Category: 3. Business

In cancer medicine, the next wave of breakthroughs often appears long before regulatory approval. These investigational therapies generate momentum at major scientific meetings, shift treatment expectations, and signal where oncology is heading. The goal is not hype, but anticipation—because patients, clinicians, and researchers want to understand which emerging medicines are already demonstrating strong efficacy, meaningful survival gains, or first-in-class mechanisms. What they share is simple: their early data suggest they are capable of reshaping treatment standards in the near future.

Last year, many of the therapies highlighted by the OncoDaily Research & Intelligence team went on to receive FDA approval in 2025, while others continue to progress toward regulatory decisions. This year, we again reviewed the latest data, and the OncoDaily Research & Intelligence Editorial Team has selected the 10 most promising cancer drugs not yet approved—agents showing meaningful survival gains, impressive response rates, and first-in-class mechanisms in some of the most difficult cancers to treat.

Most Promising Cancer Drugs 2024

10 most promising cancer drugs not yet approved in 2025

1.⁠ ⁠Vepdegestrant (ARV-471)
2.⁠ ⁠Darovasertib
3.⁠ ⁠Relacorilant
4.⁠ ⁠Daraxonrasib
5.⁠ ⁠Zanzalintinib
6.⁠ ⁠GSK5764227 (HS-20093)
7.⁠ ⁠Anbenitamab (KN026)
8.⁠ ⁠AgenT-797 (iNKT) .
9.⁠ ⁠Pasritamig (JNJ-78278343)
10.⁠ ⁠Iza-Bren (EGFRxHER3 ADC)

These programs are shaping the next wave of innovation and offering a clear preview of what may redefine oncology care in the near future.

1. Vepdegestrant (ARV-471): Rewriting Endocrine Therapy for ER-Positive Breast Cancer

Indication: ER+/HER2- Advanced or Metastatic Breast Cancer (specifically ESR1-mutated).

Mechanism of Action: An oral PROTAC (PROteolysis TArgeting Chimera) protein degrader. Unlike standard inhibitors that block the receptor, Vepdegestrant binds to the estrogen receptor (ER) and triggers the cell’s own machinery to degrade and destroy it.

Key Trial: VERITAC-2 (Phase 3).

Main Results:

• PFS Benefit in ESR1-Mutant Disease: Median PFS 5.0 vs 2.1 months vs fulvestrant (HR 0.58, P<0.001); ORR 18.6% vs 4.0%.
• Overall Population: Median PFS 3.8 vs 3.6 months (HR 0.83; P=0.07), showing strongest activity in ESR1-mutant tumors.
• Safety: Mostly grade 1–2 events; grade ≥3 AEs 23%; low discontinuation (2.9%); manageable fatigue, mild LFT elevations, and low-grade QTc prolongation.

Developer: Arvinas / Pfizer

Vepdegestrant, is an oral PROTAC estrogen receptor degrader designed to go a step beyond traditional endocrine therapy. Instead of just blocking the estrogen receptor like fulvestrant or SERDs, it recruits an E3 ligase and sends the receptor to the proteasome for destruction. The Phase 3 VERITAC-2 trial, published in The New England Journal of Medicine in May 2025, tested Vepdegestrant against fulvestrant in 624 patients with ER-positive, HER2-negative advanced breast cancer who had already received a CDK4/6 inhibitor plus endocrine therapy.

Figure From NEJM Article: Progression-free Survival as Assessed by Blinded Independent Central Review.

The clearest win was in the ESR1-mutated subgroup, which is exactly where resistance to prior endocrine therapy is most common. In these patients (about 43% of the trial), median progression-free survival was 5.0 months with Vepdegestrant versus 2.1 months with fulvestrant (hazard ratio 0.58; P<0.001), and the objective response rate jumped to 18.6% vs 4.0%. In the overall population, the difference was more modest and did not reach formal statistical significance (3.8 vs 3.6 months; HR 0.83; P=0.07), reinforcing the idea that this drug is particularly powerful in ESR1-driven disease rather than all comers.

