SPVec

SPVec encodes SMILES strings and protein sequences into continuous, lower-dimensional vectors to represent compounds and proteins for drug-target interaction prediction.

Key Features:

• Word2vec-inspired representation: Uses unsupervised representation learning inspired by Word2vec to derive embeddings from biological sequences and chemical strings.
• Input modalities: Directly processes SMILES strings for chemical compounds and amino acid sequences for proteins.
• Continuous lower-dimensional vectors: Produces continuous, information-rich, lower-dimensional vector representations of compounds and proteins.
• Sparsity and collision mitigation: Reduces sparsity and bit-collision issues associated with manually designed features such as fingerprints.
• Substructure and sequence encoding: Encodes compound substructures and protein sequence patterns implicitly within the embedding space.
• Embedding visualization: Applies visualization techniques to show that similar compounds or proteins occupy analogous regions in vector space.
• Comparative benchmarking: Evaluated against manually designed features including MACCS fingerprints and amino acid composition (AAC).
• Classifier evaluation: Assessed using Gradient Boosting Decision Trees, Random Forests, and Deep Neural Networks.
• Datasets and validation: Trained and tested on BindingDB with independent test sets from DrugBank, and used DrugBank for large-scale predictions.
• Prediction validation: Generated unlabeled DrugBank DTI predictions, with two of the top five predicted novel DTIs corroborated by external evidence.

Scientific Applications:

• Drug-target interaction prediction: Provides vector features for predicting DTIs between compounds and proteins.
• Feature representation for machine learning: Supplies continuous embeddings as alternatives to MACCS fingerprints and AAC for ML models.
• Drug discovery and reprofiling: Enables large-scale prediction of unlabeled DTIs to support drug repurposing and candidate discovery.
• Benchmarking and robustness assessment: Facilitates comparative evaluation of predictive models across BindingDB and DrugBank independent test sets.
• Similarity analysis: Uses embedding visualization to analyze biochemical and biophysical similarities among compounds and proteins.

Methodology:

Employs an unsupervised, Word2vec-inspired representation learning approach to transform SMILES strings and protein sequences into continuous lower-dimensional vectors, visualizes embedding space to assess similarity, evaluates predictive performance by training Gradient Boosting Decision Trees, Random Forests, and Deep Neural Networks versus MACCS fingerprints and AAC on BindingDB with independent DrugBank test sets, and generates predictions for unlabeled DrugBank DTIs.