An AI-generated drug molecule has posted a 40% objective response rate in a Phase 1 oncology trial for mesothelioma — one of the most treatment-resistant cancers in medicine — providing the clearest clinical evidence yet that reinforcement-learning-based generative chemistry can design molecules with genuine pharmacological activity in solid tumor oncology, not just in computational models.

Insilico Medicine (HKEX: 3696) reported Monday at ESMO 2026 in Madrid that ISM6331 — an oral, non-covalent pan-TEAD inhibitor whose chemical structure was generated and optimized entirely by the company's Chemistry42 AI platform — achieved a 40% ORR among five response-evaluable mesothelioma patients at the 130 mg once-daily dose level. The data were presented as a Rapid Oral Abstract by Dr. Melina E. Marmarelis of the Abramson Cancer Center at the University of Pennsylvania.

The result matters beyond its clinical numbers. Mesothelioma harbors inactivating mutations in the NF2 tumor-suppressor gene in approximately 70% of patients, meaning a TEAD inhibitor like ISM6331 can be deployed across the patient population without requiring molecular pre-selection by genotype — a structural advantage that most targeted oncology therapies, which depend on rare driver mutations present in only a fraction of patients, do not have.

Mesothelioma's Molecular Dependency Created a Target Too Hard for Classical Chemistry

Malignant mesothelioma — most often a cancer of the pleural lining of the lung, caused predominantly by asbestos exposure — remains one of oncology's most intractable diseases. Approximately 3,000 Americans are diagnosed with the disease each year. Median survival after diagnosis has historically been measured in months, and patients who have already progressed through first-line regimens — typically platinum-based chemotherapy combined with checkpoint immunotherapy — face vanishingly few options.

The disease's molecular biology points directly at the Hippo signaling cascade. Functioning normally, the pathway phosphorylates the transcriptional co-activators YAP and TAZ through a kinase chain (MST1/2 → LATS1/2), keeping them sequestered in the cytoplasm and preventing them from driving pro-tumor gene expression. In mesothelioma, frequent loss-of-function mutations in NF2 — the gene encoding the Merlin tumor suppressor — disable this cascade, leaving nuclear YAP and TAZ free to partner with TEAD transcription factors (TEAD1 through TEAD4) and sustain tumor proliferation, as NF2 mutation patterns in malignant pleural mesothelioma demonstrate.

The frequency of NF2 loss in mesothelioma — approximately 70% of malignant pleural cases — means ISM6331 can be studied in a largely unselected mesothelioma population and still expect meaningful target engagement across most patients. That is structurally different from, for example, an EGFR inhibitor that can only work in the minority of lung cancer patients whose tumors carry an EGFR mutation. For the TEAD inhibitor class, mesothelioma's molecular epidemiology is not incidental to why the disease was chosen for a Phase 1 proof of concept — it is the reason.

The catch was the target itself. TEAD proteins have a hydrophobic palm binding pocket that is poorly accommodating of typical drug-like molecules. This structural feature put the Hippo–TEAD axis on the industry's informal list of promising-but-unapproachable targets for years, and classical medicinal chemistry approaches struggled to generate compounds that could engage it selectively without hitting related proteins or exhibiting poor oral bioavailability. It took a different kind of chemistry engine to work through the problem.

How Chemistry42 Built a Molecule That Conventional Methods Could Not

Chemistry42 is Insilico Medicine's generative AI drug design platform, built on a multi-agent reinforcement-learning architecture that was formally described in JCIM in 2023. The system integrates more than 30 generative models — including GENTRL, a variational autoencoder with target-specific reinforcement in latent space — and trains across a corpus of approximately 10 million publicly available compounds and 100 million virtual building blocks.

