information source, motivated by overlapping and non-sequential protein and ligand words. Across both datasets, protein sequences produce the most words and MCS representations the fewest.

  • Representation rationale: The word-based representation accommodates protein motifs and domains and ligand MCS words that can overlap or lack sequential ordering.These properties motivated using words instead of character-based representations.
  • Word distributions: For both Davis and KIBA, protein sequences produce the most words, whereas MCS produces the fewest.The distributions count how many proteins or ligands contain corresponding numbers of extracted words.
  • Convolutional modules: Each information source is processed by its own module with two 1D-convolutional layers and max pooling to produce abstract features.The resulting features are then used by the model for prediction.

2.4 Prediction Model: WideDTA

WideDTA is a word-based CNN model that combines textual information from proteins and compounds, using multiple CNN modules whose outputs are integrated for prediction. It was evaluated on benchmark datasets with shared training and test folds.

  • Architecture: WideDTA combines word-based protein and ligand text inputs rather than using DeepDTA’s character-based representation.The model uses a CNN-based architecture similar to DeepDTA, but its inputs are words.
  • Architecture: The architecture combines four CNN blocks, with extracted features concatenated and passed through fully connected layers.Each text module uses two 1D-convolutional layers, max pooling, and ReLU activation; dropout layers are used between fully connected layers.
  • Input representation: WideDTA represents words with 128-dimensional dense vectors before processing the encoded inputs.The dimensionality of the input matrices varies by dataset and information source.
  • Evaluation setup: Performance was evaluated on the Davis and KIBA benchmark datasets using the same training and test folds for all experiments.Each dataset was split into six equal parts, with one part reserved for independent testing and the remainder used for hyperparameter selection by cross-validation.

2.5 Evaluation

The evaluation measures how accurately WideDTA predicts and ranks binding affinities using CI, MSE, and Pearson correlation, with statistical comparisons tested by paired t-tests.

  • Metrics: Concordance Index (CI) evaluates the ranking performance of models that output continuous values.The formulation compares predictions for pairs with larger and smaller observed affinities.
  • Metrics: Mean Squared Error (MSE) measures differences between predicted values and actual binding-affinity values.The compared values are pKd for Davis and KIBA scores for KIBA.
  • Metrics: Pearson correlation measures the relationship between true and predicted binding-affinity values.The calculation uses covariance and the standard deviations of predicted and original values.
  • Statistical testing: Statistical significance was assessed with paired t-tests using a 95% confidence interval.

2.6 Baseline

WideDTA was compared with traditional machine-learning baselines and DeepDTA, a deep-learning model that uses complete protein sequences and compound SMILES.

  • Traditional baselines: WideDTA was compared with KronRLS and SimBoost, which use traditional machine-learning strategies for binding-affinity prediction.KronRLS uses pairwise protein and compound similarity matrices, whereas SimBoost relies on engineered similarity and network-derived features.
  • Deep-learning baseline: WideDTA was also compared with DeepDTA, a character-based deep-learning model using complete protein sequences and compound SMILES.

3 Results

WideDTA combines up to four text-based protein and ligand representations and is evaluated across feature combinations and benchmark comparisons. Results show that word-based protein-sequence and ligand-SMILES inputs outperform DeepDTA on KIBA, while added domain, motif, and MCS features provide dataset-dependent effects.

  • Model and evaluation: WideDTA incorporates protein sequences, protein domains and motifs, ligand SMILES words, and ligand maximum common substructure words.These four information sources define the model's feature modules.
  • Model and evaluation: Feature ablations compare progressively combined information sources on independent-test-set results for Davis and KIBA.The experiments begin with pairs such as protein sequence plus ligand SMILES and add modules individually.
  • Feature combinations: On Davis, protein sequence plus ligand SMILES achieves CI 0.875 and MSE 0.295, while protein domains and motifs plus ligand MCS reaches CI 0.883 and MSE 0.276.The improvement over the sequence-plus-SMILES pair is not statistically significant; the four-module combination gives the best MSE and Pearson values.
  • Feature combinations: On KIBA, protein sequence plus ligand SMILES provides the best CI 0.874 and MSE 0.179 among the tested feature combinations.Adding protein domains and motifs, alone or with ligand MCS, does not significantly affect performance, despite slight improvements in larger combinations.
  • Protein representation: Domain-and-motif-only models perform similarly to full protein-sequence models, while domain and motif inclusion does not significantly improve prediction in either dataset.Most extracted domains and motifs appear in five or fewer ligands in Davis (70%) and KIBA (75%).
  • Feature combinations: Maximum common substructure words slightly improve over SMILES-based 8-character words on Davis but perform significantly worse on KIBA.This result favors ligand SMILES words over ligand MCS words for KIBA representation.
  • Benchmark comparisons: 0.875 CI: WideDTA outperforms DeepDTA's 0.863 on KIBA with statistical significance (p-value 0.001).The protein-sequence plus ligand-SMILES version also reaches 0.874 CI and significantly exceeds DeepDTA (p-value 0.0005).

4 Conclusion

WideDTA combines four textual representations of proteins and ligands for drug-target binding-affinity prediction. Word-based sequence representations outperform DeepDTA in the reported comparisons, while protein domains and motifs alone match complete protein sequences but add no significant benefit to them on the kinase datasets.

  • WideDTA combines protein sequences, domains and motifs, ligand chemical words, and ligand maximum common substructure words.Protein features use 3-residue words; ligand features use 8-character chemical words and MCS words extracted from SMILES.
  • Protein sequence and ligand SMILES text alone achieved comparable and significantly better performance than DeepDTA on the Davis and KIBA datasets.The comparison concerns WideDTA's word representations versus DeepDTA's character-based approach.
  • Adding PROSITE-derived protein domain and motif information to the full sequence did not produce a statistically significant predictive improvement.The study attributes this result to the similarity of kinase structures and motifs in the two benchmark datasets.
  • Protein domains and motifs alone performed as well as the complete protein sequence.These features represent a smaller percentage of the full sequence, while their limited overlap across ligands points to captured protein individuality.
Loading 1902.04166v1…