Data quality as the missing translational layer in computational drug repurposing

Technical Summary: Data Quality as the Missing Translational Layer in Computational Drug Repurposing

Problem Statement
Computational drug repurposing has accelerated the generation of drug-disease hypotheses through diverse data sources, including chemical structures, gene expression signatures, and real-world data. However, a critical translational gap exists between the rapid generation of candidate rankings and the availability of decision-ready evidence. Current workflows often conflate high model scores or large literature volumes with "evidence readiness." This leads to the prioritization of candidates supported by fragmented, noisy, safety-incomplete, or immature evidence. The core problem is not a lack of candidate generation methods, but the absence of a consistent, structured layer to assess the quality, maturity, and translational relevance of the evidence supporting a nomination before it escalates to expert review or validation.

Methodology
The paper proposes a framework of Evidence-Readiness Auditing, positioned as a decision-quality layer between computational nomination and downstream portfolio prioritization. Rather than predicting clinical efficacy, this audit evaluates the interpretability and robustness of the evidence base. The framework assesses six specific dimensions:

1. Terminology Quality: Evaluating the consistency of drug and disease entity mapping (e.g., distinguishing between active ingredients, salts, and brand names; ensuring disease subtypes are correctly indexed).
2. Literature Composition: Decomposing the volume of retrieved literature into therapeutic, adverse, and irrelevant categories to distinguish signal from noise.
3. Safety Burden: Auditing prior liabilities, dose relevance, and exposure profiles to determine if prior human experience is transferable to the new indication.
4. External Support: Identifying the presence of clinical trials, regulatory support, or independent validation.
5. Evidence Maturity: Classifying the state of evidence (e.g., approved use, trial-supported, preclinical, computational) without collapsing these distinct states into a single confidence score.
6. Bayesian Uncertainty: Quantifying the stability of the maturity assessment, treating uncertainty as an explicit part of the decision problem rather than a post-hoc caveat.

To validate this framework, the authors constructed a heterogeneous benchmark of 28 drug-disease pairs. This set included COVID-19 examples, classic repurposing controls, oncology controls, and computer-aided oncology hypotheses. The benchmark was designed not to prove efficacy, but to test the audit's ability to distinguish between different evidential states (e.g., established clinical use vs. sparse computational nomination).

Key Results
The application of the evidence-readiness audit to the benchmark yielded several critical insights:

• Readiness and Uncertainty are Distinct: The study demonstrated that a candidate's evidence-readiness score cannot be interpreted in isolation. A candidate with moderate readiness but narrow uncertainty (stable evidence) is fundamentally different from one with high readiness but wide uncertainty (fragile evidence). The latter often arises from sparse records, mixed therapeutic/adverse signals, or immature translation.
• Literature Volume is a Poor Proxy for Strength: High publication counts often mask the composition of the evidence. A drug-disease pair may have a large literature footprint dominated by adverse events, mechanistic speculation, or irrelevant mentions, which inflates perceived maturity without adding translational confidence. Conversely, sparse candidates may be biologically plausible but require explicit flagging for further evidence generation.
• Audit Exposes "Overcalling": When comparing expected maturity labels against observed audit classifications, the majority of cases (18 of 28) were "partial matches." This indicates that most repurposing hypotheses occupy intermediate states where some evidence exists, but critical questions regarding safety, disease-context fit, or uncertainty remain unresolved. The audit successfully identified these intermediate states, whereas simple ranking would likely obscure them.
• Safety is Context-Dependent: The audit highlighted that prior approval does not guarantee safety in a new indication. Safety liabilities must be audited against the specific proposed dose, route, formulation, and concomitant therapies of the new context.

Significance and Claims
The paper argues that data quality is a translational requirement, not merely a preprocessing detail. The primary significance of this work is the introduction of Evidence-Readiness Auditing as a mechanism to make the evidential basis of drug repurposing transparent, reviewable, and decision-ready.

The authors claim that this framework:

• Shifts the decision object from a ranked candidate to an interpretable evidence profile.
• Prevents the premature escalation of candidates based on "noisy" signals or inflated literature counts.
• Supports structured triage by distinguishing candidates ready for expert review from those requiring safety adjudication, preclinical validation, or continued surveillance.
• Aligns with decision-quality thinking in early drug discovery by making assumptions, uncertainties, and alternatives explicit before resources are committed.

The paper explicitly states that this audit does not infer clinical efficacy, estimate treatment effects, or replace experimental or clinical validation. Instead, it serves as an upstream check to ensure that the available evidence is sufficiently standardized, relevant, and safety-aware to justify the next strategic step in the drug discovery pipeline.