The key architectural feature that distinguishes Chemistry42 from earlier virtual screening or docking approaches is its reward pipeline. When the platform generates a candidate molecule, it does not evaluate that molecule on a single criterion — predicted binding affinity, for instance — and accept or reject it accordingly. Instead, a simultaneous multi-objective scoring function evaluates the molecule across potency against the target (using the known 3D crystal structure of TEAD for structure-based scoring), synthetic accessibility (can the molecule actually be made in a lab?), ADMET properties (absorption, distribution, metabolism, excretion, toxicity profile), metabolic stability, and selectivity against off-target proteins. These reward scores feed back to retrain the generative models in Chemistry42, biasing each successive iteration of generation toward structures that score higher across all objectives simultaneously — an approach analogous to how a reinforcement-learning agent improves at a game by receiving environmental feedback.

The generation cycle runs in 2 to 72 hours, and the entire process from target specification to a shortlist of synthesizable candidates replaces what classical medicinal chemistry typically executes iteratively over years of bench work.

Applied to the TEAD target, Chemistry42's structure-based design mode — using the known TEAD1 crystal structure to guide generation — was able to explore chemical space that traditional medicinal chemistry approaches had found unproductive. The platform's reward pipeline simultaneously optimized for engagement of the TEAD palm binding pocket, non-covalent binding mechanism (avoiding covalent modification approaches that can introduce off-target reactivity), and coverage of all four TEAD isoforms (TEAD1–4), all of which are expressed in mesothelioma tumor cells and all of which can compensate for each other if an inhibitor targets only one. The result was ISM6331: an oral, non-covalent pan-TEAD1–4 inhibitor that cleared preclinical development and entered first-in-human study in January 2025.

The Phase 1 Clinical Data: 40% ORR at 130 mg, Responses After as Many as Six Prior Lines

The ongoing global Phase 1 study (NCT06566079) had enrolled 36 patients across dose cohorts from 15 mg to 130 mg once daily as of the August 14, 2026 data cutoff. The patient population was heavily pretreated: median prior lines of therapy were two, but the range extended to twelve.

At the top evaluated dose of 130 mg once daily, five patients with mesothelioma were response-evaluable. Two achieved confirmed partial responses, yielding a 40% ORR at 130 mg. Computed-tomography-confirmed target-lesion shrinkage in these two patients was 58% and 36%, respectively — reductions that are atypical for first-in-human dose-escalation trials, where the primary objective is safety rather than efficacy. One of the responders had received five prior lines of treatment before enrolling; the other, six.

Across all evaluable dose levels from 30 mg to 130 mg, 4 of 22 mesothelioma patients achieved objective responses — an 18.2% ORR across dose levels — and the disease control rate (objective responses plus stable disease) reached 68.2%. Responses were documented at the 30 mg, 50 mg, and 130 mg dose levels. Three of the four responders reached partial response status at their first protocol-specified assessment at Week 8, and at data cutoff, three of the four partial responses were still ongoing. One response had been durably maintained for approximately one year.

The drug also showed glimmers of activity beyond mesothelioma. Patients with heavily pretreated non-small cell lung cancer (seven prior treatment lines) and head and neck squamous cell carcinoma (eight prior treatment lines) showed tumor shrinkage in other cancers on imaging, and patients with the rare vascular tumor epithelioid hemangioendothelioma and with colorectal cancer achieved stable disease lasting more than four months. The signal across tumor types with Hippo pathway dysregulation beyond mesothelioma is preliminary but consistent with TEAD's biological role as a convergence point for multiple cancer drivers.

Safety and Target Engagement: What the Pharmacology Data Confirm

For a first-in-human dose-escalation study, ISM6331's tolerability profile was encouraging. No dose-limiting toxicities (DLTs) were observed at any level from 15 mg through 130 mg, and no patient discontinued due to adverse events. The great majority of treatment-emergent toxicities were Grade 1 or 2. Grade 3 events, occurring in 8.3% of participants, consisted of asymptomatic elevations in the pancreatic enzymes amylase and lipase, fatigue, and drug-associated dermatitis. Treatment-related proteinuria, an adverse event of special interest for this compound class, was observed in 11.1% of patients but was low-grade and fully reversible in all cases.

