Pharma's New Lab Partner: How AI Is Compressing Drug Discovery

AI will not magically fix drug discovery. But it is already compressing some of the slowest and most expensive steps in the pipeline — and the evidence is moving from theoretical to clinical.

Bringing a new drug to market takes an average of 10 to 15 years and costs over $2 billion when you account for the cost of failures. The failure rate is staggering: roughly 90 percent of drugs that enter clinical trials never reach patients. The pipeline is long, expensive, and punishing — and it has resisted optimization for decades.

That is starting to change. Not because AI has replaced the wet lab or eliminated the need for clinical trials, but because it is compressing specific stages of the pipeline where time and money are wasted on slow, iterative processes that computation can accelerate. Target identification. Molecular design. ADMET prediction. Trial site selection. Patient recruitment. Regulatory document drafting. None of these steps disappear, but each one is getting faster — and the compression compounds.

The numbers are getting real. Over 173 AI-originated drug programs are now in clinical development — up from just 67 in 2023 and 3 in 2016. Takeda signed a multi-year AI drug-discovery deal with Iambic Therapeutics worth more than $1.7 billion in biobucks, with Iambic claiming its AI-plus-automation approach can reduce preclinical development time from roughly six years to under two. Eli Lilly has a $1.75 billion agreement with Isomorphic Labs, Alphabet's AI drug discovery subsidiary. Eli Lilly and NVIDIA jointly invested up to $1 billion in an AI co-innovation lab for drug discovery. VCs invested $2.7 billion in AI drug development firms through the first three quarters of 2025 alone.

The question is no longer whether AI belongs in drug discovery. It is which parts of the pipeline AI compresses most, and whether faster timelines translate into better drugs — or just faster failures.

Why Pharma Timelines Are So Punishing

The traditional drug discovery pipeline moves through distinct phases, each with its own timeline and failure rate.

Target identification and validation — finding the biological mechanism that a drug should act on — takes two to three years. Researchers sift through genomic data, disease biology, and published literature to identify promising targets, then validate that modulating those targets actually affects the disease.

Lead discovery and optimization — designing molecules that interact with the target — takes another two to three years. Medicinal chemists synthesize and test thousands of compounds, iterating through cycles of design, synthesis, and testing to find molecules with the right combination of potency, selectivity, and drug-like properties.

Preclinical development — testing in animal models, assessing toxicity, and preparing for human trials — adds one to two years. Most drug candidates fail here because they prove toxic, are metabolized too quickly, or do not reach the target tissue at sufficient concentrations.

Clinical trials — Phase I (safety), Phase II (efficacy), and Phase III (large-scale confirmation) — take six to eight years combined. Phase II is where the majority of failures occur, with success rates historically around 29 to 40 percent. Phase III trials are the most expensive, often costing hundreds of millions of dollars for a single program.

Regulatory review — submitting the New Drug Application and awaiting FDA approval — adds another one to two years.

The total timeline, from target identification to approval, averages 12 to 15 years. The total cost, including the cost of failed programs, exceeds $2 billion per approved drug. And the overall probability of success from Phase I to approval is roughly 10 percent. Industry estimates suggest AI could reduce R&D expenses by 30 to 40 percent, shortening timelines by one to four years and cutting clinical trial costs by up to 50 percent.

These numbers explain why pharma companies are willing to bet billions on AI: even a modest compression at any stage of the pipeline translates into enormous value — saved years, saved capital, and potentially more drugs reaching patients.

Where AI Is Actually Speeding the Pipeline

AI is not compressing the drug discovery pipeline uniformly. It is having the most impact at specific stages where the work is data-intensive, iterative, and amenable to pattern recognition.

Target identification. AI systems can analyze vast genomic, proteomic, and clinical datasets to identify disease targets that would take human researchers years to find. Machine learning models trained on multi-omics data can predict which proteins are most likely to be druggable and which genetic variants are associated with disease. Companies like Variant Bio are building agentic AI platforms on unique genetic and biological data from populations underrepresented in traditional genomic databases. This does not eliminate the need for biological validation, but it narrows the search space dramatically.

