Artificial Intelligence Will Boost Repurposing of Registered Compounds for Respiratory and Other Indications

Despite substantial investment, only a fraction of preclinical leads become licensed medicines, and many approved therapies provide modest benefit at high cost. The translational ‘valley of death’ is well recognised and reflects biological complexity as well as operational hurdles [1]. Drug repurposing partly addresses these barriers by redirecting compounds with established human safety data to mechanistically plausible, newly identified indications. Compared to de novo discovery, repurposing can reduce cost, compress timelines, and improve the chances of success. Respiratory medicine is an apt domain for repurposing. For example, the centrality, but also toxicity, of glucocorticosteroids (GCS) treatments should encourage research into steroid-sparing alternatives and compounds that can attenuate inflammation and reduce susceptibility to infection. AI tools now offer a way to integrate disparate data and generate testable, mechanism-based repurposing hypotheses. This Commentary outlines the traditional pathways for repurposing and contrasts these with contemporary AI-enabled methodologies, with a focus on respiratory medicine. Classical repurposing proceeds via two main routes. First, mechanistic extension: a drug's known pathway effects are linked to pathogenesis in a second disease. Aspirin is emblematic. Initially used for pain and fever, low-dose regimens that inhibit platelet cyclo-oxygenase-1 were shown to reduce atherosclerotic events by attenuating thromboxane-dependent platelet activation at plaque disruption sites [2]. This translational arc entails mechanistic insight, proof-of-concept, and clinical adoption, capturing the essence of repurposing. Second, therapeutic serendipity: adverse or off-target effects are reframed as primary therapeutic actions in other diseases. While many examples exist across other areas of medicine, respiratory research has often focused on mechanistic extension to identify possible candidates for repurposing because of complex biology, particularly in diseases linked to infection and immunology. Pirfenidone, an oral small molecule with antifibrotic and anti-inflammatory properties, exemplifies mechanistic extension in this context. The compound modulates profibrotic signalling (e.g., TGF-β-driven pathways) and fibroblast activity, but also inhibits cytokine production [3]. It is now an established therapy with systematic reviews supporting benefit targeting various forms of pulmonary fibrosis [4]. New delivery strategies are exploring whether the pharmacology of pirfenidone can also be harnessed in other airway diseases. Early human studies demonstrated the feasibility and short-term safety of aerosolized or nebulized pirfenidone, achieving local airway exposure with reduced systemic levels [5]. These advances invited investigation of pirfenidone as a potential steroid-sparing anti-inflammatory compound in obstructive airway diseases [6], particularly since GCS carry well-documented acute and cumulative risks [7]. In our recent preclinical studies, we mimicked contemporary clinical practice in COPD by comparing the effects of daily inhaled pirfenidone versus inhaled GCS on virus infection [6]. Reductions in airway inflammation were greater for pirfenidone and the compound was superior since it curbed virus infection whilst GCS enhanced infection (Figure 1). These findings suggest that pirfenidone may have substantial benefit if repurposed as an anti-inflammatory agent, using the same clinical pathway (regular inhaled medication and oral as needed) to replace conventional GCS treatments. In contrast to this traditional approach used to support repurposing of pirfenidone, a recent report in the New England Journal of Medicine elegantly illustrates the application of AI—in this case, to find effective treatment for a patient with idiopathic multi-centric Castleman's disease, a life-threatening disorder of unknown cause [8]. The patient had not responded to an array of treatments and was considered for palliative care. Analysis of patient bio-specimens using proteomic, transcriptomic, in vitro modelling, and computational techniques implicated tumour necrosis factor (TNF) in disease pathogenesis. In parallel, AI and machine learning approaches were used to generate predictive scores for the probability that FDA-approved medications could be repurposed to treat the patient. Adalimumab, a TNF inhibitor approved for autoimmune disorders, was deemed the treatment most likely to help. Administration induced a sustained remission lasting 24 months as of publication. Importantly, these techniques do not ‘prove’ efficacy; they generate hypotheses that can be refuted or supported experimentally and clinically. The overall arc is well illustrated by the example noted: multi-omic and computational analyses implicated key signalling pathways in a refractory inflammatory disorder, and an approved therapy identified by AI induced sustained remission [8]. While the disease context differs from respiratory medicine, the workflow using the AI-supported toolkit culminating in clinical testing can clearly map to pulmonary indications. Furthermore, repurposing ultimately requires mechanistic plausibility and regulatory-grade evidence. A central challenge for AI and repurposing will be providing a clear rationale for recommendations: ‘black-box’ predictions only are rarely sufficient for clinical decision-making or regulatory approval [9]. Transparent proof should therefore support the evidence generated by AI (which studies, which signatures, which graph relations), some indication of the level of ‘probability’ should be provided, and ultimately testable mechanistic hypotheses should be articulated with indications of how they can be refuted or verified. Regulatory considerations include demonstration of efficacy and safety in the new indication, potential for label extension, and alignment with current standards of care. Regulators increasingly encourage innovative trial designs when justified by strong mechanistic rationale and unmet need but require a reasonable rationale from model outputs to experimental plans. Equity and access warrant early attention. Data used to train models may under-represent populations most burdened by respiratory disease, potentially leading to biased predictions. Global deployment must avoid creating a two-tiered innovation ecosystem in which AI-identified therapies are evaluated and adopted only in well-resourced settings. Initiatives to share de-identified data, harmonise ontologies, and build capacity for trials in low- and middle-income regions can mitigate these risks. Incentives also matter: many attractive repurposing candidates are generic, creating a value gap for sponsors. Public-private partnerships, non-profit models, and patent exclusivity around new clinical uses can help create incentives while maintaining affordability. Limitations and guardrails are crucial. AI-driven repurposing inherits limitations from its inputs (publication bias, heterogeneous data quality) and from modelling choices (overfitting, spurious correlations). Robust external validation, preregistration of analysis plans where feasible, and independent replication can reduce false discovery. In respiratory medicine, outcome heterogeneity (e.g., exacerbation rates influenced by season and infection patterns) complicates signal detection, and careful endpoint selection and stratification are essential. Finally, AI is not a substitute for rigorous experimentation; it is an accelerator of hypothesis generation and triage. In summary, AI provides a hypothesis-generating capability linking heterogeneous biomedical data (omics, imaging, clinical trials, and electronic health records) to disease mechanisms and candidate therapeutics. In respiratory medicine, the evolution of pirfenidone from systemic anti-fibrotic to potential application as an inhaled and oral anti-inflammatory therapy shows how traditional repurposing strategies can unlock new therapeutic options. However, AI can now systematise what has historically been piecemeal: it integrates diverse data to generate mechanistically anchored, testable hypotheses—at scale. The pathway to impact runs through plausible models, efficient experimental screens, adaptive and biomarker-enriched trials, and vigilant attention to safety, equity, and incentives. With AI, respiratory medicine can shift towards repurposing and do so faster, at lower cost, and with broader access than de novo discovery alone is likely to afford. J.E.B. is an Editorial Board member of Respirology and was excluded from all editorial decision-making related to this article. P.G.B. and B.J.T. have nothing to disclose.