Elucidating the semantics-topology trade-off for knowledge inference-based pharmacological discovery

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Pharmacological Discovery

Background:

• Artificial intelligence (AI) offers significant potential for accelerating pharmacological discovery by synthesizing vast amounts of biomedical knowledge.
• Knowledge graphs representing interactions between drugs, diseases, genes, and proteins are crucial for machine learning (ML)-based discovery methods like link prediction.
• Existing predictive models are often biased by network topology, relying on high-degree nodes rather than nuanced biological understanding.

Purpose of the Study:

• To investigate the confounding effect of network topology on the semantic understanding of biological relations within knowledge graphs.
• To evaluate the impact of topological bias on the performance of drug repurposing applications.
• To identify the need for novel knowledge representation and inference methods for effective pharmacological innovation.

Main Methods:

• Development of an experimental pipeline to perform semantic and topological perturbations on biomedical knowledge graphs.
• Assessment of drug repurposing performance degradation when biological semantics are ablated under varying topological conditions.
• Quantification of the increase in performance drop due to topological bias mitigation.

Main Results:

• Ablating meaningful semantics led to a performance drop in drug repurposing.
• This performance drop increased by 21% and 38% in two different networks when topological bias was mitigated.
• Topological biases significantly confound the interpretation of biological relations in knowledge graphs.

Conclusions:

• Current methods for knowledge representation and inference in biomedical knowledge graphs are insufficient for fully utilizing biological semantics.
• Addressing topological biases is critical for improving the accuracy and reliability of AI-driven pharmacological discovery.
• New approaches are necessary to develop robust methods for knowledge representation and inference to advance pharmacological innovation.