1. Where Does AI in Drug Discovery Stand in 2025?
Between 2024 and 2025, AI in drug discovery has quietly moved from the “proof-of-concept” phase into a stage where AI-designed molecules are showing signals in human trials. At the same time, it has become equally clear that AI does not magically remove the risks of clinical development.
Public analyses suggest that by 2024, around 30 AI-designed or AI-enabled drugs from eight leading companies had entered human trials, with indications that early-phase success rates might be higher than historical norms. However, as of November 2025, no AI-designed drug has yet been approved, and Phase 3 entries remain rare.
In other words, the current reality is that AI in drug discovery is credible enough to deliver promising early-clinical signals, but the distance to final approval is still substantial.
2. Flagship Case ①: Insilico Medicine’s AI-Designed IPF Drug rentosertib
2-1. An AI-Discovered Target and AI-Designed Molecule Reaching Phase 2
One of the most symbolic success stories in 2024–2025 is Insilico Medicine’s small-molecule drug rentosertib (formerly ISM001-055 / INS018_055) for idiopathic pulmonary fibrosis (IPF). In this project:
- The target, TNIK (Traf2- and NCK-interacting kinase), was identified using omics data and generative AI.
- The molecule itself was designed by generative AI as a first-in-class TNIK inhibitor.
Insilico reported positive topline results from a Phase 2a trial in IPF patients in late 2024, and detailed clinical data were subsequently published in 2025. The drug showed a favorable safety profile and promising signals in lung-function endpoints, making rentosertib one of the first AI-designed molecules to clear a Phase 2a hurdle with encouraging results.
2-2. Why This Is Important, Yet Not the End of the Story
At the same time, rentosertib is still only in Phase 2. Long-term safety, comparative effectiveness versus existing IPF therapies, and the outcome of later-stage trials remain unknown. IPF itself is a challenging disease where many promising mechanisms have failed late in development.
The case illustrates that AI can contribute to discovering novel targets and compounds that survive into mid-stage clinical development. It does not, however, change the fundamental reality that late-stage clinical success is determined by complex biology, trial design, and patient heterogeneity, all of which remain only partially predictable.
3. Flagship Case ②: Isomorphic Labs, Novartis, and Lilly
3-1. Multi-Billion-Dollar Collaborations with Big Pharma
Isomorphic Labs, a DeepMind spin-out under Alphabet, has become a key symbol of “AI-first” drug discovery. In 2024, the company announced multi-target collaborations with Novartis and Eli Lilly that could be worth nearly $3 billion in total potential value, excluding future royalties. These deals signal a qualitatively different level of commitment by big pharma to AI-centric platforms.
In February 2025, Isomorphic Labs and Novartis further expanded their collaboration, adding several new research programs on the same financial terms as the original agreement—an indication that year-one progress was judged sufficiently encouraging by both parties.
3-2. AI-Designed In-House Programs Approaching Clinical Trials
Beyond partnerships, Isomorphic Labs is also advancing its own AI-designed oncology and cardiometabolic programs towards first-in-human studies. Company executives have repeatedly stated that they are “getting very close” to initiating trials of their in-house AI-designed drugs, backed by substantial external funding raised in 2025.
While clinical data from these programs are not yet available, they are likely to form the next wave of AI-designed molecules entering Phase 1 and 2 over the coming years.
4. Flagship Case ③: “AI Supercomputers” – Recursion × NVIDIA, Lilly × NVIDIA
4-1. Recursion’s BioHive-2: An AI Supercomputer Built for Drug Discovery
Recursion is known for generating vast quantities of high-content imaging and omics data in its own labs and using AI to learn representations of biology. In partnership with NVIDIA, Recursion built BioHive-2, which has been described as one of the largest AI supercomputers in the pharmaceutical industry.
BioHive-2 is designed to process tens of petabytes of biological and chemical data and to support training large-scale foundation models. Recursion’s leadership has emphasized that, over time, AI models will commoditize, and the real moat will lie in proprietary experimental data plus the infrastructure to exploit it.
4-2. Eli Lilly’s Supercomputer with NVIDIA
In October 2025, Eli Lilly announced a collaboration with NVIDIA to build its own DGX SuperPOD-based AI supercomputer for drug discovery and development. The platform will feed into Lilly’s TuneLab, a federated AI/ML environment intended to run millions of virtual experiments and also support manufacturing, imaging, and enterprise AI agents.
