What Are the Regulatory Pathways for AI-Discovered Therapeutics?
Regulatory pathways for AI-discovered therapeutics treat AI as a development tool: CDER received 500+ submissions with AI components from 2016–2023, but human studies still establish safety and effectiveness (FDA, 2023; FDA, 2025). Model validation, data quality, and lifecycle documentation are central. Motif scopes PMID-linked biomarker evidence for IND and qualification dossier backgrounds.
TL;DR: AI Drug Regulation
- FDA treats AI as a tool in development; human studies still establish safety and effectiveness (FDA, 2023)
- CDER received 500+ submissions with AI components from 2016 to 2023 (FDA, 2025)
- Model validation, data quality, and lifecycle documentation are central themes in FDA AI discussion papers
- Biomarker strategies for AI-enriched trials still require analytical and clinical validity (Johnson et al., 2024)
- Motif supplies PMID-linked literature evidence for biomarker backgrounds; it does not replace regulatory meetings
Note: FDA guidance evolves. Consult current CDER resources and meeting minutes for your product type and development stage.
From the Motif team: AI-discovered programs still need cited biomarker backgrounds for INDs and trial protocols. We extract literature associations with PMIDs from PubMed, PMC, and Europe PMC and score evidence with GRADE-adapted tiers. We do not submit regulatory packages or validate AI models.
AI changes how targets and compounds are proposed, but regulatory approval still rests on human evidence of safety and effectiveness. The FDA Center for Drug Evaluation and Research (CDER) states that drugs must meet the same standards regardless of whether discovery used traditional or AI-assisted methods (FDA, 2023). FDA-2023-D-067323.
What FDA Has Published So Far
In May 2023, FDA released a discussion paper on using artificial intelligence and machine learning in drug and biological product development. It is not binding guidance; it outlines considerations for data credibility, model development, and monitoring across the product lifecycle (FDA, 2023). FDA-2023-D-067323.
CDER reported more than 500 submissions with AI components from 2016 through 2023 and used public comments on the 2023 discussion paper to inform later draft guidance on AI supporting regulatory decision-making (FDA, 2025). Early engagement through INTERACT or formal meetings remains the practical path when AI outputs influence pivotal decisions.
Rajpurkar et al. (2022) emphasize that clinical AI requires validation beyond prototype performance, with attention to dataset shift and deployment context.1 The same skepticism applies when AI models rank targets or stratify patients: regulators need traceable data and pre-specified validation plans.
AI-Specific Development Considerations
Training data and documentation
Harrer et al. (2019) review AI in drug discovery and stress that biased or incomplete training data propagates into candidate selection.2 Submissions should document data provenance, quality control, and how models were locked before confirmatory studies.
Explainability and model lifecycle
Rajkomar et al. (2019) note that machine learning models in medicine need rigorous evaluation for generalization and unintended bias.3 For drug development AI, documentation typically covers algorithm description, training and test splits, performance metrics, and known failure modes.
Where human trials still decide
FDA's discussion paper frames AI as supporting hypothesis generation, trial simulation, and manufacturing optimization. Clinical benefit for patients still requires appropriately designed human studies with prespecified endpoints. Wong et al. (2019) report oncology likelihood of approval was 3.4% in their 2000 to 2015 registry sample.4 AI does not remove attrition; it may help prioritize better hypotheses earlier.
Biomarkers in AI-Enriched Programs
Many AI-discovered programs pair therapies with enrichment biomarkers or companion diagnostics. Johnson et al. (2024) review the FDA biomarker qualification program and recommend clear context-of-use statements and fit-for-purpose validation.5 Literature evidence supports gap analysis; qualification letters and IDE/IVD pathways still require prospective data.
FDA-NIH BEST separates analytical validity, clinical validity, and clinical utility (FDA-NIH, 2016).6 An AI model that predicts responders in silico does not substitute for those evidence tiers in the intended-use population.
Literature Evidence Before IND and Meetings
Teams preparing AI-assisted programs often need a cited map of biomarker literature before pre-IND briefings:
- Which markers appear as predictive associations with the target pathway and therapy class?
- Which papers report failed validations or conflicting cohorts?
- Do cross-references (ClinVar, PharmGKB, Open Targets) align with publication claims?
In Motif, that workflow uses PubMed, PMC, and Europe PMC search, structured extraction across 69 biomedical entity types, GRADE-adapted scoring, and export to Word or Excel with PMIDs.
Failure modes:
- Citing AI-generated summaries without PMIDs in briefing books
- Presenting literature association as analytical validation
- Omitting negative biomarker studies from gap analysis
- Assuming FDA discussion papers replace product-specific meetings
Read our blog on FDA biomarker validation to learn more about qualification paths. For literature-stage discovery, see Motif biomarker discovery.
Frequently Asked Questions
How does FDA regulate AI in drug development?
FDA treats AI as a tool supporting development decisions. Discussion papers emphasize model validation, training data quality, transparency, and lifecycle monitoring. AI does not replace human trials for establishing safety and effectiveness of therapeutics (FDA, 2023).
Do AI-discovered drugs follow a different approval path?
No separate pathway exists. AI-discovered molecules, targets, or biomarkers still enter standard IND and NDA/BLA processes. The AI component must be documented so reviewers can assess reliability of development decisions it informed.
How does Motif fit regulatory biomarker workflows?
Motif supplies PMID-linked literature for LOI, QP, and briefing-book gap analysis—validation cohorts, assays, and BEST categories with traceable sources. It supports evidence scoping; it does not replace FDA meetings or prospective validation studies.
References
- Rajpurkar, P., et al. (2022). AI in health and medicine. Nature Medicine, 28(1), 31-38. PMID: 35058619
- Harrer, S., et al. (2019). Artificial intelligence for clinical and drug development. Drug Discovery Today, 24(9), 1891-1902. PMID: 31557863
- Rajkomar, A., et al. (2019). Machine learning in medicine. New England Journal of Medicine, 380(14), 1347-1358. PMID: 30943338
- Wong, C.H., et al. (2019). Estimation of clinical trial success rates and related parameters. Biostatistics, 20(2), 273-286. PMID: 29394327
- Johnson, K.R., et al. (2024). The FDA biomarker qualification program: review and recommendations. Nature Reviews Drug Discovery, 23(4), 267-283. PMID: 38291248
- FDA-NIH Biomarker Working Group. (2016). BEST (Biomarkers, EndpointS, and other Tools) Resource. PMID: 27010052
- FDA. (2023). Using Artificial Intelligence and Machine Learning in the Development of Drug and Biological Products. Discussion Paper. FDA-2023-D-067323.
- FDA. (2025). Artificial Intelligence for Drug Development. CDER web resource. fda.gov/about-fda/center-drug-evaluation-and-research-cder/artificial-intelligence-drug-development