Molecular design. Generative chemistry — AI systems that design novel molecules with desired properties — is the most visible application of AI in drug discovery. Iambic's generative model NeuralPLexer, which predicts protein-ligand complexes, is central to the Takeda deal. Instead of synthesizing and testing thousands of compounds sequentially, AI models can generate candidate molecules computationally, predict their binding affinity, selectivity, and drug-like properties, and prioritize the most promising candidates for synthesis. This compresses the design-make-test cycle from months to days.

ADMET prediction. Absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties determine whether a molecule that works in a test tube will work in a living organism. Predicting these properties computationally — rather than discovering them through expensive and time-consuming animal studies — allows researchers to eliminate poor candidates earlier. AI-predicted ADMET properties now correlate well enough with experimental results that computational screening is identifying better candidates, not just more candidates.

Clinical trial optimization. AI is improving trial design, site selection, and patient recruitment. Machine learning models can identify optimal trial sites based on patient demographics, disease prevalence, and historical enrollment rates. AI-powered patient matching can identify eligible participants from electronic health records, reducing the months-long recruitment process. Adaptive trial designs, informed by real-time data analysis, allow protocols to be modified mid-trial based on interim results.

Regulatory document preparation. Drug companies must submit massive documentation packages to regulatory agencies — clinical study reports, safety summaries, integrated efficacy analyses. AI tools now draft sections of these documents, pulling from structured clinical data and formatting to regulatory standards. This does not eliminate human review, but it compresses what was previously weeks of manual writing into days.

Molecule Design, Wet Labs, and Simulation

The relationship between AI and wet-lab biology is often misunderstood. AI does not replace experiments. It makes experiments more targeted.

Generative chemistry platforms — built by companies like Recursion, Insilico Medicine, and Isomorphic Labs — can propose thousands of novel molecular structures optimized for specific targets. Recursion merged with Exscientia in 2025, integrating phenomic screening with automated precision chemistry into a full end-to-end platform. The combined company has one of the world's largest biological datasets, built through automated high-throughput experimentation, and a $1.5 billion partnership with Bayer. Their pipeline includes REC-394 (C. difficile, Phase 2 data expected Q1 2026) and REC-1245 (solid tumors, Phase 1 dose-escalation data expected H1 2026).

But a computationally designed molecule still needs to be synthesized in a physical lab and tested in biological assays. The AI compresses the design phase — generating in hours what would take chemists weeks of iterative design — but the synthesis and testing still happen in the real world.

Where the compression multiplies is in the integration of AI with automated lab platforms. Robotic systems can synthesize, purify, and test compounds at scale, feeding results back to the AI models that generated them. This closed-loop system — design, make, test, learn, redesign — operates faster than any human team. NVIDIA's BioNeMo platform provides AI-accelerated computing for molecular dynamics simulations, predicting how a molecule will behave in a biological environment. These simulations do not replace animal studies or clinical trials, but they provide data that helps researchers make better decisions earlier.

The net effect is not that the wet lab disappears. It is that the wet lab does fewer, smarter experiments — guided by AI predictions that increase the probability of each experiment succeeding.

The First AI-Designed Drug to Show Clinical Efficacy

The most significant clinical milestone for AI drug discovery came from Insilico Medicine's ISM001-055, now known by its generic name rentosertib. It is a first-in-class TNIK (TRAF2 and NCK-interacting kinase) inhibitor designed entirely by AI — target identified by AI, molecule generated by AI, and optimized through Insilico's end-to-end platform.

In a Phase IIa randomized, double-blind, placebo-controlled trial conducted across 22 sites in China, 71 patients with idiopathic pulmonary fibrosis (IPF) were enrolled. The results, published in Nature Medicine in June 2025, showed that the 60mg dose group achieved a mean forced vital capacity (FVC) improvement of 3.05 percent from baseline at Week 12 — compared to a 1.84 percent decline in the placebo group. The drug was well-tolerated with a safety profile comparable to placebo.