Safety looked manageable and very “everyday clinic” for an oral endocrine agent. Most adverse events were grade 1–2; grade ≥3 events occurred in about 23% on Vepdegestrant vs 18% on fulvestrant, with low discontinuation rates (2.9% vs 0.7%). Fatigue, mild liver enzyme elevations and occasional QTc prolongation were seen but serious cardiac events were not reported, and gastrointestinal toxicity (nausea, vomiting, diarrhea) was less frequent than what has been observed with some oral SERDs.

Taken together, VERITAC-2 positions Vepdegestrant as a first-in-class PROTAC endocrine therapy with a clear niche in ESR1-mutant, ER+/HER2– advanced breast cancer. With the NDA accepted in August 2025 and a PDUFA date of June 5, 2026, it is one of the most likely near-term approvals to genuinely shift second-line endocrine practice from injections to an oral, mutation-focused degrader.

2. Darovasertib: A Vision-Saving Therapy for Uveal Melanoma

Indication: Uveal Melanoma (Neoadjuvant and Metastatic).

Mechanism of Action: A selective PKC (Protein Kinase C) inhibitor. PKC is a key driver in the GNAQ/GNA11 mutations found in approximately 90% of uveal melanoma cases.

Key Trial: Phase 2 OptimUM-09 trial

Main Results:

• Tumor Shrinkage & Eye Preservation: 83% had tumor shrinkage; ≥20% shrinkage in 54%, enabling 57% eye preservation, rising to 95% in those with ≥20% shrinkage.
• Vision Protection: 70% of plaque brachytherapy candidates had reduced predicted radiation dose, and 65% showed lower predicted risk of severe long-term vision loss; over half improved visual acuity during treatment.
• Safety & Regulatory: Generally well tolerated (grade ≥3 TRAEs 16.8%); granted FDA Breakthrough Therapy Designation as the first systemic neoadjuvant therapy with proven potential to prevent enucleation.

Developer:IDEAYA Biosciences

Darovasertib, developed by IDEAYA Biosciences in collaboration with Servier, is a potent, selective inhibitor of PKC, and is one of the most important new drugs in ocular oncology. Primary uveal melanoma is one of the few solid tumors where patients still routinely face the prospect of enucleation—surgical removal of the eye—or high-dose plaque brachytherapy that often results in major vision loss. With no approved systemic therapy for the localized disease setting, Darovasertib’s neoadjuvant strategy represents a genuine paradigm shift.

Presented at ESMO 2025 in a Proffered Paper session, the Phase 2 OptimUM-09 trial delivered some of the strongest functional-preservation data ever reported in this cancer. Across 94 evaluable patients, 83% experienced measurable tumor shrinkage and 54% achieved ≥20% shrinkage. Among patients originally recommended for enucleation, the therapy preserved the eye in 57%, and in those with ≥20% shrinkage prior to local therapy, the eye-preservation rate soared to 95%. This is unprecedented in a disease where surgical removal is often the only option.

The benefits extended beyond tumor control. In patients eligible for plaque brachytherapy, Darovasertib reduced the predicted radiation dose to critical eye structures in 70%, with two-thirds experiencing a lower predicted risk of severe vision loss at three years. Importantly, more than half of patients across both cohorts improved their visual acuity during neoadjuvant treatment—gaining an average of 17 letters (enucleation cohort) and 10 letters (brachytherapy cohort), a clinically meaningful restoration of functional vision.

Read More About Ocular Melanoma on Oncodaily

These findings, combined with a manageable safety profile—grade ≥3 treatment-related events occurred in 16.8%, with low discontinuation rates—led the FDA to grant Breakthrough Therapy Designation for Darovasertib in the neoadjuvant treatment of primary uveal melanoma. No other systemic therapy has ever demonstrated this degree of tumor control, eye preservation, and functional vision benefit in this population.

With the registrational Phase 3 OptimUM-10 trial now underway and pivotal progression-free survival data from the metastatic UM program expected in late 2025 or early 2026, Darovasertib stands out as a likely first-in-class therapy capable of redefining how uveal melanoma is treated—shifting care from organ removal to organ preservation.

3. Relacorilant: A New Option for Platinum-Resistant Ovarian Cancer

Indication: Platinum-Resistant Ovarian Cancer (PROC).
Mechanism of Action: A selective Glucocorticoid Receptor (GR) Antagonist that blocks cortisol-mediated chemotherapy resistance and restores tumor susceptibility to apoptosis.
Key Trial: ROSELLA (Phase 3).

