Can AI drugs survive human trials?

Introduction

AI can now suggest drug targets, generate candidate molecules, predict protein structures and narrow years of laboratory search into months. That has made drug discovery one of the strongest real-world examples behind the broader idea that advanced AI could accelerate science and eventually help extend healthy human life. Yet the hardest question is not whether AI can design molecules. It is whether those molecules survive contact with human biology.

The central bottleneck remains the clinical trial system. Most experimental medicines fail not because researchers cannot generate ideas, but because treatments that appear promising in cells, animal models or computer simulations often do not help real patients enough to justify approval. AI may compress the search phase of medicine, but phase 2 and phase 3 trials still determine whether a drug is safe, effective and worth using. The future of AI-enabled longevity medicine therefore depends less on producing candidate drugs and more on improving the long, expensive process of proving that they work.

What AI speeds up in early discovery

Traditional drug discovery is partly a search problem. Researchers must identify a biological target, understand how it behaves, design molecules that interact with it and then refine those molecules through repeated rounds of testing. Much of this process has historically involved large amounts of trial and error.

Recent AI systems have become useful because they reduce parts of that blind search.

A major example is AlphaFold, developed by Google DeepMind and collaborators. AlphaFold 2 dramatically improved the prediction of protein structures, while AlphaFold 3 expanded this capability to model interactions involving proteins, DNA, RNA and small molecules. That matters because many diseases involve molecular interactions that are difficult to observe directly. Better structural predictions can help researchers identify potential drug targets and design compounds more efficiently.[Nature]nature.comAccurate structure prediction of biomolecular interactions…by J Abramson · 2024 · Cited by 14582 — Here we describe our AlphaFol…

The practical gains appear in several stages:

The attraction is obvious. Drug development often takes more than a decade and costs billions of pounds. If AI can eliminate large numbers of weak candidates earlier, the entire pipeline could become faster and cheaper.

Several companies now claim substantial reductions in discovery timelines. Insilico Medicine, for example, reported moving its idiopathic pulmonary fibrosis drug candidate rentosertib from target identification to clinical testing far faster than conventional timelines. The company argues that AI reduced the time needed to reach a preclinical candidate by helping identify targets and generate molecules more efficiently.[Insilico Medicine]insilico.comInsilico MedicineCase Study: Insilico's TransformationWith Rentosertib now finished Phase IIa clinical trial… A critical bottleneck in…

This is why drug discovery occupies such an important place in the wider AI bloom story. Scientific acceleration is easier to imagine when there is a concrete mechanism: machine intelligence helping researchers search enormous biological possibility spaces that humans alone cannot explore efficiently.

Why phase 2 and phase 3 remain hard

The difficulty is that discovering a molecule and proving a medicine are different problems.

A model may predict that a compound binds strongly to a protein. That does not mean the compound reaches the right tissue in the body, remains stable long enough to work, avoids dangerous side effects, interacts safely with other biological systems or improves patient outcomes.

This gap explains why the industry’s biggest failure rates occur after promising laboratory results.

Biology is messier than prediction

Human bodies are not static molecular diagrams. Diseases involve immune systems, metabolism, genetics, age, lifestyle and environmental influences interacting across many levels.

Even highly accurate structure prediction does not solve these complexities. Researchers have praised AlphaFold 3 as a major advance while also highlighting limitations involving stereochemistry, molecular dynamics and prediction accuracy in genuinely novel biological situations. Nature[Cell]cell.comDrug development in the AI era: AlphaFold 3 is coming!by Y Shi · 2024 · Cited by 24 — We also noticed that AF3 has limitations with r…

A molecule that looks excellent on a computer may behave very differently inside a living patient.

This is especially important for ageing-related diseases. Conditions such as Alzheimer’s disease, fibrosis, cardiovascular disease and many cancers involve multiple interacting biological systems. A drug may successfully affect one target while leaving the broader disease process largely unchanged.

The phase 2 problem

Drug developers often describe phase 2 as the industry’s graveyard.

Phase 1 trials primarily test safety in small groups of people. A drug that survives phase 1 has shown that it can usually be administered without unacceptable immediate toxicity.