These results matter for two reasons. First, IPF is a progressive, fatal lung disease with limited treatment options — any drug that stabilizes or improves lung function is clinically meaningful. Second, rentosertib is the first AI-discovered and AI-designed drug to demonstrate efficacy in a randomized controlled trial. The molecule went from target identification to Phase IIa results in under four years — a timeline that would have taken a decade or more through traditional drug discovery.

Insilico has stated that it plans to advance rentosertib into a larger Phase IIb/III trial. If those results hold, it could become the first AI-designed drug to reach the market — a milestone that would validate the entire AI drug discovery thesis.

The caveat is important: Phase IIa is still early. Many drugs that show promising Phase IIa results fail in larger trials. The 71-patient sample size is small. But the signal — FVC improvement versus placebo decline — is clear enough that it passed peer review in one of the world's top medical journals. This is no longer theoretical.

Clinical Trials and Regulatory Paperwork as Hidden AI Wins

The most visible applications of AI in drug discovery — molecular design, protein structure prediction — get the most attention. But some of the largest time and cost savings come from less glamorous applications: clinical trial logistics and regulatory documentation.

Patient recruitment is one of the biggest bottlenecks in clinical trials. Finding patients who meet the inclusion and exclusion criteria, convincing them to enroll, and retaining them through the trial duration accounts for a significant portion of trial timelines and costs. AI-powered matching systems can scan electronic health records across health systems to identify eligible patients, predict which patients are most likely to enroll and complete the trial, and automate outreach. Companies like Tempus and Deep 6 AI specialize in this matching. Drugmakers report shaving weeks off labor-intensive recruitment processes.

Site selection determines where the trial is conducted. AI models can analyze historical enrollment data, patient demographics, investigator performance, and geographic accessibility to recommend optimal sites — reducing the risk of slow enrollment that can delay trials by months or years.

Adaptive trial design uses AI to analyze interim data and adjust the trial protocol in real time. This might mean increasing enrollment in a subgroup that shows particular benefit, dropping a dosage arm that is not working, or modifying endpoints based on observed outcomes. The FDA has issued guidance encouraging adaptive designs, and AI makes them practical at scale. Adaptive designs can reduce trial duration and cost while maintaining statistical rigor.

Regulatory drafting is the most unglamorous AI win — and one of the most practical. Preparing a New Drug Application for the FDA involves assembling tens of thousands of pages of clinical data, safety analyses, manufacturing documentation, and labeling information. AI tools can draft sections of these documents, cross-reference data tables, flag inconsistencies, and format submissions to regulatory standards. This does not change the regulatory requirements, but it compresses the labor of meeting them.

Why Data Advantage Matters More Than Model Hype

In pharma AI, the competitive advantage is not the model. It is the data.

Large language models and generative chemistry platforms are increasingly commoditized — the architectures are well-known, and the computational resources to train them are available to any well-funded company. Over 210 AI drug discovery companies now claim to offer AI-based services, platforms, and tools. What cannot be commoditized is proprietary biological data: experimental results from millions of assays, clinical trial outcomes, patient genomic data, and multi-omics datasets that capture the complexity of human disease.

The implication is that the winners in pharma AI will not necessarily be the companies with the most sophisticated models. They will be the companies with the most comprehensive, high-quality biological data — and the infrastructure to continuously generate more of it. This is why partnerships between traditional pharma companies (which have decades of clinical data) and AI-native biotechs (which have the computational platforms) are becoming the dominant model.

The Rise of AI-Native Biotech Partnerships

The Takeda-Iambic deal — worth more than $1.7 billion in potential milestone payments plus royalties — represents a model proliferating across the industry. The deal covers oncology, gastrointestinal, and inflammation therapeutic areas, with Takeda leveraging Iambic's software and drug discovery assets to advance small molecule programs. Traditional pharma companies are not trying to build AI capabilities in-house. They are partnering with AI-native biotechs that have purpose-built platforms, proprietary data, and computational infrastructure.