Main Results (Updated ESMO 2025 LBA45 Data):

• PFS & OS Benefit: Relacorilant + nab-paclitaxel achieved PFS HR 0.70 (median 6.54 vs 5.52 mo) and an OS HR 0.69 with median 15.97 vs 11.50 mo.
• Strong PARPi Subgroup Activity: In PARP inhibitor–exposed patients, median PFS reached 7.36 months (HR 0.60) and 7.36 months in PARPi-progressors (HR 0.56).
• Favorable Safety: Grade ≥3 and serious AEs in PARPi patients were comparable to overall ITT (71% vs 74%; 32% vs 35%), with no new safety signals.

Developer: Corcept Therapeutics.

Relacorilant is a selective glucocorticoid receptor antagonist designed to counteract cortisol-driven chemotherapy resistance—a mechanism long recognized in ovarian cancer biology but never effectively targeted. By blocking GR signaling, relacorilant enhances taxane-induced apoptosis, directly addressing a key driver of platinum-resistant ovarian cancer (PROC), where therapeutic progress has been slow for over a decade.

At ESMO 2025, the Phase 3 ROSELLA trial (LBA45) delivered evidence supporting this approach. Adding relacorilant to weekly nab-paclitaxel produced a 30% reduction in the risk of progression (HR 0.70) and a 31% reduction in the risk of death (HR 0.69), with median OS improving from 11.50 to 15.97 months. This marks the first regimen to improve both PFS and OS over weekly taxane therapy in PROC.

Presented by Domenica Lorusso at ESMO 2025 Congress

A key highlight of the updated analysis involved patients previously treated with PARP inhibitors—a subgroup known for markedly poor outcomes after PARPi progression. ROSELLA demonstrated that relacorilant can partially reverse this resistance biology. Median PFS in PARPi-exposed patients reached 7.36 months, with hazard ratios of 0.60 (prior PARPi) and 0.56 (progressed during PARPi therapy). These results compare favorably with benchmark data such as PAOLA-1, where post-PARPi patients typically show significantly shorter benefit from subsequent chemotherapy.

Critically, relacorilant achieved these improvements without adding toxicity. Safety remained consistent with nab-paclitaxel monotherapy, with no new adverse events and similar rates of grade ≥3 and serious AEs, even in the PARPi subgroup. This distinguishes relacorilant as a rare agent that meaningfully increases chemotherapy effectiveness without compromising tolerability.

Based on the strength of the ROSELLA findings, Corcept is expanding the Phase 2 BELLA program to evaluate relacorilant-containing regimens in platinum-resistant ovarian cancer, platinum-sensitive PARPi-progressors, and endometrial cancer. This expansion reflects growing confidence that GR antagonism may have broader relevance across gynecologic oncology.

With the FDA’s acceptance of the New Drug Application for PROC and a PDUFA date of July 11, 2026, relacorilant is positioned to become the first therapy to deliver dual survival improvements in this setting—and potentially the most meaningful advance in platinum-resistant ovarian cancer management in many years.

4. Daraxonrasib (RMC-6236): The First Real Pan-RAS Inhibitor

Indication: Pancreatic Cancer (and other RAS-mutated solid tumors).

Mechanism of Action: A Pan-RAS Inhibitor (RAS(ON) multi-selective inhibitor). It targets the active form of all three major RAS variants (KRAS, NRAS, HRAS), addressing a mutation historically considered “undruggable.”

Key Trial: RASolute 303 (Phase 3) and Phase 1/2 updates.

Main Results:

• High activity in PDAC: ORR 47% with monotherapy and 55% with daraxonrasib + GnP in first-line metastatic PDAC; DCR ~90%.
• Durable benefit in 2L+: ORR 29–35%, median PFS ~8 months, median OS 13–16 months, with >90% disease control.
• Favorable safety: Mostly rash and mild GI events; no treatment discontinuations due to toxicity; dose intensity >80%.