Two pharmacological findings provided mechanistic reassurance. Circulating levels of CCN2 — also known as connective tissue growth factor (CTGF), a secreted protein whose transcription is directly driven by YAP/TAZ–TEAD complex activity — declined in treated patients, confirming TEAD target engagement in human tissue, not only in preclinical models. This biomarker signal is essential for a novel mechanism: it demonstrates the drug is doing in people what the computational platform designed it to do.

Additionally, ISM6331's terminal plasma half-life of approximately 18 to 32 hours supports once-daily oral dosing without the accumulation risks associated with very long-lived compounds. This pharmacokinetic profile — which Chemistry42's reward pipeline was explicitly designed to optimize for alongside potency — is consistent with a drug designed for continuous daily suppression of a constitutively active oncogenic pathway.

TEAD Inhibitors Are Now a Clinical Competition — How ISM6331 Fits

ISM6331 is entering a nascent but suddenly active drug class. Vivace Therapeutics' VT3989, a TEAD inhibitor targeting the autopalmitoylation pocket through a covalent mechanism, reported clinical proof-of-concept for the Hippo–TEAD axis in refractory mesothelioma at ESMO 2025 — VT3989 achieved 32% response rate (7 of 22 patients achieving partial responses) with an 86% disease control rate. That data marked the first human demonstration that the TEAD binding pocket is pharmacologically accessible. Orion Pharma's ODM-212 is progressing through its own Phase I/II TEADES trial (NCT06725758) in mesothelioma and related Hippo-dysregulated tumors.

ISM6331 differentiates itself on two dimensions: mechanism and design provenance. Unlike VT3989's covalent mechanism, ISM6331 binds non-covalently, which may offer a different selectivity and safety profile as the compound moves to higher exposure levels in larger studies. And ISM6331 is the only TEAD-class compound to date that was generated and multi-parameter-optimized entirely by a generative AI platform — its pharmacokinetic profile, its pan-isoform coverage, and its synthetic accessibility were all shaped simultaneously by Chemistry42's reward pipeline rather than iteratively by bench chemists.

The broader TEAD inhibitor competitive landscape also included IAG933 from Novartis, which entered Phase 1 trials with a different structural class targeting the YAP–TEAD protein-protein interaction directly. However, Novartis quietly halted enrollment in the IAG933 Phase 1 trial in October 2025 after the compound showed limited efficacy — an 13% ORR in heavily pretreated pleural mesothelioma patients — underscoring the difficulty of the target and, by contrast, the clinical meaningfulness of ISM6331's and VT3989's results. The class as a whole represents the pharmaceutical industry's first systematic attempt to drug the Hippo pathway transcriptional output, an approach that — if it succeeds in Phase 2/3 trials — could extend well beyond mesothelioma to any solid tumor with NF2 loss or aberrant YAP/TAZ activation.

Insilico Medicine: AI Drug Discovery Platform With Active China Operations and HKEX Listing

The ISM6331 data come from a company whose corporate structure and geography are worth understanding in context. Insilico Medicine is incorporated in the Cayman Islands and operationally headquartered in Cambridge, Massachusetts, with its principal R&D operations based in mainland China and a Hong Kong subsidiary. The company listed on HKEX Main Board on December 30, 2025, under stock code 3696. Its investor base includes Chinese venture firms including Qiming, WuXi AppTec, Baidu, and Sinovation Ventures, alongside US and international funds including Warburg Pincus, Sequoia, and OrbiMed. The Hong Kong Investment Corporation, a HKSAR government-owned fund, made a strategic investment in September 2025.

In March 2026, Eli Lilly entered into a partnership valued at $2.75 billion (with $115 million paid upfront) to apply the company's generative chemistry platform to drug targets in Lilly's pipeline. The deal is the most significant validation of Chemistry42 as a commercial platform to date from a major Western pharmaceutical company.

Insilico's drug discovery pipeline is built on three computational modules that together constitute its Insilico Pharma.AI drug platform: Biology42 for target identification, Chemistry42 for generative drug design, and Medicine42 for clinical development optimization.

