Abstract
Traditional drug discovery is a high-risk, high-cost endeavor, often requiring over $2 billion and 15 years per successful compound. Drug repurposing—identifying new therapeutic uses for existing drugs—presents a more efficient pathway. However, computational repurposing methods typically rely on either isolated chemical structural data or textual mining of medical literature. This paper proposes a novel hybrid embedding approach that integrates IUPAC-based chemical functional groups into a Word2Vec natural language processing framework. By representing drugs as collections of functional units within clinical contexts, the model achieves superior performance in predicting clinical success (AUC 0.84) and successfully identifies novel repurposing candidates validated by real-world evidence.
Introduction
The pharmaceutical industry faces a staggering 80% failure rate in human trials. Drug repurposing offers a strategic alternative, leveraging compounds with established safety profiles. While modern computational methods have moved beyond serendipitous discovery, they often suffer from a "silo" effect: chemical models focus on molecular binding but lack clinical context, while textual models mine PubMed for associations but cannot generalize to novel molecules. This research bridges this gap by creating a unified representation that captures both the chemical "ground truth" and the clinical application of a drug.
Methodology
The researchers evaluated four distinct embedding strategies to represent drugs as mathematical vectors:
- Text-based (Word2Vec): Trained on 5.8 million sentences from PubMed abstracts.
- Mol2Vec: A chemical approach treating molecular fingerprints as "words."
- IUPAC One-Hot: A representation using large functional groups (e.g., hydroxyl, phenyl) derived from formal IUPAC names.
- Hybrid (IUPAC word2vec): The core innovation. This model replaces drug names in sentences with their constituent IUPAC functional groups. For example, "Aspirin treats pain" is translated to "Acetoxybenzoic acid treats pain."
The models were trained using SemMedDB (PubMed literature) and benchmarked against RepoDB, which contains over 10,000 drug-indication pairs. Final validation was conducted using Electronic Health Records (EHR) from 1.5 million patients.
Key Insights and Results
A primary finding of the study is that the granularity of chemical representation matters significantly. Larger functional groups (IUPAC) proved more effective than tiny molecular fragments (Mol2Vec), suggesting that these groups represent the actual "functional units" of biological interaction.
Performance Metrics
The models were tested on two primary tasks:
- Predicting Study Interest: The IUPAC One-Hot model performed best (AUC 0.75), indicating that researcher interest is often driven by familiar chemical structures.
- Predicting Clinical Success: The Hybrid IUPAC word2vec model achieved the highest accuracy (AUC up to 0.84). By combining chemical structure with clinical context, it outperformed pure text or pure chemical models.
Real-World Validation
The hybrid system identified Doxazosine, an alpha-1 blocker typically used for prostate issues, as a candidate for treating Type 2 Diabetes. Analysis of the 1.5 million patient EHR dataset confirmed this prediction: patients on alpha-1 blockers showed a significantly higher success rate in blood sugar management (61% vs 53%, p < 0.0004).
Discussion and Future Work
While successful, the current approach relies on 1D/2D representations (SMILES/IUPAC), which omit 3D spatial information vital for protein-ligand docking. Additionally, text-based models are inherently limited by existing literature biases. Future research aims to integrate EHR data directly into the training phase to capture more nuanced clinical signals and further refine the predictive power of hybrid embeddings.
Summary updated