Developer: Revolution Medicines

Daraxonrasib (RMC-6236),  represents one of the most advanced efforts to target RAS—often described as the most difficult oncogenic driver in solid tumors. Unlike mutation-specific drugs such as KRAS G12C inhibitors, daraxonrasib is a multi-selective RAS(ON) inhibitor, designed to suppress a broad spectrum of pathogenic RAS variants including G12X, G13X, and Q61X, which together drive the overwhelming majority of pancreatic ductal adenocarcinomas (PDAC).

Pancreatic cancer remains a malignancy with extremely poor survival outcomes, where median survival in the metastatic setting is measured in months and chemotherapy remains the only systemic standard. Against this backdrop, the breadth and consistency of daraxonrasib’s activity in 2025 generated significant clinical optimism.

In the second-line metastatic PDAC cohort, long-term follow-up showed durable activity rarely seen in this setting. At the 300 mg daily dose, daraxonrasib achieved ORR 29–35%, disease-control rates over 90%, and median PFS of 8.1–8.5 months. Median OS reached 13.1–15.6 months, with a median follow-up of approximately 17 months. The safety profile remained manageable: rash and mucositis/stomatitis were the most common toxicities, grade ≥3 events occurred in about one-third of patients, and importantly, no patients discontinued therapy due to toxicity, with dose intensity maintained at 86%.

Early first-line metastatic PDAC data further strengthened the signal. Daraxonrasib monotherapy yielded a 47% response rate, and the combination of daraxonrasib with gemcitabine/nab-paclitaxel delivered an ORR of 55%, with ~90% disease control and no unexpected safety concerns. Mean dose intensity remained above 80%, indicating good tolerability even in combination regimens. These response levels exceed historical expectations for first-line chemotherapy in PDAC and suggest that RAS-directed therapy may be viable even in untreated disease.

These findings support the ongoing RASolute 303 Phase 3 trial, which is enrolling first-line metastatic PDAC patients into three arms—daraxonrasib monotherapy, daraxonrasib + GnP, and GnP alone—with PFS and OS as dual primary endpoints. In parallel, the RASolute 302 Phase 3 trial is evaluating daraxonrasib monotherapy in second-line PDAC, and an additional Phase 3 study is planned in the adjuvant setting.

In October 2025, the FDA granted Orphan Drug Designation to daraxonrasib for pancreatic cancer, reflecting both the severity of the disease and the strength of emerging clinical evidence.

If the ongoing Phase 3 trials confirm the activity seen in early studies, daraxonrasib could become the first broadly active RAS inhibitor for pancreatic cancer—potentially redefining the therapeutic landscape of a disease historically dominated by cytotoxic therapy alone.

Read More About RASolute Trial on OncoDaily GI

Zanzalintinib: A New Potetial Standard of Care for Advanced Colorectal Cancer?

Indication: Refractory Colorectal Cancer (CRC).

Mechanism of Action: A next-generation Multi-kinase Inhibitor targeting VEGFR, MET, and the TAM family kinases (Tyro3, Axl, Mer), which are involved in tumor growth and immune evasion.

Key Trial: STELLAR-303 (Phase 3).

Main Results:

• OS Benefit: Median OS 10.9 vs 9.4 months vs regorafenib (HR 0.80, p=0.0045) in refractory MSS metastatic colorectal cancer.
• PFS Benefit: Median PFS 3.7 vs 2.0 months (HR 0.68), with consistent benefit across all subgroups—including patients with or without liver metastases.
• Manageable Safety: Grade 3/4 TRAEs in 59%, mainly hypertension (15%), fatigue (6%), diarrhea (6%), and proteinuria (6%); no new safety signals.

Presented by Anwaar Saeed at ESMO 2025 Congress

The phase III STELLAR-303 trial delivered the first randomized evidence that a multitargeted tyrosine kinase inhibitor combined with immunotherapy can extend survival in microsatellite-stable metastatic colorectal cancer (MSS mCRC), a population historically resistant to checkpoint inhibition. In this global study of 901 previously treated patients, zanzalintinib plus atezolizumab achieved a statistically significant improvement in overall survival compared with regorafenib, the current standard of care. Median OS reached 10.9 months with the combination versus 9.4 monthswith regorafenib (HR 0.80; p=0.0045), and this advantage was consistent across all predefined subgroups, including patients with the traditionally immune-refractory profile of active liver metastases.

