ReSimNet

ReSimNet: Drug Response Similarity Prediction via Siamese Neural Networks

ReSimNet predicts transcriptional drug response similarity by learning compound embeddings in transcriptional response space and estimating Connectivity Map (CMap) scores between compound pairs using a Siamese neural network architecture.

Key Features:

• Siamese Neural Network Architecture: Learns embedding vectors for chemical compounds to model transcriptional response similarity and predict CMap scores.
• Embedding Vector Optimization: Trains the network to minimize the difference between cosine similarity of compound embedding vectors and corresponding CMap scores.
• Quantitative Performance Evaluation: Achieves a Pearson correlation coefficient of 0.518 and precision@1% of 0.989 in ensemble models compared to baseline machine learning methods.
• ZINC15 Database Validation: Evaluates predicted similarities against compounds from the ZINC15 database to identify candidates related to prototype drugs with known mechanisms of action.

Scientific Applications:

• Drug Discovery: Identifies structurally distinct compounds with similar transcriptional responses to support novel therapeutic discovery and mechanism-of-action analysis.

Methodology:

Pairs of chemical compounds are processed through a Siamese neural network to generate embedding vectors in transcriptional response space. The model is trained by minimizing the discrepancy between cosine similarity of embeddings and experimentally derived CMap scores. Performance is assessed using Pearson correlation and precision@1%, with qualitative validation performed using compounds from the ZINC15 database.