BLUE BELL, PA — Unisys researchers are calling for drug-repurposing AI to be judged on whether its recommendations can withstand biological, safety, and evidentiary scrutiny — not predictive accuracy alone — as regulators increase expectations for transparency in AI-assisted drug development.
The argument appears in a peer-reviewed Perspective published July 30 in Frontiers in Pharmacology by Markus Bertl and Salvatore Sinno, researchers affiliated with Unisys’ Advanced Research and Innovation Group in Blue Bell. The paper proposes a four-part framework for evaluating whether AI-generated drug hypotheses are credible enough to support experimental and translational decisions.
The framework, called MURP, assesses mechanistic coherence, uncertainty, robustness and provenance. The researchers argue those measures should supplement conventional performance metrics when AI systems identify existing drugs as candidates for new uses.
That distinction matters because a model can score highly against historical datasets without establishing that its recommendation is biologically meaningful, clinically actionable or safe, according to the paper. The authors point to potential problems including database incompleteness, label leakage, biased network structures and computational shortcuts.
The researchers propose greater use of biomedical knowledge graphs and neurosymbolic AI, which combines statistical learning with explicit rules or structured knowledge. Under that approach, a neural model could identify a potential drug-disease relationship while a symbolic layer tests explanatory pathways against known mechanisms, tissue-specific expression, contraindications and other pharmacological constraints.
The goal is to produce an inspectable evidence chain connecting a drug with targets, pathways, phenotypes and potential clinical outcomes, rather than supplying researchers with a prediction score alone.
The paper’s focus also reflects changing regulatory expectations. Bertl and Sinno cite joint principles issued by the US Food and Drug Administration and European Medicines Agency in January that emphasize transparency, lifecycle risk management, communication of model limitations and traceable evidence for AI used in drug development.
“The next phase of AI innovation will be shaped not only by performance, but by credibility,” Sinno, Unisys’ vice president of innovation for Enterprise Computing Solutions, said in the company’s release.
The researchers do not argue that explainability can substitute for laboratory or clinical validation. They say statistically strong predictions can remain useful during early-stage exploration, while mechanistic explanations become increasingly important when AI recommendations influence expensive experiments, safety assessments, translational investment or regulatory-facing decisions. Ultimately, pharmacological credibility still depends on empirical evidence.
The paper is also narrower than the company’s description of a “study” may suggest. It is a Perspective article rather than a report of a new experimental test of the proposed framework. The authors explicitly state that they did not develop a new benchmark or executable pipeline, conduct a prospective experiment, or perform a systematic literature review. Instead, they synthesize existing research to propose a standard that future studies would need to test.
Among those unanswered questions is whether evidence-chain evaluation actually improves experimental hit rates, resource allocation, safety assessment, or translational decisions compared with approaches centered primarily on predictive accuracy.
Both authors were employed by Unisys, according to the paper’s conflict-of-interest disclosure. The authors reported receiving no financial support for the work or its publication. They also disclosed using Microsoft Copilot to improve the manuscript’s structure, language, and readability, with the authors reviewing and editing the resulting material.
Support the local news that supports Chester County. MyChesCo delivers reliable, fact-based reporting and essential community resources—free for everyone. If you value that, click here to become a patron today.