30 Months, $6M: How Insilico Medicine's AI Rewrites Drug Discovery

The average new drug costs $2.6 billion and takes 12 years to reach patients — and 90% of candidates fail before approval. In 2025, a drug designed entirely by AI, from target identification to final molecule, completed Phase IIa trials with statistically significant results, was published in Nature Medicine, and cost approximately $6 million to discover and design. The program took 30 months from standing start to clinical evidence. That is not a research paper. That is a working drug, in patients, with published data.

Idiopathic pulmonary fibrosis (IPF) kills roughly 40,000 Americans per year and has almost no effective treatment options. Traditional drug discovery spent decades looking for the right target with limited success. Insilico Medicine's AI system, Pharma.AI, identified a completely novel biological target — TNIK — that human researchers had largely deprioritized, then designed the molecule that would become Rentosertib (ISM001-055). Both steps, the ones that traditionally consume the most time and capital, were fully automated.

Insilico is not a speculative AI startup. The company currently serves 13 of the top 20 global pharmaceutical companies. Its software revenue grew 23.8% year-over-year in 2025, with its subscription customer base growing 18.3%. Its internal pipeline contains 31 active drug programs — 5 of them now in clinical stages — all initiated and designed using AI. The Rentosertib results are not an anomaly. They are the proof that the model works at scale.

The $2.6 Billion Problem That Has Paralysed Drug Innovation for Decades

The mathematics of traditional drug discovery are structurally broken. To bring a single drug to market, a pharma company must fund dozens of programs through discovery, preclinical, Phase I, Phase II, and Phase III trials — with an industry-average failure rate above 90% at each gate. The result: $2.6 billion in R&D expenditure per approved drug (DiMasi et al., Tufts CSDD), most of it spent on programs that never reach a patient.

The bottleneck is not effort or resources. It is information processing. Identifying the right biological target requires integrating genomic, proteomic, transcriptomic, and clinical data across thousands of disease variants simultaneously. Designing an optimal therapeutic molecule means exploring a chemical space of approximately 10⁶⁰ possible compounds. No human team can do this at speed or at scale. This is precisely the computational problem that large language models and generative AI were built to solve — and Insilico Medicine built an end-to-end platform to do exactly that.

Companies that adopted AI-driven discovery between 2021 and 2023 are now reporting Phase II data. Companies that did not are still running the same $150M programs on 8-year timelines. The performance gap is now measurable, documented, and widening every quarter.

How Pharma.AI Works: Three Engines, One Closed Loop

Pharma.AI integrates three specialized AI engines into a continuous discovery workflow. Each engine solves a distinct bottleneck in the traditional pipeline.

PandaOmics is the target identification engine. It ingests multi-omics data — genomic, transcriptomic, proteomic, and clinical datasets — and applies machine learning to surface biological targets that conventional analysis overlooks. For the Rentosertib program, PandaOmics analyzed patterns across hundreds of human tissue samples and IPF disease cohorts and identified TNIK (TRAF2 and NCK-interacting protein kinase) as a high-confidence novel target. Human researchers had deprioritized TNIK; the AI detected a network-level signal invisible in siloed datasets. This is the step that cost traditional pharma teams years of manual literature review and hypothesis testing.

Chemistry42 is the generative molecular design engine. Once a validated target exists, Chemistry42 uses a generative model trained on millions of known molecular structures to design novel drug candidates optimized simultaneously for potency, selectivity, bioavailability, and safety profile. For ISM001-055, Chemistry42 generated, screened, and ranked candidate molecules with approximately 60–200 molecules synthesized and tested across the program — compared to thousands in conventional medicinal chemistry campaigns that last 2–3 years.

LabClaw (deployed 2026) closes the feedback loop between computational prediction and physical wet-lab validation. The system automatically dispatches experimental tasks to laboratory equipment — cell culture, high-throughput screening, next-generation sequencing — creating a self-improving cycle in which experimental results feed directly back into the AI's next prediction iteration. This is the architectural step that transforms Pharma.AI from an analysis tool into an autonomous discovery system capable of learning from its own experiments.

Rentosertib: The Drug That Proved the Model in Front of the World

Rentosertib entered clinical trials in 2023 and completed a Phase IIa randomized, double-blind, placebo-controlled trial across 21 sites in China, enrolling 71 patients with confirmed IPF. Results were presented at the American Thoracic Society (ATS) 2025 International Conference and simultaneously published in Nature Medicine in June 2025 — one of the highest-impact peer-reviewed journals in medicine.

The primary endpoint — safety and tolerability across all dose levels — was met. Secondary efficacy analysis showed that patients receiving 60 mg once daily experienced a mean improvement of +98.4 mL in forced vital capacity (FVC) over 12 weeks, versus a mean decline of −20.3 mL in the placebo group. FVC is the primary functional lung measure in IPF — a disease in which lung capacity typically declines relentlessly. Exploratory biomarker analyses independently validated TNIK as the mechanistic target, confirming that the AI's original identification was biologically correct.

