AI in drug discovery: 2025 in review

For those that hoped artificial intelligence would help to deliver drugs that actually work, the hope lives on. Here, Dr Raminderpal Singh reflects on the sobering reality check that 2025 delivered to the industry.

The year 2025 will be remembered as the moment AI drug discovery transitioned from promise to proof – albeit with significant caveats that temper the triumphalism. After a decade of escalating investment and bold predictions, the field delivered its first genuine clinical validation while simultaneously experiencing setbacks that exposed the technology’s contemporaneous limitations. For those of us working at the intersection of AI and pharmaceutical research, 2025 represented not a revolution but a sobering reality check: AI can accelerate early-stage discovery, but it has not yet solved the fundamental challenge of clinical success rates.

The headline achievement was the first drug with both target and molecule designed entirely by AI completing Phase IIa trials. Published in Nature Medicine in June, the trial demonstrated dose-dependent improvement in forced vital capacity for idiopathic pulmonary fibrosis patients. The drug reached preclinical candidate nomination in just 18 months from target identification – a timeline that would have required three to four years through traditional methods.

However, the trial enrolled only 71 patients and the efficacy signal, while promising, requires validation in larger cohorts. This exemplifies a pattern throughout 2025: genuine progress accompanied by the reality that clinical validation remains years away for most AI-discovered compounds.

Regulatory framework takes shape

Perhaps the most consequential development of 2025 occurred on 6 January, when the US Food and Drug Administration (FDA) issued draft guidance titled, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. The framework establishes a seven-step credibility assessment based on “context of use,” requires lifecycle maintenance plans and mandates transparency about model architectures and training data.

In December, the FDA qualified its first AI-based tool approved for use in drug development clinical trials – a cloud-based platform that helps pathologists score liver biopsies in NASH/MASH trials.

Notably, the guidance excludes AI used in early discovery – focusing solely on applications affecting regulatory decisions. This distinction matters: while AI can accelerate target identification and lead optimisation, the FDA remains appropriately cautious about AI’s role in regulatory-critical activities. The agency has reviewed 500+ AI-related submissions since 2016 without approving any AI-discovered drugs and reports suggest early implementations of the FDA’s own generative AI tools experienced ‘hallucination’ issues.

In December, the FDA qualified its first AI-based tool approved for use in drug development clinical trials – a cloud-based platform that helps pathologists score liver biopsies in NASH/MASH trials. This represents formal regulatory acceptance of AI in the drug development process, though notably for a quality assessment tool rather than drug design itself.

The economics of AI drug discovery

Total deal value for AI drug discovery partnerships exceeded $15 billion in announced ‘biobucks’ during 2025, but the financial reality proves far more modest. Multiple major partnerships announced at $5+ billion included upfront payments of just 2 percent of the headline value. The full amounts depend on multiple drug candidates achieving clinical and commercial milestones, each with historically high failure rates.

This 50:1 ratio between announced biobucks and actual upfront payments reveals the industry’s appropriate caution about AI’s unproven ability to deliver clinical success.

The year also exposed financial fragility among AI drug discovery companies. Multiple firms ‘deprioritised’ clinical programmes, announced 20-30 percent workforce reductions and reported substantial net losses. Several well-funded companies shut down entirely or pursued delisting. The pattern suggests smaller AI drug discovery companies face existential pressures in a market that demands clinical validation rather than technological sophistication.

Technology advances and persistent limitations

On the technology front, major advances in protein structure prediction extended capabilities to DNA, RNA, ligands, antibodies and their interactions. New models achieved 50 percent greater accuracy than traditional methods for drug-like interactions on standard benchmarks and successfully predicted thousands of post-translational modification protein–ligand complexes.

On the technology front, major advances in protein structure prediction extended capabilities to DNA, RNA, ligands, antibodies and their interactions.

However, competition results showed that newer AI models did not significantly outperform older methods for protein–ligand interaction prediction. This underscores a critical limitation: accurate structure prediction does not guarantee druggable targets or successful molecules. Current models struggle with conformational changes and show persistent biases towards particular receptor states.

In antibody design, new models achieved unprecedented 16-20 percent experimental hit rates in zero-shot de novo design across dozens of novel targets using only 20 candidates per target – a 100-fold improvement over existing methods. This represents genuine progress in biologics design, though the compounds remain years from clinical validation.

