TL;DR
This paper introduces Affinity2Vec, a novel graph-based method for predicting drug-target binding affinity without relying on 3D structural data. By integrating drug and target similarities with known binding data into a heterogeneous graph, Affinity2Vec leverages representation learning and graph mining to achieve state-of-the-art performance in predicting binding strength, crucial for drug discovery and repositioning.
What Problem Does This Paper Address?
Predicting how strongly a drug will bind to a target protein (drug-target binding affinity, DTBA) is a critical step in drug discovery and repositioning. While many methods exist, a significant challenge is the requirement for the 3D structures of target proteins, which are often unavailable. Existing non-structure-based methods typically predict a binary outcome (whether a drug binds or not) rather than the actual binding strength. This paper addresses the need for a robust, non-structure-based method that can accurately predict continuous binding affinity values, overcoming the limitations of missing 3D structural data and the lack of nuanced binding strength information.
Methods
Affinity2Vec employs a multi-stage approach:
- Input Representation and Graph Construction: The method starts by representing drugs using SMILES strings and proteins using amino acid sequences. It then constructs a weighted heterogeneous graph by integrating three types of information:
- Drug-Target Binding Affinity (DTBA) graph: Known binding affinity values.
- Drug-Drug Similarity (DDsim) graph: Similarities between drugs, derived from either 2D chemical structures or SMILES embeddings.
-
Target-Target Similarity (TTsim) graph: Similarities between target proteins, derived from sequence alignment or sequence embeddings. Crucially, similarity edges are filtered to remove weak connections, and binding affinity edges are transformed to ensure consistent meaning (higher values indicate stronger relationships).
-
Embedding Techniques:
- Drug Encoders: A sequence-to-sequence (seq2seq) model is used to generate drug embeddings from SMILES strings. This unsupervised deep learning approach captures critical drug features.
- Protein Encoders: The ProtVec method, a pre-trained model for protein sequence embeddings, is used to generate feature representations for target proteins.
- Feature Extraction: Three types of features are extracted for drug-target pairs:
- Meta-path Score Features: Graph mining techniques are applied to the constructed heterogeneous graphs (G1 and G2, using different similarity sources) to compute "meta-path scores." These scores represent the probability of traversing specific paths (e.g., Drug-Drug-Target) within the graph, capturing complex relationships.
- Embedding-Based Features: The drug and protein embeddings generated in the previous step are concatenated to form a feature vector for each drug-target pair.
- Hybrid Features: A combination of meta-path score features and embedding-based features.
- Machine Learning Regression: The extracted features are fed into a supervised machine learning regression model. The authors primarily use Extreme Gradient Boosting (XGBoost) due to its speed, parallelizability, and flexibility, and also explore a feedforward Artificial Neural Network (ANN).
Datasets
The study utilizes three benchmark datasets for training and evaluation:
- Davis dataset: Contains selectivity assay data for kinase proteins, with binding affinity measured by kinase dissociation constant (Ka).
- KIBA dataset: Provides Kinase Inhibitor BioActivity (KIBA) scores, which combine Ka, K₁, and IC50 into a single bioactivity score.
- PDBBind Refined v.2015 dataset: A widely used dataset for protein-ligand complexes, providing experimentally measured binding affinity (pKd) values. While this dataset includes 3D structural information, the authors intentionally excluded it to demonstrate their method's non-structure-based capabilities.
Key Insights and Ideas
- Graph-based DTBA is Powerful: Formulating DTBA prediction as a graph-based problem, rather than just using individual drug/protein features, significantly enhances performance. This leverages the "guilt-by-association" principle where similar drugs interact with similar targets.
- Post-processing Embeddings Matters: Simply concatenating raw embeddings from DL models is less effective than post-processing them (e.g., calculating cosine similarity) and integrating these refined features into the graph structure. This adds more meaning and reliability to the features.
- Similarity Source is Crucial: The choice of similarity metric for drugs and targets (e.g., 2D chemical structure vs. SMILES embeddings, sequence alignment vs. sequence embeddings) significantly impacts the performance of graph-based features, suggesting careful selection is needed.
- Meta-path Features Dominate: The meta-path score features, derived from graph mining, appear to be the most influential, often outperforming models that rely solely on concatenated embeddings, even with a smaller feature set.
Key Results
- Superior Performance: Affinity2Vec consistently outperformed state-of-the-art non-structure-based methods across all three benchmark datasets (Davis, KIBA, and PDBBind Refined) on key evaluation metrics like MSE, CI, rm², and AUPR.
- KIBA Dataset Advantage: The method showed particularly strong performance on the larger KIBA dataset, suggesting its effectiveness with large-scale biomedical data.
- Competitive on Smaller Datasets: On the smaller Davis dataset, Affinity2Vec achieved competitive results, often matching or exceeding other methods.
- Robustness: Y-Randomization tests confirmed the statistical significance of the model's predictions, demonstrating that the performance was not due to chance.
- Specific Metric Improvements: For example, on the Davis dataset, the Affinity2Vec_Hybrid model achieved the lowest MSE (0.24) and highest rm² (0.693). On the KIBA dataset, Affinity2Vec_Pscore achieved the lowest MSE (0.112) and highest rm² (0.783).
Limitations and Future Work
- Limited to Known Pairs: Affinity2Vec is currently limited to predicting interactions for known drugs and targets and cannot identify interactions for entirely new drugs or targets.
- Drug Representation: The method uses SMILES strings for drug representation, while molecule graph representations might offer more informative features.
- Future Directions:
- Explore alternative embedding techniques (e.g., graph embedding, knowledge graph embedding).
- Enhance model interpretability to identify critical features.
- Apply Affinity2Vec to real-world drug repurposing case studies.
- Extend the network-based approach to other biomedical domains like drug-disease interactions and protein-protein interactions.
- Investigate methods for predicting interactions of novel drugs and targets.
Summary updated