Source-linked AI summary
DeepGS: Deep Representation Learning of Graphs and Sequences for Drug-Target Binding Affinity Prediction
Xuan Lin
TL;DR
DTA prediction is important but existing approaches may require difficult-to-obtain target structures or omit local chemical context. DeepGS jointly learns sequence context and drug topology with distributed embeddings and deep neural networks, and experiments report promising performance and competitiveness against established baselines.
Problem
Existing DTA methods may require expert knowledge or target 3D structures, while deep-learning approaches omit local chemical context in amino-acid and SMILES sequences.
Source: §1 · p. 1Method
DeepGS uses Smi2Vec and Prot2Vec with CNN, BiGRU, and GAT components to learn target context, drug context, and drug topology for affinity prediction.
Source: §2.1 · p. 2, §2.2.1 · p. 2Results
DeepGS reports promising performance and competitiveness, outperforming the evaluated baselines across all metrics on the Davis dataset and performing better on almost all metrics on KIBA.
Source: §3.2 · p. 5Takeaways & Limitations
Jointly modeling local chemical context and molecular topology is reported as effective within the DeepGS framework.
Source: §1 · p. 2KIBA data heterogeneity may negatively affect DeepGS’s CI metric, for which its performance remains comparable to DeepDTA.
Source: §3.2 · p. 5
Abstract
from arXiv · showhide
Accurately predicting drug-target binding affinity (DTA) in silico is a key task in drug discovery. Most of the conventional DTA prediction methods are simulation-based, which rely heavily on domain knowledge or the assumption of having the 3D structure of the targets, which are often difficult to obtain. Meanwhile, traditional machine learning-based methods apply various features and descriptors, and simply depend on the similarities between drug-target pairs. Recently, with the increasing amount of affinity data available and the success of deep representation learning models on various domains, deep learning techniques have been applied to DTA prediction. However, these methods consider either label/one-hot encodings or the topological structure of molecules, without considering the local chemical context of amino acids and SMILES sequences. Motivated by this, we propose a novel end-to-end learning framework, called DeepGS, which uses deep neural networks to extract the local chemical context from amino acids and SMILES sequences, as well as the molecular structure from the drugs. To assist the operations on the symbolic data, we propose to use advanced embedding techniques (i.e., Smi2Vec and Prot2Vec) to encode the amino acids and SMILES sequences to a distributed representation. Meanwhile, we suggest a new molecular structure modeling approach that works well under our framework. We have conducted extensive experiments to compare our proposed method with state-of-the-art models including KronRLS, SimBoost, DeepDTA and DeepCPI. Extensive experimental results demonstrate the superiorities and competitiveness of DeepGS.
1 Introduction
DTA prediction is important for drug discovery, but existing simulation and machine-learning approaches face unavailable structural information or omit local chemical context. DeepGS addresses this gap by jointly learning local context and molecular topology, with experiments showing promising performance.
1 Introduction
- Simulation-based methods depend on expert knowledge or target 3D structures, while traditional machine-learning methods use handcrafted features, descriptors, and pairwise similarities.Source: §1 · p. 1
- DeepGS targets DTA prediction by combining local chemical context from amino-acid and SMILES sequences with drug molecular topology.The paper identifies these complementary information types as important for DTA prediction.Source: §1 · p. 1
- DeepGS uses CNN, BiGRU, and GAT blocks with Smi2Vec and Prot2Vec embeddings to learn representations of targets and drugs.Source: §1 · p. 1
- The framework requires neither expert knowledge nor target 3D structure, distinguishing it from simulation-based approaches.Source: §1 · p. 1
- Experiments on small and large benchmark datasets report promising performance, effective joint modeling of context and topology, and a molecular-structure approach that works within DeepGS.Source: §1 · pp. 1–2
2 The Proposed Method
DeepGS encodes target and drug sequences into distributed representations, models drug topology alongside local chemical context, and predicts affinity from their learned representations. Its architecture combines CNN, BiGRU, and GAT components in an end-to-end regression framework.
