TL;DR
This paper introduces REx, a novel reinforcement learning approach that generates scientifically grounded explanations for drug repurposing predictions. By rewarding explanations for both accuracy (fidelity) and meaningfulness (relevance), REx significantly improves predictive performance and provides more insightful justifications than existing methods.
What Problem Does This Paper Address?
The increasing reliance on AI for scientific discovery, particularly in areas like drug repurposing, necessitates not just accurate predictions but also understandable explanations. For AI to be truly accepted as a scientific tool, its outputs must be interpretable and align with existing biomedical knowledge. Current methods for generating explanations, especially those based on link prediction in knowledge graphs (KGs), often fall short. They tend to focus on identifying relevant features or triples but fail to provide the deeper, causal, and contextually rich explanations that scientists require for validation and hypothesis generation. This gap hinders the adoption of AI in critical scientific domains where trust and understanding are paramount.
Methods
REx tackles this problem by framing explanation generation as a reinforcement learning (RL) task. The core innovation lies in its dual reward mechanism that guides the RL agent. This mechanism rewards paths that not only correctly connect the drug and disease entities (fidelity) but also possess high information content (IC), a measure of the specificity and meaningfulness of the entities and relations within the path.
To compute IC, the authors introduce Clustered IC, which refines traditional IC calculations by considering node degrees within clustered subgraphs. This helps mitigate biases arising from varying granularities and research focus within the KG. Furthermore, REx incorporates an early stopping mechanism for the RL agent, promoting simpler and more concise explanations without sacrificing relevance. Finally, the generated explanatory paths are enriched with domain-specific ontologies (like NCIT and ChEBI) to provide a richer, more coherent, and scientifically grounded context.
Datasets
The experiments were conducted on three established biomedical knowledge graph benchmarks:
- Hetionet: An integrative KG combining data from 29 sources, covering genes, compounds, and diseases.
- PrimeKG: A precision medicine-focused KG spanning multiple biological scales.
- OREGANO: A KG specifically designed for drug repurposing, integrating drug-disease associations.
The authors also mention enriching these KGs by aligning them to the National Cancer Institute Thesaurus (NCIT) and the Chemical Entities of Biological Interest (ChEBI) ontologies.
Key Insights and Ideas
- The power of dual rewards: Simply rewarding accuracy isn't enough for scientific explanations. Incorporating a measure of "meaningfulness" (like IC) is crucial for generating explanations that are both correct and insightful.
- Information Content as a proxy for relevance: The paper posits that paths involving less frequent entities are more likely to reveal meaningful relationships, effectively guiding the agent towards more scientifically relevant explanations.
- Clustering for better IC: Refining IC calculations by clustering entities based on semantic similarity and relation type helps to overcome biases in KGs and provides a more accurate measure of an entity's significance.
- Simplicity matters: The early stopping mechanism in the RL agent demonstrates that concise explanations are preferred and contribute to better performance, highlighting the trade-off between completeness and understandability.
- Ontologies are key for grounding: Integrating external ontologies significantly enhances the scientific grounding and coherence of the generated explanations.
Key Results
REx demonstrated superior performance across the three benchmark datasets (Hetionet, PrimeKG, and OREGANO) in terms of predictive accuracy, achieving the highest Mean Reciprocal Rank (MRR) in all cases. For instance, on Hetionet, REx achieved an MRR of 0.427, outperforming state-of-the-art methods like MINERVA (0.402) and PoLo (0.402).
Qualitatively, expert evaluations showed that REx-generated explanations were rated higher than those from MINERVA. Furthermore, an ablation study revealed that both the relevance reward and the early stopping mechanism significantly contributed to REx's performance, with their removal leading to noticeable drops in MRR. The analysis of IC distributions also showed that REx generated paths with higher IC compared to baselines, indicating more relevant explanations.
Limitations and Future Work
The authors acknowledge that the predictive performance of both REx and other state-of-the-art methods remains "fairly modest," suggesting that in many cases, no explanation can be found due to KG incompleteness or a lack of scientific evidence.
Future work could involve extending REx to other domains beyond drug repurposing and exploring its application to hypotheses derived from alternative data sources. The authors also note that REx's methodology can be adapted to integrate with a wide range of RL frameworks, allowing it to benefit from future advancements in the field.
Summary updated