Introduction
Part I of this article proposed a framework to help guide the discussion around repurposing, which we define very broadly. Now, in Part II, we extend the framework and highlight just a few of the AI platforms and tools which enable the modern reevaluation of repurposing and the value it can quickly create.
The 2×2 Framework
In Part I, we proposed this framework. We repeat it here, redefining the entries to guide this AI-focused discussion.
| Quadrant | Primary AI Applications | Impact Level |
| Q1: Old Indication / Old Formulation | Asset characterization; safety signal mining; opportunity identification | Foundation-building |
| Q2: Old Indication / New Formulation | Formulation prediction; excipient compatibility; stability modeling | High impact on efficiency |
| Q3: New Indication / Old Formulation | Target prediction; indication discovery; EHR mining; literature analysis | Transformative |
| Q4: New Indication / New Formulation | Integrated indication + formulation optimization; clinical trial simulation | Emerging |
Quadrant 1: Building the Foundation with AI
Even at the baseline stage, AI provides value by comprehensively characterizing drug assets:
Mechanism Mapping: AI tools analyze available data to map a drug’s complete mechanism of action, including off-target effects that may not have been appreciated during original development. These off-target effects often become the seeds of Quadrant 3 opportunities.
Safety Signal Mining: Natural language processing (NLP) algorithms scan FDA Adverse Event Reporting System (FAERS) data, published literature, and electronic health records to identify unexpected effects—both adverse and beneficial. A drug that appears to reduce the incidence of an unrelated disease in treated patients becomes a repurposing candidate.
Asset Scoring: Machine learning models can score shelved compounds based on their repurposing potential, considering factors like mechanism breadth, safety profile quality, and manufacturing complexity. Computational approaches can mine and analyze real world data to identify new indications for established MOAs.
Quadrant 2: AI-Driven Formulation Design
Reformulation has traditionally been a trial-and-error process, with formulation scientists iterating through dozens or hundreds of experiments to optimize drug delivery. AI is dramatically compressing this process:
Solubility and Bioavailability Prediction: Approximately 70% or more of drug candidates have poor aqueous solubility, limiting oral bioavailability. Machine learning models, including random forests, gradient boosting, and deep neural networks, can predict which co-formers, salt forms, or excipients will improve solubility, reducing experimental iterations by 3-fold or more.
Excipient Compatibility: AI can predict drug-excipient interactions that might cause stability problems or alter release profiles, allowing formulators to avoid problematic combinations before bench work begins.
Dissolution and Release Modeling: For extended-release or complex delivery systems, AI models can predict drug release profiles based on formulation parameters, enabling in silico optimization before expensive in vitro and in vivo studies.
Process Optimization: Machine learning optimizes manufacturing processes (spray drying, hot melt extrusion, nanoparticle synthesis) to achieve target product attributes.
Examples of companies offering AI-formulation tools as a service include Persist AI, DeepCeutix, and BioVIA.
Quadrant 3: AI as the Engine of Indication Discovery
Quadrant 3 is where AI has achieved the most dramatic results. The sheer scale of biomedical data exceeds human cognitive capacity. AI excels at finding patterns in this complexity.
Network-Based Deep Learning: Algorithms like deepDTnet embed drugs, targets, and diseases into a shared mathematical space, enabling prediction of novel drug-disease relationships. These models have been validated against known repurposing successes and can prioritize candidates for experimental validation.
Real-World Evidence Mining: AI systems analyze electronic health records (EHRs) from millions of patients to identify drugs associated with better outcomes in diseases they weren’t prescribed for. If patients taking Drug A for Disease X have unexpectedly low rates of Disease Y, that’s a repurposing signal worth investigating.
Literature Mining at Scale: NLP models scan the entire PubMed corpus (35+ million articles) plus patents, clinical trial registries, and preprints to extract drug-target-disease relationships. Connections buried across hundreds of papers become visible.
Multi-Omics Integration: By combining genomics, transcriptomics, proteomics, and metabolomics data, AI can match a drug’s mechanism to the molecular signature of a disease, identifying candidates that would never emerge from traditional hypothesis-driven research.
Virtual Screening: Computational methods screen libraries of approved drugs against new targets at scales impossible for wet-lab experiments. A campaign that would take years in the laboratory can be completed in days in silico.
Quadrant 4: Integrated Optimization
Quadrant 4 represents the frontier for AI in repurposing. Here, AI must simultaneously solve two problems: identifying a new indication AND determining the optimal formulation for that indication.
Coupled Indication-Formulation Reasoning: Emerging AI systems can reason why a drug might work in a new indication AND what delivery modifications would be required. For example, an AI might identify that:
- Drug X has off-target activity relevant to Disease Y
- Disease Y is localized to a specific tissue
- Systemic delivery of Drug X causes unacceptable side effects
- Therefore, a topical or local delivery formulation is required
PK/PD Modeling: Physiologically-based pharmacokinetic (PBPK) models, increasingly powered by machine learning, can predict how a reformulated drug will distribute to target tissues. This is critical when the new indication requires drug levels at a different site than the original indication.
Clinical Trial Simulation: AI-powered “virtual patients” can simulate clinical trial outcomes for novel indication/formulation combinations, identifying likely responders and informing trial design before a single real patient is enrolled.
To our knowledge, no system fully integrates all these capabilities today. But it is quite clear that AI will increasingly enable end-to-end Quadrant 4 optimization.
Summary
| Method | Description | Primary Application |
| Network-based deep learning | Embeds drugs, targets, diseases in shared space; predicts novel links | Q3: Indication discovery |
| Natural language processing | Extracts relationships from text (literature, EHRs, patents) | All quadrants |
| Graph neural networks | Models molecular structures and interaction networks | Q3: Target prediction |
| Machine learning regression/classification | Predicts solubility, stability, bioavailability | Q2: Formulation design |
| Reinforcement learning | Optimizes multi-parameter objectives | Q2/Q4: Process optimization |
| Generative models | Proposes novel formulation compositions | Q2/Q4: Formulation design |
| PBPK modeling (ML-enhanced) | Predicts drug distribution in body | Q4: PK in new indication/formulation |
| Clinical trial simulation | Models virtual patient outcomes | All quadrants |
We are in the earliest phases of what we believe is a true revolution in how pharmaceutical products are conceptualized and developed. Digitization of drug development is quickly moving well beyond computational chemistry (a major achievement in and of itself) and enabling rapid innovation across the entire drug development value chain.
Some final thoughts:
- Every drug development program should be treated as a potential platform for future repurposing in some way.
- Repurposing and reformulation strategies should be developed very early in product development, and long before COM patent expiry.
- Our industry already does a good job when it comes to expanding candidates into new indications. This will only accelerate, and become an increasingly important gating factor when it comes to both portfolio and in-licensing strategies.
- Drug Discovery, Clinical, and Formulation teams need to be in the same room, especially when considering Quadrant 4 opportunities.
- Some indications and/or modalities will lend themselves to a repurposing mindset. Oncology and Autoimmune diseases come to mind.
One final note. The 505(b)(2) route, while seemingly attractive, is not for everyone. Companies (and their investors) seeking a biobucks mega-license will find it challenging when this is the development pathway.
However, for the right opportunities, taking these all the way to market may deliver excellent ROIC for shrewd investors and investees who can leverage these AI tools and have a truly rapid, low-cost development plan.