Clinical reality check

The year delivered sobering clinical setbacks alongside the Phase IIa success. Multiple AI-designed drugs were deprioritised, shelved after Phase II, or showed no efficacy signal in trials. Deep Genomics founder Brendan Frey stated bluntly: “AI has really let us all down in the last decade when it comes to drug discovery – we’ve just seen failure after failure.”

This clinical attrition underscores the fundamental challenge: AI has not demonstrably improved the pharmaceutical industry’s ~90 percent clinical failure rate. While AI can compress early discovery timelines by 30-40 percent and reduce preclinical candidate development to 13-18 months (versus traditional three to four years), clinical trial duration, regulatory review timelines and manufacturing scale-up remain bound by biology, patient enrolment and regulatory requirements – none of which AI can bypass.

Commentary from scientific publications questioned whether AI improves clinical success rates, noting that AI-discovered compounds show progression rates similar to traditionally discovered compounds. Until AI-discovered drugs achieve regulatory approval and commercial success, the field remains in a ‘proof-of-concept’ phase rather than a proven paradigm shift.

Infrastructure and integration challenges

Major pharmaceutical companies announced construction of industry-leading supercomputers powered by thousands of advanced GPUs, operational in early 2026. Several companies launched platforms for sharing AI models with biotech partners, providing access to models trained on proprietary data from hundreds of thousands of molecules.

A survey of tech executives found 68 percent cite poor data quality and governance as the main reason AI initiatives fail.

Some companies deployed humanoid AI scientists in robotic laboratories, while others raised substantial funding to build autonomous AI-robot labs. These ‘self-driving laboratories’ accelerate the design–make–test–learn cycle but have not yet demonstrated ability to autonomously discover validated drug candidates. Integration of wet-lab and dry-lab operations remains a significant organisational challenge.

A survey of tech executives found 68 percent cite poor data quality and governance as the main reason AI initiatives fail. High-quality, rigorously curated datasets with biological, pharmacological and clinical annotations remain scarce due to costs, privacy regulations and data-sharing restrictions. The industry’s fundamental challenge is not algorithmic sophistication but data availability.

Conclusion: cautious optimism

The year 2025 advanced AI drug discovery from speculative technology to early clinical validation. The first Phase IIa results for a fully AI-designed drug demonstrate that the approach can show clinical efficacy, while regulatory frameworks from the FDA and EU AI Act create clearer pathways for development.

However, no AI-discovered drug has achieved FDA approval as of December 2025 – a reality that frames both the achievements and challenges ahead. The most important question for 2026 is not whether AI can accelerate preclinical timelines – we know it can –

but whether it can improve clinical success rates. Until that question receives a positive answer through Phase III data and regulatory approvals, the pharmaceutical industry’s cautious approach to AI investment appears entirely justified.

For laboratory scientists and pharmaceutical researchers, 2025 confirmed that AI is a powerful tool for early-stage discovery but not a panacea for drug development’s fundamental challenges. The technology has earned its place in the R&D toolkit while simultaneously demonstrating its current limitations. This balanced perspective – progress without hyperbole – represents the maturation of the field.

Meet the author

Dr Raminderpal Singh

Dr Raminderpal Singh is a recognised visionary in the implementation of AI across technology and science-focused industries. He has over 30 years of global experience leading and advising teams, helping early- to mid-stage companies achieve breakthroughs through the effective use of computational modelling. Raminderpal is currently the Global Head of AI and GenAI Practice at 20/15 Visioneers. He also founded and leads the HitchhikersAI.org open-source community and is Co-founder of the techbio, Incubate Bio. 

Raminderpal has extensive experience building businesses in both Europe and the US. As a business executive at IBM Research in New York, Dr Singh led the go-to-market for IBM Watson Genomics Analytics. He was also Vice President and Head of the Microbiome Division at Eagle Genomics Ltd, in Cambridge. Raminderpal earned his PhD in semiconductor modelling in 1997 and has published several papers and two books and has twelve issued patents. In 2003, he was selected by EE Times as one of the top 13 most influential people in the semiconductor industry.

Related topics
Antibodies, Artificial Intelligence, Biologics, Clinical Trials, Computational techniques, Drug Discovery, Drug Discovery Processes, Target Validation, Targets, Technology, Translational Science