The Insilico team has been transparent about what this means: a completely novel drug target, identified by AI in a disease area where human researchers had spent decades without identifying a high-confidence mechanism, validated by a clinical trial published in Nature Medicine. The drug is now advancing toward Phase IIb/III.

The Numbers Reshaping Pharma Investment Decisions in 2026

Rentosertib is the most documented proof of concept, but the industry-level data tells a larger structural story. According to Axis Intelligence's 2026 analysis, 173+ AI-originated drug programs are currently in clinical development — up from approximately 24 programs in late 2023. The market for AI in drug discovery is valued at $5 billion in 2026 and projected to reach $12.56 billion by 2034 at a CAGR of approximately 12.2% (Grand View Research).

Eli Lilly formalized its commitment with a $2.75 billion multi-year partnership with Insilico Medicine — one of the largest AI-pharma deals on record — signalling that AI-native discovery is no longer an experimental initiative but strategic infrastructure for the world's most valuable drug pipelines. Novo Nordisk, Pfizer, and Sanofi have made comparable commitments with competing AI platforms. The 81% of pharma companies now deploying AI are not running proofs of concept — they are funding programs with clinical endpoints, regulatory strategies, and commercial projections.

The financial logic is structural. AI reduces preclinical development costs by 30–70% by eliminating the failed synthesis cycles and compressed timelines that have historically consumed the majority of early-stage R&D budgets. For an organization running 20 active programs, the compound effect is hundreds of millions in recovered capital that can be reinvested into additional pipeline. The companies that adopted this infrastructure in 2022–2023 are now two Phase cycles ahead of those that did not.

The Technology Stack Behind the Breakthrough

The complete Insilico platform runs on enterprise software subscriptions. No proprietary hardware is required beyond standard computational infrastructure.

Pharma.AI is currently accessed via enterprise partnership and subscription. Insilico serves 13 of the top 20 global pharmaceutical companies as software customers. Academic and smaller biotech access is available via external collaboration programs. [REQUIERE VERIFICACIÓN: current academic licensing terms — check insilico.com/enterprise for latest.]

Who Should Be Paying Attention Right Now

The primary audience for this case study is drug discovery executives, computational biology leads, R&D strategy directors, and CTO-level leadership at pharma and biotech organizations with active preclinical pipelines. If your organization is currently funding programs that will take 8+ years to generate Phase II data, the Insilico model represents a direct competitive threat from any competitor or partner that adopts AI-native discovery first.

This is equally material for biotech investors and institutional capital allocators. The valuation frameworks for AI-native drug companies differ fundamentally from traditional pharma because the cost-per-program and time-to-clinical-signal have structurally changed. An AI-first biotech can generate Phase II evidence in 30 months — the same period a traditional competitor spends in lead optimization. This compresses the capital efficiency equation in ways that are now visible in public market valuations and deal multiples.

This case study is not directly applicable to medical device manufacturers, CROs without internal discovery assets, or organizations without active small-molecule or biologics programs.

Frequently Asked Questions

Is Rentosertib the only AI-discovered drug with Phase II clinical results?

No — Rentosertib is among the most prominent because the full discovery narrative (AI target identification + AI molecule design + positive Phase IIa results) is documented in peer-reviewed publications. But 173+ AI-originated drug programs are in clinical development as of 2026. Recursion Pharmaceuticals, Exscientia, and Absci all have programs in Phase I/II. Insilico's Rentosertib stands out because every step — target identification, molecule design, and clinical validation — was AI-driven end to end, and all three steps have been independently validated in the scientific literature.

Does AI replace medicinal chemists and biologists in this model?

No. The Pharma.AI model augments scientific teams rather than replacing them. Medicinal chemists validate AI-generated candidates and design synthesis routes. Biologists design and execute assays that feed results back into the system. Clinicians design and run trials. LabClaw automates the physical dispatching of experimental tasks, but experimental design and scientific interpretation remain human-led. The efficiency gain comes from eliminating the iterative design–synthesize–test–fail–repeat cycles that historically consumed years of chemist time on unproductive molecular scaffolds.

How long does it take to implement Pharma.AI in an existing drug discovery organization?

Enterprise onboarding typically involves integration with existing data systems — genomics databases, existing compound libraries, EHR data — followed by a supervised discovery sprint on an initial program. Time to first AI-generated target shortlist is typically measured in weeks to a few months, depending on data availability and quality. Full platform deployment including LabClaw wet-lab integration is a multi-quarter implementation. [DATO A VERIFICAR — source: insilico.com/enterprise]

The trajectory is unambiguous: AI-native drug discovery is compressing decades of pharmaceutical R&D into months. The organizations building on this infrastructure today are not running pilots — they are filing INDs, generating Phase II data, and signing billion-dollar partnership agreements. The tools are proven. The Phase IIa results are published. The only variable left is timing — and in pharma, timing is measured in competitive advantage.