2.1 Overview of DeepGS
- DeepGS encodes symbolic tokens, learns sequence and molecular-graph representations, then predicts affinity from the combined encodings.Source: §2.1 · p. 2
2.2.1 Local Chemical Context Modeling
- Smi2Vec and Prot2Vec replace label or one-hot representations with distributed embeddings for drug and target sequence symbols.Source: §2.1 · p. 2, §2.2.1 · p. 2, §2.3 · p. 3
- A BiGRU models local chemical context in fixed-length SMILES matrices, using zero-padding when sequences are shorter than the chosen input length.Source: §2.2.1 · pp. 2–3
2.2.2 Molecular Structure Modeling
- A GAT models drug topology by propagating neighboring-atom information and aggregating r-radius subgraph representations into a molecular vector.Source: §2.2.2 · p. 3
2.3 Representation Learning for Target/Protein
- Protein sequences are partitioned into fixed-length 3-grams, embedded from a pre-trained dictionary, and processed by a CNN to capture local context.Source: §2.3 · p. 3
2.4 Drug-target Binding Affinity Prediction
- The drug and target representations are concatenated and passed through fully connected layers to output a real-valued affinity, trained with MSE.Source: §2.4 · p. 3
3 Experiments
The experiments compare DeepGS with state-of-the-art models and include analyses of performance, sensitivity, and the effectiveness of its main strategies.
3 Experiments
- The evaluation compares DeepGS with state-of-the-art models on drug-target affinity prediction.Source: §3 · p. 3
- Additional experiments analyze prediction performance and model sensitivity.Source: §3 · p. 3
- An ablation study investigates the effectiveness of the paper’s main proposed strategies.Source: §3 · p. 3
3.1 Experimental Setup
DeepGS is evaluated on Davis and KIBA using regression metrics, binary AUPR thresholds, established baselines, and specified embedding and network settings.
3.1.1 Datasets
- 3.1.1 Datasets: Davis contains 68 drugs, 442 targets, and 30,056 interactions, while KIBA contains 2,111 drugs, 229 targets, and 118,254 interactions.Source: §3.1.1 · pp. 3–4
- 3.1.1 Datasets: Random six-way splits use five subsets for training and one for testing, with pKd for Davis and KIBA-values for KIBA as affinity labels.Source: §3.1.1 · p. 4
3.1.2 Evaluation Metrics
- 3.1.2 Evaluation Metrics: Performance is assessed with MSE, CI, r2_m, and AUPR, measuring error, ranking, QSAR external prediction, and binary classification.Source: §3.1.2 · p. 4
3.1.3 Baseline Methods
- 3.1.3 Baseline Methods: DeepGS is compared with KronRLS, SimBoost, DeepCPI, and DeepDTA, using adapted regression settings for DeepCPI.Source: §3.1.3 · p. 4
3.1.4 Implementation Details
- 3.1.4 Implementation Details: Smi2Vec and Prot2Vec use 100-dimensional representations, while the model uses 10-head GAT, BiGRU, CNN, PyTorch, TensorFlow, and ADAM.Source: §3.1.4 · p. 4, Table 1 · p. 4
3.2 Comparison Results
DeepGS outperforms the compared baselines across all reported Davis metrics and performs better on almost all KIBA metrics, with comparable CI to DeepDTA.
3.2 Comparison Results
- The Davis comparison associates DeepGS’s advantage with jointly modeling molecular topology and local chemical context through GAT and advanced embeddings.Source: §3.2 · p. 5
- DeepGS outperforms KronRLS, SimBoost, DeepCPI, and DeepDTA in all reported Davis metrics.Source: §3.2 · p. 5, Table 2 · p. 5
- On KIBA, DeepGS performs better on almost all metrics, although its CI remains comparable to DeepDTA.Source: §3.2 · p. 5, Table 3 · p. 5
- KIBA’s multiple-source composition, integrating Ki, Kd, and IC50, may negatively affect DeepGS’s CI metric.Source: §3.2 · p. 5
3.3 Model Analysis
DeepGS predictions closely track measured affinities on Davis and KIBA, while sequence-length analyses show that truncation mainly matters when lengths fall below typical sequence sizes.