Insilico Plans Phase 2a Dose-Optimization Study After Completing Dose Selection

Dose escalation in the NCT06566079 Phase 1 study is still ongoing. Once a recommended Phase 2 dose is selected, Insilico plans to advance to Phase 2a randomized dose-optimization study in mesothelioma.

"TEAD sat for years in the industry's 'too hard' category," said Alex Zhavoronkov, PhD, Insilico's Founder and CEO. "Today we reported first-in-human data for our pan-TEAD inhibitor designed using generative AI with a 40% ORR at 130 mg in heavily pretreated mesothelioma, deep lesion reductions of 58% and 36%, disease control in more than two-thirds of evaluable participants, and responses by week eight — including after five and six prior lines. No dose-limiting toxicities so far, and falling blood CCN2 tells us we are hitting TEAD in patients. That is generative chemistry becoming a clinical signal."

Halle Zhang, PhD (Med), Insilico's Vice President of Clinical Development in Oncology, noted that the activity and safety profile together support the transition to Phase 2 evaluation: "The 40% ORR at 130 mg and manageable safety profile in previously treated mesothelioma support continued investigation. As we complete dose selection, we plan to advance to Phase 2a randomized dose optimization."

The trial (NCT06566079) is also evaluating ISM6331 in combination with the PD-1 checkpoint inhibitor pembrolizumab — a strategy that could augment the single-agent activity the drug already appears to be generating through an immune-independent mechanism, per the trial's combination cohort.

The FDA has granted ISM6331 Orphan Drug and Fast Track designations for mesothelioma (Orphan Drug in June 2024, Fast Track in July 2026), making it Insilico's first program to achieve the latter status. Both designations reflect the agency's recognition that effective therapies for the disease remain an unmet medical need.

ISM001-055's Phase 2a Result in 2025 Established Generative Chemistry's First Efficacy Milestone

ISM6331 builds on a landmark Insilico achieved in 2025, when ISM001-055 — a TNIK inhibitor designed by the same generative AI framework to treat idiopathic pulmonary fibrosis — produced positive results in a Phase 2a randomized clinical trial: the first time an AI-designed drug had been shown to meet efficacy endpoints in a controlled human study. The results were published in Nature Medicine in 2025 and represent the clearest prior demonstration that Chemistry42 can produce molecules capable of producing measurable treatment effects in patients — not only in early-phase signals.

ISM6331 extends that demonstration into solid tumor oncology, a setting where the path from a Phase 1 signal to an approved therapy is considerably longer and harder. A 40% ORR in a Phase 1 dose-escalation trial in mesothelioma — a disease as resistant as any in the oncology field, in patients who had already failed multiple prior therapies — is a signal that oncologists rarely see from unoptimized early-stage molecules. The argument the data support is not that ISM6331 will succeed in Phase 3, but that the molecules Chemistry42 produces arrive in the clinic with measurably higher intrinsic activity than the baseline expectation for unoptimized clinical candidates.

How Does AI Drug Discovery Move From a 40% Phase 1 Signal to an Approved Therapy?

A 40% ORR in Phase 1 is a signal, not an outcome. Phase 1 trials are designed for safety and dose escalation; efficacy findings are secondary and often fail to replicate in larger, randomized Phase 2 and Phase 3 studies. The historical attrition rate for oncology drugs from Phase 1 to approval is approximately 5–7%. A strong Phase 1 signal in a small population raises the probability of success, but does not determine it.

The debate that ISM6331's data feed is not about this particular molecule's approvability — it is about whether the generative AI layer that produced the molecule meaningfully changes the probability distribution of outcomes in the development pipeline that follows. If Chemistry42's multi-objective optimization reliably produces molecules with better ADMET profiles, better target selectivity, and better pharmacokinetic behavior than the historical baseline for medicinal chemistry outputs, then the success rate across Insilico's portfolio should, over time, differ detectably from the industry average.