These partnerships are structured to compress timelines. Iambic claims its platform can reduce preclinical development from six years to under two. Eli Lilly's $1.75 billion deal with Isomorphic Labs targets novel drug candidates using AlphaFold-derived structural biology. Over $10 billion has been invested in AI drug discovery in the last five years, and the market is projected to grow from $2.9 billion in 2025 to $5.1 billion in 2026.

The partnership model works because drug discovery requires both computational capability and biological expertise. An AI platform that can design molecules but cannot navigate regulatory requirements, manage manufacturing, or run clinical trials needs a pharma partner. A pharma company that can run trials and commercialize drugs but is limited by the speed of traditional medicinal chemistry needs an AI partner.

What Investors and Pharma Leaders Are Really Betting On

The investment thesis for pharma AI is not that AI will discover drugs that humans cannot. It is that AI will discover drugs faster and cheaper — compressing timelines, reducing failure rates, and allowing more programs to be pursued simultaneously.

This is a capital efficiency argument, not a capability argument. If AI can reduce the cost of bringing a drug to market from $2 billion to $1 billion — or reduce the timeline from 15 years to 8 — the economics of drug development change fundamentally. More drugs become commercially viable. Smaller patient populations become addressable. Rare diseases that were too expensive to pursue become tractable.

The evidence is building. AI-discovered molecules show an 80 to 90 percent success rate in Phase I trials — substantially higher than the historical 40 to 65 percent for traditionally discovered drugs. Phase II success rates are approximately 40 percent, comparable to traditional methods. With rentosertib's Phase IIa results now published in Nature Medicine, there is peer-reviewed evidence that an AI-designed molecule can demonstrate clinical efficacy in a randomized trial. No AI-discovered drug has achieved FDA approval as of early 2026, but the first approval is projected for 2026 to 2027 with approximately 60 percent probability.

What investors are watching most closely is not the number of AI-discovered drugs in trials, but the hit rates. If AI compression comes at the cost of lower success rates — faster but less reliable — the economics do not work. If AI compression maintains or improves success rates — faster and more reliable — the value creation is enormous.

Why "Faster" Only Matters If It Still Improves Hit Rates

Speed without accuracy is waste at scale.

The fundamental risk of AI-compressed drug discovery is that faster timelines could produce more candidates that fail in later stages — shifting costs from early discovery to expensive late-stage clinical trials. If AI makes it easy to generate thousands of molecular candidates but the quality of those candidates is no better than traditional approaches, the pipeline just moves faster while failing at the same rate.

The early evidence is encouraging but not yet conclusive. The 80 to 90 percent Phase I success rate is promising, but the sample sizes are still relatively small. Rentosertib's Phase IIa results — 3.05 percent FVC improvement versus 1.84 percent decline on placebo, across 71 patients — represent the strongest clinical signal for an AI-designed drug to date, but the leap from Phase IIa to registration-quality Phase III results remains the hardest step in drug development. And the most definitive test — whether AI-discovered drugs succeed in Phase III trials and reach patients — is still years away for most programs.

The pharma industry has seen waves of technology promise before: combinatorial chemistry, high-throughput screening, genomics-guided target selection. Each delivered value, but none transformed the economics as dramatically as proponents predicted. AI may be different — the breadth of its application across the pipeline, the speed of improvement in the underlying models, and the integration with automated experimentation all suggest a more fundamental shift. Multiple AI-discovered drug programs have been shelved after Phase II failures, including some from Recursion that were deprioritized in 2025. These failures are statistically expected but serve as a reminder that AI does not eliminate biology's inherent unpredictability.

What is clear today is that AI is not replacing drug discovery. It is compressing it — making specific stages faster, cheaper, and more targeted. The companies that will benefit most are the ones that combine AI capability with biological rigor, that use computation to guide experimentation rather than replace it, and that measure success not by the speed of their pipeline but by the quality of what comes out the other end.

---

At AIReady.fit↗, we help professionals and teams build productive AI workflows. Our AI Foundations track covers how AI is reshaping industries from pharma to finance — practical skills for anyone adapting to the next generation of workplace AI tools.