Prediction analysis
- Samples for both Davis and KIBA lie close to the p = m line, supporting good prediction performance on both datasets.Davis predictions concentrate around pKd values of 5–6, while KIBA scores concentrate around 10–14.Source: §3.3 · p. 5, Figure 3 · p. 5
Sequence-length sensitivity
- Lds = 50 degrades performance relative to Lds = 100, whereas the gap between Lds = 100 and Lds = 500 is very small.Most Davis SMILES sequences are longer than 50 but shorter than 100, so truncation at 50 removes information while larger limits usually do not.Source: §3.3 · pp. 5–6, Figure 2 · p. 5
- Lps = 500 performs noticeably worse than Lps = 1000, while increasing Lps from 1000 to 2000 produces a less obvious difference.The average protein-sequence length is 788, making truncation common at 500 but less consequential above 1000.Source: §3.3 · pp. 5–6, Figure 4 · p. 6
- The results suggest that setting sequence length above the dataset’s average can limit performance degradation from truncation.The sensitivity experiment fixes sequence lengths by cutting longer inputs and zero-padding shorter ones.Source: §3.3 · pp. 5–6
3.4 Ablation Study
The ablation study finds that both local chemical context and the proposed GAT-based molecular structure modeling contribute to DeepGS performance.
Local chemical context
- DeepGS1 is inferior to DeepGS on CI, MSE, r_m^2, and AUPR, indicating that removing Smi2Vec and Prot2Vec context representations reduces performance.DeepGS1 removes local chemical context from both drug and protein inputs.Source: §3.4 · p. 6, Table 4 · p. 6
Ablation design
- The variants separately test local chemical context and the molecular structure component under the same experimental settings.DeepGS1 investigates local chemical context, whereas DeepGS2 justifies the GAT choice.Source: Table 4 · p. 6, §3.4 · p. 6
Molecular structure modeling
- DeepGS2 is obviously poorer than DeepGS, supporting the choice of the proposed GAT-based molecular structure modeling approach over the evaluated GNN-based variant.The authors suggest molecular structure modeling contributes substantially and is sensitive to changes when local chemical context is also included.Source: §3.4 · p. 6, Table 4 · p. 6
4 Related Work
Related DTA work spans simulation, traditional machine learning, and deep learning, while DeepGS combines local chemical context with drug topological information.
Classical methods
- Simulation-based and classical machine-learning methods may require expert knowledge, target 3D structures, handcrafted descriptors, or similarity-based representations.KronRLS uses similarity matrices and SimBoost uses gradient boosting to address limitations of earlier approaches.Source: §4 · pp. 6–7
Deep learning methods
- Deep learning methods have modeled drug-target relationships with heterogeneous networks, graph structures, molecular fingerprints, protein sequences, and attention mechanisms.Examples include DBN, NeoDTI, graph convolutional models, two-way attention, multi-view graph auto-encoders, and Decagon.Source: §4 · p. 7
Closest deep-learning baselines
- DeepDTA uses one-hot sequence encodings with CNNs, whereas DeepCPI combines fingerprint-based GNN drug representations with CNN protein encoding.Both methods address drug-target prediction but differ in how molecular and sequence information are represented.Source: §4 · p. 7
DeepGS positioning
- DeepGS differs from DeepDTA and DeepCPI by jointly modeling local chemical context and drug topology using Smi2Vec, Prot2Vec, and a new molecular structure approach.The paper positions this combination as its distinguishing representation-learning design.Source: §4 · p. 7
5 Conclusions
DeepGS is an end-to-end DTA framework that combines molecular topology with local chemical context encoded from drug and protein sequences. Experiments report promising performance against state-of-the-art models, while the authors identify further integration and application directions.
Framework
- DeepGS combines GAT-based molecular-graph topology with BiGRU-derived drug context and Smi2Vec and Prot2Vec encodings for symbolic sequences.The framework predicts drug-target binding affinity end to end.Source: §5 · p. 7
Results
- Extensive comparisons with state-of-the-art models demonstrate promising performance for DeepGS in DTA prediction.The conclusion does not report a specific aggregate metric in this passage.Source: §5 · p. 7
Future directions
- The authors propose future work on integrating other techniques, such as generative adversarial networks, and extending DeepGS to other data-mining and bioinformatics problems.Source: §5 · p. 7