Phase 2 is different. Researchers must demonstrate that the treatment actually helps patients.

That requires answering difficult questions:

Many candidate drugs fail at this stage because a promising biological theory turns out not to translate into meaningful clinical improvement.

The emerging evidence suggests that AI-discovered drugs are not exempt from this reality. While AI may improve candidate selection, there is little evidence so far that it has eliminated the fundamental challenge of demonstrating efficacy in humans. Some observers note that AI-originated compounds appear to encounter phase 2 failure rates similar to those seen across the broader pharmaceutical industry.[DeepCeutix]deepceutix.comDeep Ceutix Your AI Can Design a MoleculeIt Can't Formulate a Drug.23 Feb 2026 — AI-discovered compounds show 80-90% Phase I success rates but only ~40% Phase II success, indisti…

Phase 3 is even more demanding

A successful phase 2 result does not guarantee approval.

Phase 3 trials often involve hundreds or thousands of patients across multiple sites. They are designed to answer whether a treatment works reliably in diverse real-world populations.

This stage introduces additional challenges:

An AI-generated molecule faces exactly the same evidential burden as any other medicine.

This explains why many companies that promote AI drug discovery still speak cautiously about timelines. Even firms built around advanced AI platforms continue to measure progress in years rather than months once clinical testing begins. Isomorphic Labs, for example, has repeatedly emphasised the challenge of moving from computational breakthroughs to human trials, delaying its projected clinical timeline while continuing to expand investment in AI-driven discovery.[Reuters]reuters.comGoogle-backed Isomorphic Labs delays clinical trial timelineThis update was shared by founder and CEO Demis Hassabis during the World Economic Forum in Davos, Switzerland. The company, established…

Rentosertib: an early test of the idea

The most closely watched case in this field is rentosertib, previously known as INS018_055, developed by Insilico Medicine for idiopathic pulmonary fibrosis.

The importance of the programme is not simply that AI contributed to its discovery. The significance is that it reached human testing and produced clinical data.

In 2023, the company announced that the drug had entered phase 2 trials, describing it as the first generative-AI-discovered drug to reach that stage.[Insilico Medicine]insilico.comInsilico MedicineCase Study: Insilico's TransformationWith Rentosertib now finished Phase IIa clinical trial… A critical bottleneck in…

More recently, researchers reported results from a randomised phase 2a study suggesting that the AI-discovered target-and-drug combination was safe and showed signs of efficacy in patients with idiopathic pulmonary fibrosis. The study has been widely discussed because it represents one of the clearest attempts to validate an end-to-end AI drug discovery process using clinical evidence rather than laboratory demonstrations.[PubMed]pubmed.ncbi.nlm.nih.govA generative AI-discovered TNIK inhibitor for idiopathic…by Z Xu · 2025 · Cited by 128 — Despite substantial progress in artific…

The key point is not that rentosertib proves AI has solved drug development. It does not.

Instead, it provides a more realistic milestone. AI-generated hypotheses can survive long enough to reach meaningful human testing. That is a stronger claim than saying AI can generate attractive molecular structures.

At the same time, even a successful phase 2a study is not equivalent to regulatory approval. The larger and more difficult stages still lie ahead.

Can AI help with the trial bottleneck itself?

The optimistic case is that AI eventually improves not only discovery but also clinical development.

Several routes are being explored.

Better patient selection

Many trials fail because researchers recruit patients whose disease differs in important ways from the biological mechanism being targeted.

AI systems may help identify subgroups more likely to respond to treatment by analysing genetics, biomarkers, imaging and health records.

In principle, this could make treatment effects easier to detect and reduce the number of patients required for successful studies.

Better prediction before trials begin

Researchers are increasingly exploring multimodal models that combine biological, chemical and clinical information.

Recent work has shown that AI systems can use preclinical data to predict aspects of clinical outcomes, including adverse effects and drug-combination risks. The long-term goal is to identify weak candidates before expensive human trials begin.[arXiv]arxiv.orgMultimodal AI predicts clinical outcomes of drug combinations from preclinical dataMarch 4, 2025…Published: March 4, 2025

If such approaches become reliable, they could reduce the number of doomed compounds entering phase 2 and phase 3 studies.

