Why disciplined narrative - not algorithms alone - will determine whether AI-led repurposing creates durable value
A note before we begin: This article offers a governance lens for leaders navigating AI-generated signals - not scientific claims about AI's efficacy in drug discovery. The field is young, the evidence is evolving, and the goal here is to ask better questions, not provide definitive answers. We acknowledge that for every signal AI generates, many more lead nowhere. Survivorship bias is real, and we name it deliberately.
Finding new clinical indications - whether for approved drugs or pipeline candidates - has long been a strategic aspiration in pharmaceutical R&D. The promise is compelling: extract additional value from existing molecules, shorten development timelines, and reduce risk by leveraging known safety profiles.
In recent years, artificial intelligence (AI) has been positioned as a powerful accelerator of this ambition. By integrating biological data, clinical evidence, literature, and real-world datasets, AI systems can surface non-obvious drug-disease relationships at a scale and speed that were previously impractical.
Yet despite growing investment and activity, AI-led indication finding has produced far fewer regulatorily validated outcomes than its visibility might suggest. Understanding why requires separating signals from substance - and applying a disciplined lens to survivorship bias.
This article uses that triad - Signals, Substance, Survivorship Bias - not to critique past efforts, but to help leaders ask better forward-looking questions about how AI is deployed, governed, and defended in high-stakes decisions. As Dilip Shanghvi of Sun Pharma recently observed, "When our competitors are using AI tools to make their drug development faster, we must know how to come up with a new idea or a structure. There is no option for us but to use AI in drug discovery and research" (Economic Times, 2026). At the same forum, Glenn Saldanha of Glenmark Pharma offered a complementary perspective, noting that while AI's near-term impact will be strongest in "the small molecule drug discovery space," it will ultimately play a role across "the entire research cycle — from clinical trials to molecular modelling or drug design" (Economic Times, 2026). These dual perspectives — competitive necessity from Sun Pharma, and practical scope from Glenmark — frame the challenge this article addresses.
What AI Reliably Delivers Today: Strong Signals, Faster
Across industry and academia, AI has demonstrated emerging strengths in indication discovery:
- Integrating heterogeneous datasets (omics, pathways, phenotypes, literature)
- Knowledge-graph and network-based inference across biological systems
- Prioritizing hypotheses in data-sparse domains (notably rare diseases)
- Compressing exploratory timelines dramatically
In short, AI excels at hypothesis generation and prioritization - but this is where its proven contribution largely begins and ends.
What it does not yet do reliably is convert those hypotheses into defensible clinical development decisions without substantial human interpretation, validation planning, and governance discipline. For every AI-generated signal that advances, many more quietly fade - a reality that survivorship bias often hides.
In plain terms: AI can draw knowledge networks fast and suggest a set of indications. But after that, it is up to human experimental validation, judgment, and intelligence to decide where to spend money. The work that follows AI-driven discovery typically requires investment of time and capital many orders of magnitude larger.
The AI Repurposing Value Chain (Where Rigor Matters)
Below is a simplified representation of where value is created - and where defensibility must be established.
At each transition, leaders must ask not "Is the signal interesting?" but "Is the narrative sufficiently robust to justify the next commitment of time, capital, and credibility?"
Structural Challenges That Shape Outcomes (Industry-Wide, Not Company-Specific)
A word on nuance: The challenges below are real, but they play out differently depending on therapeutic area, data availability, and organizational context. We surface them not as settled conclusions, but as areas where leaders should proceed with caution and curiosity.
The Data Foundation Challenge
AI's performance is fundamentally gated by the quality, completeness, and interoperability of data ecosystems. Fragmented clinical data, inconsistent phenotyping, and limited longitudinal outcomes often constrain what AI can reasonably infer - regardless of model sophistication. Even the most elegant algorithms cannot compensate for absent or biased data.
The Regulatory Landscape
Regulators increasingly recognize AI's role, but standards for validating AI-derived repurposing claims remain emergent. Initiatives such as the FDA's AI/ML Action Plan signal direction, not settled doctrine. This uncertainty places greater responsibility on sponsors to document assumptions, evidence pathways, and decision logic - and to avoid overinterpreting preliminary signals.
The Organizational and Cultural Bridge
AI outputs are probabilistic; clinical development decisions are necessarily deterministic. Bridging this gap requires translation, not escalation - ensuring that AI-generated signals are framed in terms clinicians, regulators, and governance bodies can interrogate and defend. This is as much about communication as it is about technical rigor.
Documented AI-Assisted Repurposing Initiatives (With Evidence Context)
The following examples are often cited in discussions of AI-led repurposing. We include them with important caveats: they illustrate where AI can contribute, not where it has proven clinical success. Most remain at exploratory stages, and none should be read as validation of AI's predictive power.
Baricitinib for COVID-19 (BenevolentAI)
AI contribution: Knowledge-graph and machine-learning analysis identified baricitinib (approved for rheumatoid arthritis) as a candidate with both anti-inflammatory effects and potential inhibition of viral entry pathways.