Progression-free survival also favored the combination, with a median PFS of 3.7 months compared with 2.0 months for regorafenib (HR 0.68). Although the absolute magnitude of benefit was modest, the findings remain clinically meaningful in the highly refractory MSS setting. Importantly, the safety profile aligned with the known effects of VEGF/MET/TAM pathway inhibition. Grade 3–4 treatment-related toxicities occurred in 59% of patients receiving the combination, most commonly hypertension, diarrhea, fatigue, and proteinuria, but no new or unexpected safety signals emerged.

Read More About STELLAR-303 Trial on OncoDaily

Overall, STELLAR-303 positions zanzalintinib plus atezolizumab as a potential new therapeutic approach for patients with refractory MSS mCRC, supporting further biomarker-driven evaluation and ongoing final OS analyses in patients without liver metastases.

6. GSK5764227 (HS-20093): A Long-Awaited Breakthrough B7-H3 ADC for Sarcomas

Indication: Relapsed/Refractory Osteosarcoma and Soft Tissue Sarcoma (STS).

Mechanism of Action: A B7-H3–targeted antibody-drug conjugate (ADC) carrying a topoisomerase I inhibitor payload.

Key Trial: ARTEMIS-001 (Phase 1/2).

Main Results:

• Osteosarcoma: ORR 20%, DCR 86% in heavily pretreated disease (historical ORR <5% with chemotherapy).
• Soft Tissue Sarcoma: ORR 23%, DCR 92%, with responses across multiple STS subtypes.
• Safety: Manageable toxicity profile with no unexpected safety signals; consistent with topo-I ADC class.

Developer:

GSK5764227 (formerly HS-20093) is a first-in-class B7-H3–directed antibody–drug conjugate (ADC) incorporating a topoisomerase I inhibitor payload, designed to target B7-H3–expressing solid tumors. B7-H3 is highly expressed in osteosarcoma and multiple soft tissue sarcoma (STS) subtypes, making it an attractive therapeutic target in diseases with extremely limited systemic therapy responsiveness. Following GSK’s acquisition of Hansoh Bio’s ADC platform, GSK5764227 has advanced into global development as a next-generation cytotoxic ADC for sarcomas.

Updated results from the ARTEMIS-001 Phase 1/2 study were presented at ESMO 2025 and demonstrated clinically meaningful activity in heavily pretreated sarcoma populations. In relapsed/refractory osteosarcoma, GSK5764227 achieved an Objective Response Rate (ORR) of 20% and Disease Control Rate (DCR) of 86%, markedly exceeding the historical ORR of <5% associated with salvage chemotherapy. In the soft tissue sarcoma cohort, ORR reached 23%with a DCR of 92%, with responses observed across multiple sarcoma subtypes.

The safety profile was consistent with topoisomerase-based ADCs; treatment-related adverse events were manageable, and no new safety signals emerged. Importantly, the drug demonstrated durable tumor shrinkage and disease stabilization in a population with few alternative treatment options and poor prognoses.

Collectively, the ARTEMIS-001 data position GSK5764227 as a potentially transformative therapeutic candidate in bone and soft tissue sarcomas, offering one of the most substantial improvements in response rates seen in this disease area in years. Expansion into later-phase studies is anticipated.

7. Anbenitamab (KN026): A Next-Generation HER2 Bispecific Antibody

Indication: HER2-Positive Gastric Cancer (GC) and GEJ

Mechanism of Action: A bispecific HER2 antibody binding two non-overlapping HER2 epitopes (ECD2 + ECD4), enhancing HER2 blockade, receptor internalization, and ADCC.

Key Trial: KC-WISE (KN026-001, Phase 3).

Main Results:

• PFS Benefit: Median PFS 7.1 vs 2.7 months vs chemotherapy (HR 0.25, p=5.44×10⁻¹²), representing a 75% reduction in risk of progression or death.
• OS Benefit: Median OS 19.6 vs 11.5 months (HR 0.29, p=1.56×10⁻⁶), a 71% reduction in risk of death; OS not yet mature.
• High Response Rates: ORR 55.8% vs 10.8%; DCR 80% vs 41.9%; median DoR 8.2 vs 2.9 months.
• Manageable Safety: Grade ≥3 TEAEs 60.6%, primarily neutropenia, leukopenia, anemia, diarrhea, and asthenia; low cardiotoxicity (3.2%), comparable to control.