Faster trial operations

Clinical trials generate huge amounts of data.

AI may help:

These gains would not remove the need for trials, but they could lower costs and shorten timelines.

The distinction matters. Much public discussion assumes AI will somehow replace clinical testing. A more plausible outcome is that AI makes trials more efficient while leaving the requirement for human evidence intact.

What success would mean for longevity medicine

The strongest longevity implications do not come from a single miracle anti-ageing drug.

Instead, they come from the possibility of compounding improvements across many diseases.

If AI substantially lowers the cost and time required to discover and validate treatments, researchers could explore far more biological hypotheses than today’s pharmaceutical system allows.

That could matter especially for diseases associated with ageing:

Many of these areas suffer from a basic economics problem. Testing new ideas is expensive, and the failure rate is high. If AI increases the number of ideas that can be explored and improves the quality of candidates entering trials, the result may be a larger pipeline of therapies rather than one dramatic breakthrough.

In the broader AI bloom framework, this is where the stakes become unusually large. A civilisation that can generate medical knowledge faster may not merely add a few treatments. Over decades, it could build a much richer understanding of ageing, disease and prevention. The gains would accumulate across generations.

But that future depends on solving both halves of the problem. Accelerating discovery alone is not enough. Scientific abundance only reaches patients when evidence, regulation, manufacturing and healthcare systems can keep pace.

The deeper lesson: intelligence is speeding up faster than validation

One emerging pattern across AI is that idea generation is becoming cheaper than verification.

Models can now propose molecules, proteins, hypotheses and experimental directions at unprecedented scale. Yet the physical world still imposes constraints.

A proposed medicine must be synthesised.

A biological mechanism must be tested.

A patient must be treated.

A clinical outcome must be measured.

These steps consume time because reality, unlike a simulation, cannot be parallelised indefinitely.

That makes the clinical trial bottleneck one of the most revealing tests of the wider AI bloom thesis. If advanced AI can genuinely improve both discovery and validation, medicine could become one of the first domains where scientific acceleration translates into longer, healthier lives. If validation remains stubbornly slow, AI may generate far more medical possibilities than society can efficiently prove.

The future of AI-discovered drugs therefore rests on a deceptively simple question: can humanity accelerate the process of learning what actually works in real patients? The answer may determine whether AI’s medical revolution remains a laboratory achievement or becomes a genuine expansion of human health and longevity.

Amazon book picks

Further Reading

Books and field guides related to Can AI drugs survive human trials?. Use these as the next step if you want deeper reading beyond the article.

Book

Deep Medicine

By Eric Topol

Explains AI's promise and limits in real medical practice.

See on Amazon

Book

The Drug Hunters

By Donald R. Kirsch, Ogi Ogas

Explains the long path from discovery to approved therapies.

See on Amazon

Book

The AI Revolution in Medicine

By Peter Lee, Carey Goldberg et al.

Places AI medical tools within clinical workflows and evidence demands.

See on Amazon

Book

The Billion-Dollar Molecule

By Barry Werth

Shows why moving from molecule to medicine is difficult.

See on Amazon

Browse more on Amazon:Deep MedicineThe Drug HuntersThe AI Revolution in Medicine

As an Amazon Associate I earn from qualifying purchases.

eBay marketplace picks

Marketplace Samples

Example marketplace items related to this page. Use the search link to explore similar finds on eBay.

Shop location

UsingUSA USA USAUKCanadaAustraliaIreland

Example eBay listing

Soviet Poster ORIGINAL First Aid kit Medical AI-2 NBC Chernobyl stalker USSR

SearcheBay.co.uk: medical AI poster

Browse similar oneBay.co.uk

Example eBay listing

Soviet Russian Poster First Aid Medical Kit AI Army Forces, Military AV 18x24"

SearcheBay.co.uk: medical AI poster

Browse similar oneBay.co.uk

Example eBay listing

Soviet Russian Poster First Aid Individual Medical Kit AI Military 24x36"

SearcheBay.co.uk: medical AI poster

Browse similar oneBay.co.uk

Example eBay listing

Soviet Russian Poster First Aid Individual Medical Kit AI Military Army 18x24"

SearcheBay.co.uk: medical AI poster

Browse similar oneBay.co.uk

Browse more oneBay.co.uk

Example items shown for inspiration; availability and pricing can change. Branchoria may earn a commission if you purchase through outbound eBay links.