ISM001-055's Phase 2a success in pulmonary fibrosis provides one data point in favor of that hypothesis. ISM6331's Phase 1 data in mesothelioma add a second. Neither result resolves the question; a portfolio of Phase 2 and Phase 3 outcomes across multiple molecules will be needed to distinguish "generative AI reliably improves candidate quality" from "Insilico has had two early successes." That is the experiment now underway in Cambridge, Hong Kong, and clinical sites worldwide — and mesothelioma patients who have exhausted standard options are, for the first time, among its subjects.

ISM6331 is an investigational compound and has not been approved by any regulatory authority. All data are preliminary and subject to change as the trial continues.

Frequently Asked Questions

What is mesothelioma, and why does it have so few treatment options?

Mesothelioma is a rare cancer of the thin tissue layers (mesothelium) lining the lungs, abdomen, and other organs, caused predominantly by asbestos exposure. Approximately 3,000 Americans are diagnosed each year. The disease has historically resisted treatment because it is typically diagnosed at an advanced stage, because it frequently acquires mutations that disable normal tumor-suppressor pathways (particularly the NF2/Hippo pathway), and because cytotoxic chemotherapy agents — the standard of care for decades — rarely produce durable responses. The approval of immunotherapy-based regimens in the front-line setting improved initial response rates, but patients who progress through front-line treatment have very limited options, which is why a Phase 1 signal like ISM6331's 40% ORR at 130 mg represents meaningful news for this patient population.

How does Chemistry42 design a drug using AI, and is it fundamentally different from searching existing drug databases?

Chemistry42 does not search a database of existing molecules — it generates novel molecular structures that have never been synthesized, using multiple deep generative models (including variational autoencoders and transformer-based architectures) trained on approximately 10 million compounds and 100 million virtual building blocks. The critical distinction from classical virtual screening is the reinforcement-learning reward loop: as the platform generates candidate molecules, each one is scored against multiple simultaneous objectives — potency, synthetic accessibility, metabolic stability, selectivity — and those scores feed back to bias the next generation cycle toward better candidates across all dimensions at once. Classical medicinal chemistry optimizes these properties sequentially, which takes years; Chemistry42 addresses them in parallel within hours to days. The result is molecules shaped from the start to perform well in a human body, not just in a computational binding prediction.

What does the NF2 mutation mean for mesothelioma patients — and why does it matter for TEAD inhibitors specifically?

NF2 is a tumor-suppressor gene that, when functioning, keeps the Hippo signaling pathway active — which in turn prevents YAP and TAZ proteins from partnering with TEAD transcription factors to drive tumor growth. In approximately 70% of malignant pleural mesothelioma cases, NF2 is mutated or lost, disabling this pathway and leaving TEAD constitutively active as a driver of cancer cell proliferation. This high prevalence means most mesothelioma patients — regardless of whether they have been genotyped — harbor the molecular alteration that makes TEAD inhibition potentially therapeutic. Unlike targeted therapies that require a rare driver mutation (EGFR amplification in 10–15% of lung cancers, for example), a pan-TEAD inhibitor in mesothelioma can be studied across the patient population without pre-selecting by genotype, which is a practical advantage both in clinical trials and, if approved, in eventual clinical practice.

What comes next for ISM6331, and what would it take for an AI-designed drug to reach mesothelioma patients?

Dose escalation in NCT06566079 is ongoing. Once a recommended Phase 2 dose is selected, Insilico plans to advance ISM6331 to a Phase 2a randomized dose-optimization study. A Phase 2a trial typically involves a larger patient cohort randomized across two or more dose levels, with efficacy as the primary or co-primary endpoint — the study that would generate the data needed to justify a Phase 3 pivotal trial. Full regulatory approval would require a Phase 3 trial demonstrating statistically significant improvement in overall survival or progression-free survival versus the standard of care, a process that typically takes three to seven years from Phase 2a initiation for an oncology drug. ISM6331 also carries FDA Orphan Drug Designation and Fast Track Designation, both of which provide regulatory support (including more frequent FDA interaction) and are designed to accelerate development timelines for drugs treating diseases with limited existing therapies. Neither guarantees approval.