ResearchPapers, evidence & method
Drugmaker CEOs Push Back on Silicon Valley Claims That AI Will Cure Disease
Pharma leaders, including AstraZeneca's Pascal Soriot and Vertex's Jeff Leiden, say no AI-discovered drug has won FDA approval and that biology, not computing, is the bottleneck. Takeda's computationally designed psoriasis pill now faces an FDA decision due in Q1 2027.
From the WSJ · Technology article of November 4, 2026 and cited sources: Pharma executives say AI-discovered drugs have zero FDA approvals so far. [1] [2]
SOURCES[1] Silicon Valley Promises AI Will Cure Disease. Drugmaker… wsj.com[2] Benzinga: Drugmaker CEOs Say AI Alone Can't Cure Diseases benzinga.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- WSJ · Technology article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 4, 2026 article published
- Covered
- Nov 4 edition
- Evidence
- 9 sources across 9 sites
- 1Silicon Valley Promises AI Will Cure Disease. Drugmaker…wsj.com
- 2Benzinga: Drugmaker CEOs Say AI Alone Can't Cure Diseasesbenzinga.com
- 3Takeda/Business Wire: FDA accepts zasocitinib NDA under…secure.businesswire.com
- 4PubMed: A generative AI-discovered TNIK inhibitor for IPF…pubmed.ncbi.nlm.nih.gov
- 5Artificial Science: AI Drug Discovery's Moment of Truthartificialscience.org
- 6Bio-IT World: Insilico launches Phase III trialbio-itworld.com
Reach: not measured yet
Pharma executives say AI-discovered drugs have zero FDA approvals so far. They describe AI as a co-pilot for choosing targets, optimizing molecules and selecting patients, not as an end-to-end discovery engine. Leiden said about 50 AI companies screened a Vertex dataset of 10,000 compounds and none produced a working drug. [1] [2]
The zero-approvals count may soon change, depending on how AI is defined. Nimbus designed zasocitinib using Schrödinger's physics-based simulation, not generative AI. Takeda's application for it is under FDA Priority Review. Takeda reports that about 70% of nearly 3,000 patients had clear or almost-clear skin at week 16. That figure is company-reported, and the FDA review is still pending. [3] [8]
The leading generative-AI candidate is still early. Insilico's rentosertib, whose target and molecule both came from generative AI, had a 71-patient, 12-week Phase 2a trial with safety as its primary endpoint. Its lung-function gain was exploratory: about 98 mL at the top dose versus a roughly 20 mL decline on placebo. A 52-week Phase III trial began in China in July 2026. [4] [6]
The record so far is mixed. A 2024 analysis found AI-discovered molecules passed Phase I 80–90% of the time, versus a historical 40–65%. In Phase II, only 4 of 10 succeeded, which is in line with industry norms. Experts say standardized, high-quality biological data is the real bottleneck. [9] [7]
Evidence so far shows AI-derived molecules clearing early safety trials more often, but not yet beating industry norms on efficacy.
Read the full assessment
The likely implication is that proprietary biological data and Phase II–III results, not model claims, will decide where AI drug discovery pays off.
Executive brief
Pharma leaders say no drug discovered or designed with AI has yet won FDA approval, and that the main obstacle is biology, not computing power (WSJ feed excerpt). The sharpest example comes from Vertex executive chairman Jeff Leiden. He reportedly said about 50 AI companies screened a Vertex dataset of 10,000 compounds and none of them produced a working drug (Benzinga). The "zero approvals" count could change soon. The FDA is due to decide by Q1 2027 on Takeda's computationally designed psoriasis pill, zasocitinib (Takeda/Business Wire).
What changed and event timeline
AI-discovered drug shows early efficacy
Nature Medicine published a 71-patient Phase 2a trial of Insilico's rentosertib, which blocks a protein called TNIK, in idiopathic pulmonary fibrosis (IPF), a scarring lung disease. Both the target and the molecule came from generative AI ().
Lilly licenses Insilico's platform
Reported terms are $115M upfront and up to $2.75B in milestones for exclusive access to Insilico's Pharma.AI platform ().
Rentosertib enters Phase III
The first patients were dosed in GENESIS-IPF-3, a 52-week, placebo-controlled trial at sites in China ().
FDA accepts Takeda filing
Zasocitinib, an oral TYK2 inhibitor, got Priority Review with an FDA decision due in Q1 2027. Europe's EMA also accepted a filing ().
Biotech veterans push back on the hype
At a journalism panel moderated by a WSJ editor, former Amgen CTO David Reese and Insitro's Mary Rozenman disputed promises that AI will cure "all disease" within a decade ().
WSJ publishes the CEOs' skepticism
AstraZeneca CEO Pascal Soriot and Vertex's Leiden challenged tech-sector claims about AI curing disease (;).
Capabilities and access
- Rentosertib (formerly ISM001-055 / INS018_055) came from Insilico's Pharma.AI platform (Bio-IT World). Lilly now holds an exclusive platform license (Artificial Science).
- Zasocitinib (TAK-279) came from Nimbus Therapeutics using Schrödinger's physics-based computational design, not generative AI. Takeda paid $4B upfront for it (Business Model Analyst).
- None of these AI platforms is publicly available as a model with a version number.
Read the full section
- Rentosertib (formerly ISM001-055 / INS018_055) came from Insilico's Pharma.AI platform (Bio-IT World). Lilly now holds an exclusive platform license (Artificial Science).
- Zasocitinib (TAK-279) came from Nimbus Therapeutics using Schrödinger's physics-based computational design, not generative AI. Takeda paid $4B upfront for it (Business Model Analyst).