Outcome: Rapid clinical testing during the COVID-19 pandemic led to Emergency Use Authorization.
Critical context: This remains an exceptional, emergency-driven case, enabled by unprecedented regulatory flexibility rather than routine repurposing pathways. It should not be read as a template for normal circumstances.
Primary reference: The Lancet Digital Health (2020) - AI-enabled identification of baricitinib for COVID-19.
Lifitegrast for Endometriosis (Insilico Medicine)
AI contribution: Multi-omics and pathway-level analysis suggested a potential mechanistic link between Lifitegrast (approved for dry-eye disease) and inflammatory pathways relevant to endometriosis.
Current status: Exploratory and preclinical hypothesis generation; no clinical validation to date.
Why it matters: Illustrates AI's strength in cross-domain hypothesis generation - not proof of translational success. It remains an unvalidated signal.
Evidence anchor: Peer-reviewed and industry analyses summarized in AI-repurposing literature; no completed clinical trial record currently available.
Network-Based AI Repurposing for Alzheimer's Disease
AI contribution: Graph neural networks integrating drug-target interactions, gene expression, and pathway data to prioritize repurposable drugs and combinations.
Current status: Research-stage and preclinical prioritization. No clinical data yet available.
Value: Demonstrates biologically grounded signal generation without overstating clinical readiness - a useful reminder that computational promise and clinical reality remain far apart.
Primary reference: Bioinformatics (2019) — Network-based deep learning framework for drug repositioning.
AI-Assisted Repurposing in Rare Diseases
Recent peer-reviewed work highlights AI's utility in rare disease contexts, where traditional discovery is constrained by limited data and small patient populations. Examples include prioritization efforts in ultra-rare neurological and oncologic conditions using semantic enrichment and systems-biology modeling. However, these remain largely exploratory.
Evidence base: Frontiers in Medicine (2024) - AI in drug repurposing for rare diseases.
AI in Repurposing: A Spectrum of Evidence Maturity
| Case Example | Primary AI Contribution | Current Evidence Stage | Key Defensibility Question |
|---|---|---|---|
| Baricitinib (COVID-19) | Knowledge-graph inference | Clinical validation (emergency context) | Would this pathway hold outside crisis conditions? |
| Lifitegrast (Endometriosis) | Cross-domain pathway discovery | Preclinical / exploratory | What evidence justifies advancement to clinic? |
| DeepDrug (Alzheimer's) | Network-based prioritization | Research / preclinical | How will biological plausibility be validated? |
| Rare disease workflows | Data integration & hypothesis ranking | Research-stage | How will sparse data risk be managed? |
The table is descriptive - not judgmental - and reflects where the field realistically stands. Most entries remain far from clinical validation.
What the Academic Literature Reinforces
Authoritative reviews converge on a consistent message:
- AI is increasingly embedded in early discovery and repurposing workflows.
- The largest gap lies between signal discovery and clinical translation.
- Validation, governance, and documentation - not model accuracy alone - determine impact.
- The field remains young, and claims of success should be met with skepticism until validated in humans.
Key references include:
- Nature Reviews Drug Discovery (2020): Artificial intelligence in drug discovery and development
- Nature Biotechnology (2022): Best practices for AI validation and translational rigor
Together, these underscore that AI's bottleneck is no longer computational - it is decisional.
Reframing Survivorship Bias as a Leadership Discipline
Survivorship bias is often discussed retrospectively - focusing on the few visible successes while ignoring the many quiet dead ends.
For leaders, its real value is prospective.
Understanding survivorship bias equips decision-makers to ask better questions before outcomes are known:
- What evidence threshold justifies moving forward?
- What assumptions are we making explicit?
- What would cause us to stop?
- Can we defend this decision two years from now, regardless of outcome?
This is not about avoiding risk.
It is about making risk explainable.
Where Narrative Defensibility Becomes Decisive
At this point, the question is no longer whether AI should be used in indication discovery - but how its outputs are translated into decisions that can withstand scrutiny.
This is where Narrative Defensibility matters.
Applying the six stress-tests of the EYQA Narrative Defensibility Framework - Evidence, Logic, Perspective, Risk Awareness, Contextual Fit, Actionability - helps leaders convert AI insights from compelling slides into defensible decision assets.
For readers interested in applying this lens:
- How Serious Leaders Stress-Test Narrative Defensibility
- EYQA Narrative Defensibility Stress Tests (role-based toolkit)
These are designed as professional self-assessment tools, not post-mortems.
Closing Perspective
AI has meaningfully expanded what is possible in finding new clinical indications. It has not eliminated the need for disciplined judgment, validation strategy, or narrative clarity.
The organizations that will extract durable value from AI-led repurposing will be those that pair advanced analytics with defensible decision narratives - capable of standing up to scientific, regulatory, and strategic scrutiny alike.
That capability, more than any individual model, will define leadership in the next phase of AI-enabled R&D. But we offer this perspective with humility: the field is young, the evidence is incomplete, and the most important questions remain unanswered.
Post a Comment