Abstract
EnSidNet is a novel hybrid Siamese Neural Network designed to automatically group clinical trials into their respective drug-development pathways. By combining text-based deep learning with structured numerical data, the model achieves nearly 97% accuracy in one-shot evaluation, significantly reducing the manual effort required for pharmaceutical market analysis and success-rate modeling.
Introduction
Tracking a drug from its first human trial to market approval is a complex challenge. A single "pathway" can span over a decade, cost hundreds of millions of dollars, and involve dozens of trials across different countries. Currently, linking these trials automatically is difficult due to messy public databases where companies use inconsistent drug names, skip trial phases, or fail to update records. While proprietary hand-curated datasets exist, they are expensive and labor-intensive. EnSidNet addresses this "unmet healthcare task" by algorithmically reconstructing these pathways to help researchers calculate drug success probabilities and analyze competitive landscapes.
Methodology
The authors propose EnSidNet (Enhanced hybrid Siamese Deep Neural Network), a Siamese architecture consisting of two identical branches that process clinical trial pairs to determine similarity.
Key technical features include:
- Hybrid Architecture: EnSidNet processes character-based text (drug names, conditions, sponsors) using three layers of Bidirectional LSTMs with an Attention mechanism, alongside numerical data (trial phases, dates, country counts).
- Feature Engineering: The model calculates "difference vectors" between trials, such as the gap in days between start dates and the number of common sponsors.
- Enhanced Component: A fully connected layer added before the final prediction significantly boosts performance over simpler hybrid models.
- Iterative Pipeline: Designed for semi-supervised environments, the model identifies likely matches to speed up human expert labeling, which is then fed back into the training set.
Datasets and Evaluation
The researchers utilized several sources for validation:
- DrugBank: Primary source for over 142,000 raw trials.
- FDA Snapshot: Dataset of pivotal trials for FDA-approved drugs.
- Hand-Curated Pathways: High-quality dataset for ground truth validation.
- Balanced & Imbalanced Sets: Testing included a balanced set (50/50 match/no-match) and a realistic 1:32 imbalanced set to mimic real-world conditions.
Results and Discussion
EnSidNet demonstrated superior performance across multiple evaluation metrics. On a balanced validation set, it achieved 91.3% accuracy, significantly outperforming a standard Siamese network (76.3%).
One-Shot Learning Performance
The model excels at "one-shot" learning—identifying if a new, unseen trial belongs to an existing pathway after seeing only a single example. In a 32-way one-shot evaluation, EnSidNet reached 96.97% accuracy.
Robustness on Imbalanced Data
In realistic scenarios with highly imbalanced data (1:32 ratio), the model maintained high precision and recall, achieving an F1-score of 0.90. This confirms the model's ability to distinguish true matches without over-predicting the majority class.
Key Insights
- Metadata Importance: SHAP analysis revealed that the number of common sponsors was the most critical feature for distinguishing pathways.
- Hybrid Advantage: Adding numerical differences and metadata increased 1-shot performance by nearly 30% over pure text models.
- Discovery Potential: The algorithm identified 1.5 times more potential pathways in DrugBank than were in the original training set.
Conclusion
EnSidNet provides a robust, efficient solution for clinical trial pathway reconstruction, processing over 34,000 trials in under four hours. While highly accurate, it functions best in a semi-supervised loop where human experts verify borderline cases. Future work will focus on refining de-novo pathway creation and further automating data cleaning for inconsistent international drug nomenclatures.
Summary updated