By Stuart Kerr, Technology Correspondent, LiveAIWire
AI and drug discovery moved from theoretical promise to partial clinical proof in June 2025, when Insilico Medicine published Phase IIa data for rentosertib, an AI-designed drug candidate for idiopathic pulmonary fibrosis, in Nature Medicine. Patients on the highest dose showed a mean lung function improvement of 98.4 millilitres over twelve weeks, against a mean decline of 20.3 millilitres on placebo.
It was the first peer-reviewed Phase IIa result for a drug candidate that AI both identified and designed. As of mid-2026, more than 173 AI-discovered drug programmes are in clinical development, up from just 3 in 2016 and 67 in 2023, according to industry pipeline tracking. No AI-discovered drug has yet received FDA approval. The field has reached the point where AI and drug discovery stop being judged on laboratory potential and start being judged on the only bar that matters in medicine: clinical efficacy and safety in human patients.
The gap between what AI and drug discovery can do in the lab and what they have delivered for patients is real, but it is closing. How fast it closes, and which patients benefit first, depends on scientific, regulatory and economic factors that are not all moving in the same direction.
Where AI and Drug Discovery Are Genuinely Compressing the Timeline
The traditional drug development timeline runs 10 to 15 years from target identification to regulatory approval. AI is compressing the earliest phases specifically. Models trained on protein structure data, genomic information and published research can now identify promising drug targets and generate candidate molecules in months rather than years, and the growth in AI-originated candidates entering clinical stages reflects that: from 3 in 2016 to 67 in 2023 to more than 173 as of 2026.
Our own coverage of how AI is reshaping medicine more broadly traced this back to a specific inflection point: AlphaFold2’s protein structure predictions, which won a Nobel Prize in Chemistry in 2024, gave AI and drug discovery a genuine scientific foundation rather than just a marketing narrative, accelerating target identification across multiple therapeutic areas at companies well beyond Insilico.
Phase 1 success rates for AI-discovered drugs are tracking at 80 to 90 percent, compared with roughly 40 to 65 percent for traditionally discovered candidates. Phase 2 success rates for AI-discovered drugs run somewhat higher too, though the sample size at that stage is still small enough that the numbers carry real statistical uncertainty. A 2025 RAND Corporation study, published in JAMA Network Open, separately challenged the widely cited 2.6 billion dollar average cost per approved drug, finding a median research and development cost of 708 million dollars across a cohort of 38 recently approved drugs once high-cost outliers were excluded, a figure that AI’s efficiency gains in the earliest phases are beginning to compress further.
What the FDA Is Actually Doing About AI and Drug Discovery
The FDA has moved carefully but definitely to accommodate AI’s growing role in the field. In January 2025 the agency issued its first draft guidance on AI used to support regulatory decisions for drug and biological products, establishing a seven-step credibility assessment framework. In January 2026, the FDA and the European Medicines Agency jointly published a set of Guiding Principles of Good AI Practice in Drug Development, extending similar thinking across the full product lifecycle.
In December 2025, the FDA qualified its first AI-based drug development tool, a cloud-based system that helps pathologists score liver biopsies in trials for metabolic dysfunction-associated steatohepatitis. The agency has reviewed more than 500 AI-related submissions since 2016, most of them concentrated in oncology and neurology.
The regulatory framework’s caution is appropriate but it creates a specific timeline pressure worth understanding. The FDA’s January 2025 guidance applies to AI used in activities that affect a regulatory decision. It explicitly does not apply to AI used in early discovery. That distinction means AI can dramatically accelerate the preclinical work that generates a drug candidate without touching a single regulatory constraint, while the clinical trial and approval process that follows remains primarily human-governed and largely uncompressed.
The first AI-discovered drug to clear every trial phase and win FDA approval will still have spent years inside the part of the pipeline that AI and drug discovery, on current evidence, cannot meaningfully speed up.
The Failures That Deserve Equal Billing
An honest account of AI and drug discovery has to sit the failures next to the successes rather than treat them as a footnote. Exscientia’s DSP-1181, the first AI-designed molecule to enter human clinical trials in 2020, completed Phase 1 with a favourable safety profile but was discontinued shortly after, with no progression to Phase 2. That outcome is the clearest evidence that faster candidate generation does not guarantee clinical success, whatever the speed of the discovery process behind it.
Additional clinical failures across the current pipeline are close to statistically inevitable given the historical attrition rates that apply to any drug candidate, AI-discovered or otherwise, even if the per-candidate failure rate for AI-discovered molecules is running lower than the historical average so far.