- None of these AI platforms is publicly available as a model with a version number.
Technical analysis for researchers and developers
- Rentosertib Phase 2a design: three dose arms plus placebo over 12 weeks. That was an exploratory result in roughly 18 patients per arm (PubMed).
- Data quality is the bottleneck: biology models need standardized, high-quality lab data, and poor data produces errors rather than answers (AHCJ).
- "AI-designed" means different things: generative target-and-molecule design (Insilico) and physics-based simulation (Schrödinger/Nimbus) are different methods.
Read the full section
- Rentosertib Phase 2a design: three dose arms plus placebo over 12 weeks. The primary endpoint was safety (the share of patients with any adverse event), and it was similar across arms: 70.6–83.3%. Lung capacity (FVC) rose 98.4 mL at 60 mg once daily versus a 20.3 mL decline on placebo. That was an exploratory result in roughly 18 patients per arm (PubMed).
- Data quality is the bottleneck: biology models need standardized, high-quality lab data, and poor data produces errors rather than answers (AHCJ).
- "AI-designed" means different things: generative target-and-molecule design (Insilico) and physics-based simulation (Schrödinger/Nimbus) are different methods. Tallies of "AI drugs" depend on which one is counted.
Claims and evidence
- Zero FDA approvals for AI-discovered drugs, cited by executives (WSJ). These figures are uncorroborated.
- Leiden's 50 companies / 10,000 compounds anecdote: reported only as his own statement, with no independent detail (Benzinga).
- Zasocitinib efficacy: about 70% of patients had clear or almost-clear skin at week 16, across nearly 3,000 patients.
Read the full section
- Zero FDA approvals for AI-discovered drugs, cited by executives (WSJ). Secondary tallies put the field at about 175 AI-originated programs in human trials and about $60B invested since 2019 (Artificial Science). These figures are uncorroborated.
- Leiden's 50 companies / 10,000 compounds anecdote: reported only as his own statement, with no independent detail (Benzinga).
- Zasocitinib efficacy: about 70% of patients had clear or almost-clear skin at week 16, across nearly 3,000 patients. This is company-reported; the FDA review is pending (Takeda/Business Wire).
- Rentosertib speed and cost: Insilico says it developed the drug in under 18 months for about $2.6M (Artificial Science). This is a company claim.
Context and prior work
- Track record so far: a 2024 analysis in Drug Discovery Today (Jayatunga et al.) found AI-discovered molecules passed Phase I 80–90% of the time, well above the historical 40–65%.
- Notable failure: Exscientia's DSP-1181, billed as the first AI-designed molecule tested in humans, was dropped after Phase I (Artificial Science).
- The tech-sector claim: the optimistic benchmark is Anthropic CEO Dario Amodei's suggestion that most human disease could be cured within 5–10 years (Benzinga).
Read the full section
- Track record so far: a 2024 analysis in Drug Discovery Today (Jayatunga et al.) found AI-discovered molecules passed Phase I 80–90% of the time, well above the historical 40–65%. In Phase II, 4 of 10 succeeded (about 40%), in line with industry norms (Gary Marklevin summary).
- Notable failure: Exscientia's DSP-1181, billed as the first AI-designed molecule tested in humans, was dropped after Phase I (Artificial Science).
- The tech-sector claim: the optimistic benchmark is Anthropic CEO Dario Amodei's suggestion that most human disease could be cured within 5–10 years (Benzinga).
Limitations, safety and contested findings
- Coverage of the WSJ article: only the feed excerpt and secondary summaries were reviewed, so full quotes and the complete list of executives are unconfirmed.
- Disputed framing: "no AI-designed drug approved" conflicts with describing zasocitinib as AI-crafted (Business Model Analyst). The answer depends on whether physics-based design counts as AI.
- Rentosertib evidence is thin: the Phase 2a was small and short, and its efficacy measure was not the primary endpoint (PubMed).
Read the full section
- Coverage of the WSJ article: only the feed excerpt and secondary summaries were reviewed, so full quotes and the complete list of executives are unconfirmed.
- Disputed framing: "no AI-designed drug approved" conflicts with describing zasocitinib as AI-crafted (Business Model Analyst). The answer depends on whether physics-based design counts as AI.
- Rentosertib evidence is thin: the Phase 2a was small and short, and its efficacy measure was not the primary endpoint (PubMed).
- Baseline: overall clinical success rates are still about 8–10% (AHCJ).
Business and practitioner implications
- Where AI helps today: executives describe it as a co-pilot for picking targets, optimizing molecules and selecting patients, not as an end-to-end discovery engine (Benzinga).
- Where to look for proof: the Phase II efficacy hurdle, not Phase I pass rates, is the test that matters.
- Near-term catalysts: the zasocitinib FDA decision in Q1 2027, and rentosertib's 52-week Phase III readout.
Read the full section
- Where AI helps today: executives describe it as a co-pilot for picking targets, optimizing molecules and selecting patients, not as an end-to-end discovery engine (Benzinga).
- Where to look for proof: the Phase II efficacy hurdle, not Phase I pass rates, is the test that matters.
- Near-term catalysts: the zasocitinib FDA decision in Q1 2027, and rentosertib's 52-week Phase III readout.
- Investment priority: differentiation is likely to come from proprietary, standardized biological data more than from model architecture (AHCJ).
The source trail.
Sources (9)
Silicon Valley Promises AI Will Cure Disease. Drugmaker CEOs Aren’t Buying It.
Feed excerpt only.
wsj.com