Lilly’s Chief AI Officer has described the shift as moving from “AI as a tool” to “AI as a collaborative scientific partner”, highlighting that the competition is no longer just about algorithms but about integrated systems of data, compute, and domain expertise.
5. Flagship Case ④: Emerging Startups and Better Hit Rates
2025 also saw rising attention on newer AI-native startups. One example is Chai Discovery, backed by major AI-focused investors. In 2025, the company raised around $70 million at a valuation of roughly $550 million, citing internal benchmarks in which its latest model reportedly achieved successful binding hits against roughly one in six proteins, compared with traditional hit rates on the order of one in a thousand.
These hit-rate numbers are still early and come from internal experiments rather than clinical trial results. Yet they support the broader narrative that AI can substantially improve the efficiency of hit finding and early validation in small molecules, antibodies, and beyond.
6. Failures and What Has Not Yet Been Achieved
6-1. BEN-2293: An AI-Derived Molecule Failing in Phase 2
On the other side of the ledger, the atopic dermatitis drug BEN-2293, often cited as an AI-derived molecule, failed to show efficacy in a Phase 2a trial, leading to the discontinuation of development.
This case is important because it reminds us that AI-designed molecules are not exempt from the usual attrition of drug development. What matters is not that every AI-derived drug succeeds, but rather:
- Whether AI can raise early-phase success rates and shorten cycle times at a portfolio level.
- Whether failures can be turned into faster learning cycles through better data capture and model updates.
6-2. No AI-Designed Drug Has Been Approved Yet
As of November 22, 2025, no AI-designed drug has completed Phase 3 and gained regulatory approval. A number of programs are advancing through Phase 1 and 2, but the final verdict at the level of approved drugs is still pending.
Thus, AI in drug discovery has clearly advanced to the point where it can deliver promising early-clinical assets, but it is too early to claim a proven impact on the probability of approval or on long-term safety.
7. The 2025 Reality: What We Can Say with Some Confidence
Looking across these cases, the 2025 reality for AI in drug discovery can be summarized as follows:
- (1) For small molecules, AI is now delivering practical value in target identification and lead optimization. Rentosertib and similar programs show that AI-designed molecules can survive into mid-stage development with encouraging signals.
- (2) For antibodies, RNA, cell, and gene therapies, AI is mostly improving exploration efficiency and risk stratification. It helps in sequence design, delivery, and manufacturing conditions, but does not yet guarantee clinical success.
- (3) The real competitive edge lies in the combination of data, infrastructure, and deep modality expertise. Recursion, Lilly, and others highlight a shift toward “AI supercomputers plus proprietary data” rather than algorithms alone.
- (4) Failures are now visible and informative. Cases like BEN-2293 demonstrate that AI-designed drugs fail for familiar reasons—biology is complex—and that the industry must learn from these failures systematically.
My Thoughts and Future Outlook
If we try to characterize AI in drug discovery in 2025, it looks like a phase where a few highly visible successes and a growing list of quiet failures are finally appearing together. Rentosertib shows that AI-originated targets and molecules can generate meaningful signals in human trials, while BEN-2293 reminds us that AI does not abolish clinical risk. In that sense, the question is no longer “Can AI ever work in drug discovery?” but rather “Under what conditions does it work well enough to matter at scale?”.
Over the next 5–10 years, the key will be to move beyond case-by-case storytelling and quantify where (which modalities), in which phases, and with what level of data and infrastructure AI actually shifts portfolio-level outcomes. AI has the potential to raise early-phase success rates and accelerate learning cycles, but its realized impact will depend heavily on how thoughtfully organizations invest in data design, infrastructure, and cross-functional teams. Those who treat AI as a long-term systems investment rather than a short-term hype wave are likely to see the most durable benefits.
As an extra chapter, this snapshot is meant to capture “where we stand as of November 22, 2025” rather than to provide final answers. When we revisit the same question a few years from now, some of the programs discussed here may have succeeded or failed in late-stage trials, and new players may have emerged. That trajectory itself—who learns fastest and how they structure their AI–biology–data loop—will probably tell us more about the true power of AI in drug discovery than any single success story.
This article has been edited by the Morningglorysciences team.
Related Articles
Comments