Endnotes

1. Source: nature.com
Link:https://www.nature.com/articles/s41586-024-07487-w

2. Source: insilico.com
Link:https://insilico.com/casestudy

3. Source: insilico.com
Title: first phase2
Link:https://insilico.com/blog/first_phase2

4. Source: nature.com
Link:https://www.nature.com/articles/d41586-025-00111-5

5. Source: cell.com
Link:https://www.cell.com/the-innovation/fulltext/S2666-6758%2824%2900123-1

6. Source: deepceutix.com
Title: Deep Ceutix Your AI Can Design a Molecule
Link:https://deepceutix.com/insights/ai-formulation-gap

7. Source: reuters.com
Title: Google-backed Isomorphic Labs delays clinical trial timeline
Link:https://www.reuters.com/business/healthcare-pharmaceuticals/google-backed-ai-drug-discovery-startup-isomorphic-labs-delays-clinical-trial-2026-01-20/

8. Source: reuters.com
Link:https://www.reuters.com/legal/litigation/google-backed-isomorphic-raises-21-billion-scale-ai-driven-drug-discovery-2026-05-12/

9. Source: arxiv.org
Link:https://arxiv.org/abs/2503.02781

Published: March 4, 2025

10. Source: nature.com
Link:https://www.nature.com/articles/d41586-024-01383-z

11. Source: nature.com
Link:https://www.nature.com/articles/d41586-024-01463-0

Published: May 2024

12. Source: pubmed.ncbi.nlm.nih.gov
Link:https://pubmed.ncbi.nlm.nih.gov/40461817/

13. Source: reruption.com
Link:https://reruption.com/en/knowledge/industry-cases/insilico-medicine-ai-drug-to-phase-ii-in-30-months

Additional References

14. Source: researchgate.net
Link:https://www.researchgate.net/publication/392366960_A_generative_AI-discovered_TNIK_inhibitor_for_idiopathic_pulmonary_fibrosis_a_randomized_phase_2a_trial

15. Source: prescouter.com
Link:https://www.prescouter.com/2024/05/alphafold-3/

16. Source: semanticscholar.org
Link:https://www.semanticscholar.org/paper/AI-enabled-drug-discovery-reaches-clinical-Zitnik/bbca07d1a74b4c5311411ff41d354b274d944300

17. Source: loonbio.com
Link:https://loonbio.com/reflections/ai-drug-discoverys-60-billion-reality-check-hype-failures-and-the-market-access-blindspot

18. Source: science.org
Title: limits access deepmind s new protein program trigger backlash
Link:https://www.science.org/content/article/limits-access-deepmind-s-new-protein-program-trigger-backlash

Published: May 2024

19. Source: forbes.com
Link:https://www.forbes.com/sites/calumchace/2023/06/30/the-first-ai-developed-drug-reaches-phase-2-clinical-trials-with-alex-zhavoronkov/

20. Source: isomorphiclabs.com
Title: alphafold 3 predicts the structure and interactions of all of lifes molecules
Link:https://www.isomorphiclabs.com/articles/alphafold-3-predicts-the-structure-and-interactions-of-all-of-lifes-molecules

Published: May 2024

21. Source: genengnews.com
Title: alphafold 3 angst limited accessibility stirs outcry from researchers
Link:https://www.genengnews.com/topics/artificial-intelligence/alphafold-3-angst-limited-accessibility-stirs-outcry-from-researchers/

22. Source: linkedin.com
Title: ai enabled clinical trials 2025 evidence engineering framework z3xde
Link:https://www.linkedin.com/pulse/ai-enabled-clinical-trials-2025-evidence-engineering-framework-z3xde

23. Source: empowerswiss.org
Link:https://empowerswiss.org/en/blog/ai-drug-discovery-breakthroughs-move-from-promise-to-proof-in-2025