What This Actually Changes for Patients
The question that matters to a patient is not whether AI and drug discovery can move faster in the laboratory. It is whether that speed reaches a patient sooner. The clinical trial phases that follow AI-accelerated preclinical work are not themselves accelerated by AI in most current deployments. A candidate entering Phase 1 in year three rather than year six still faces the same multi-year clinical evaluation timeline that AI does not currently compress.
On current evidence, the net patient benefit of AI and drug discovery is measured in years of timeline reduction, not decades, meaningful but not transformative on the timescale that matters to a specific patient waiting for a treatment for idiopathic pulmonary fibrosis or an aggressive cancer where disease progression will not wait for a regulatory timeline to shorten.
Our own coverage of what the evidence actually shows about AI cancer diagnosis found a similar pattern one step earlier in the care pathway: genuine, trial-backed progress concentrated in specific, well-studied applications, sitting alongside a much larger volume of AI health products whose evidence has not caught up with the marketing. AI and drug discovery deserves the same discipline: judge each programme against its own trial data, not against the category’s headline potential.
The same evidence-first standard runs through our reporting on where the proof for AI physiotherapy tools is genuinely strong and where it is still catching up to the marketing, a pattern that recurs across almost every AI health category once the trial data is examined closely rather than taken on a vendor’s word. Applied to AI and drug discovery specifically, that means treating a Phase 1 success rate as encouraging but provisional, and reserving real confidence for Phase 3 data that has not yet arrived for most of the current pipeline.
The Rare Disease Exception
The clearest near-term exception may be rare diseases, where small patient populations make traditional drug discovery economically unviable regardless of the underlying science. AI’s ability to compress the cost and timeline of early-phase discovery changes that calculation directly, making more rare disease programmes viable to fund in the first place.
Insilico’s own rentosertib programme, now advancing toward Phase 3 following its Phase IIa results, targets idiopathic pulmonary fibrosis, a condition with a five-year survival rate worse than many cancers and a limited treatment landscape. If the later-phase data holds up, it will represent one of the clearest direct examples of what AI and drug discovery can do for a specific patient population, not because AI made the drug cheaper for that patient directly, but because AI made developing the drug economically viable to attempt at all.
The Patient Access Question Nobody Is Answering
Even a successful, FDA-approved drug born from AI and drug discovery does not guarantee patient access. Pricing is set by the market exclusivity and pricing power that regulatory approval confers, not by the cost efficiencies AI introduced during development. A drug that cost meaningfully less to develop because AI compressed the preclinical timeline is not automatically a cheaper drug at the pharmacy counter, unless competitive dynamics or a specific pricing intervention produce that outcome.
In the United States, where prescription drug pricing is largely unregulated, cost reductions from AI and drug discovery may accrue primarily to pharmaceutical company margins rather than to the patients whose treatment those savings were meant to make more viable in the first place.
The Part of the Pipeline AI Is Also Quietly Changing
Beyond candidate discovery itself, AI is reshaping clinical trial operations in ways that compress timelines independently of preclinical acceleration. Patient recruitment, historically one of the slowest and most expensive parts of any trial, is increasingly supported by systems that match eligible patients to trials using electronic health records, genetic profiles and prior treatment histories. AI-powered adverse event monitoring that flags early safety signals before they become protocol-stopping events, and adaptive trial designs that adjust dosing or eligibility criteria in real time, are addressing a persistent and costly problem: patient dropout during a trial that is already underway.
Whether these operational gains translate into faster approvals depends on a distinction that is easy to miss. Earlier identification of a safety signal is only valuable if it leads to earlier, better-informed intervention. If it simply produces earlier awareness of a problem that would have stopped the trial regardless, the calendar-time benefit of AI and drug discovery in this part of the pipeline is smaller than the operational efficiency numbers alone suggest.
The Next 36 Months Will Answer the Real Question
Fifteen AI-discovered drug programmes are currently in Phase 3, the stage where drugs most often die: historically, 50 to 65 percent of drugs entering Phase 3 fail, either on efficacy or on a safety signal that did not appear earlier. Whether AI’s improvements in target identification and candidate optimisation translate into meaningfully better Phase 3 outcomes, rather than simply feeding more candidates faster into the same historical filter, is a question the current pipeline will answer empirically over the next two to three years, not one that can be settled by pointing to Phase 1 success rates alone.
The balanced forecast worth holding in mind is validation and disappointment in roughly equal measure. The validation, when it arrives, will be real. So will the disappointments. What determines whether AI and drug discovery has been worth the investment, for patients rather than for pipeline headlines, is the net balance of approved treatments that emerges from those 15 Phase 3 programmes, and from the ones that follow them.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.