Developer: Alphamab Oncology

The first interim analysis of the Phase III KC-WISE (KN026-001) trial presented at the ESMO Congress 2025 demonstrated that anbenitamab (KN026) combined with chemotherapy provides a significant and clinically meaningful benefit in previously treated HER2-positive gastric and gastroesophageal junction (GC/GEJ) cancers after progression on trastuzumab-containing regimens.

In this randomized, double-arm study, patients with HER2-positive GC/GEJ who experienced disease progression following trastuzumab-based therapy were assigned to receive anbenitamab + chemotherapy or placebo + chemotherapy. Baseline characteristics were well-balanced, with most patients presenting with ECOG PS 1 and stage IVB disease. The median follow-up was approximately 9.7 months in both arms.

Presented By Jianming Xu, ESMO 2025

Treatment with anbenitamab yielded a statistically significant improvement in progression-free survival (PFS) compared with chemotherapy alone. Median PFS was 7.1 months in the anbenitamab arm versus 2.7 months in the control arm, corresponding to a hazard ratio (HR) of 0.25 (P = 5.44 × 10⁻¹²), indicating a 75% reduction in risk of disease progression or death. Overall survival (OS) also favored anbenitamab, with a median OS of 19.6 months (not yet mature) versus 11.5 months in the control arm, representing an HR of 0.29 (P = 1.56 × 10⁻⁶) and a 71% reduction in risk of death. Both primary endpoints of the trial—PFS and OS—were met with high statistical robustness.

Objective response outcomes were similarly improved. The ORR was 55.8% with anbenitamab compared with 10.8% in the control group, and the disease control rate (DCR) reached 80.0% vs. 41.9%, respectively. The median duration of response (DoR) was 8.2 months for anbenitamab versus 2.9 months for control therapy, indicating more durable benefit.

Safety findings showed that anbenitamab plus chemotherapy was tolerable and consistent with expected HER2-directed therapy toxicities. Grade ≥3 TEAEs occurred in 60.6% of patients receiving anbenitamab versus 51.6% in the control group, with neutropenia, leukopenia, anemia, diarrhea, and asthenia being the most common high-grade events. Importantly, cardiotoxicity was low (3.2%) and comparable between arms, despite the HER2-targeting mechanism.

Presented By Jianming Xu, ESMO 2025

These interim data indicate that anbenitamab significantly improves PFS, OS, ORR, and durability of response in trastuzumab-pretreated HER2-positive GC/GEJ, positioning it as a strong potential contender for second-line and later-line therapy. The magnitude of benefit compares favorably with results from DESTINY-Gastric04 (trastuzumab deruxtecan), suggesting possible efficacy and safety advantages. Based on these findings, additional trials are planned to expand anbenitamab development into first-line and perioperative settings.

8. iNKT Cell Therapy: An “Off-the-shelf” cell therapy with Transformative Potential

Indication: Refractory Solid Tumors (specifically Metastatic Germ Cell Tumors and Gastric Cancer).

Mechanism of Action: An allogeneic, “off-the-shelf” invariant Natural Killer T (iNKT) cell therapy. It targets CD1d (a lipid-presenting molecule on tumor cells) to trigger a dual attack: direct tumor lysis via perforin/granzyme and rapid modulation of the suppressive tumor microenvironment through the release of cytokines (like IFN-gamma) to recruit host immune cells.

Key Trial: Phase I Trial in Patients With Relapsed/ Refractory Solid Tumors

Main Results:

• Germ Cell Tumor Breakthrough: Achieved a confirmed Complete Remission (CR) in a patient with metastatic disease refractory to 7 prior lines of therapy (including high-dose chemotherapy and stem cell transplant), as published in Oncogene (July 2025).
• Gastric Cancer Activity: Demonstrated a Disease Control Rate (DCR) of ~70% in heavily pre-treated gastric cancer patients, showing efficacy in tumors typically resistant to standard checkpoint inhibitors.
• Safety Profile: Highly tolerable with no Graft-versus-Host Disease (GvHD) or severe neurotoxicity observed, validating its potential as a scalable therapy that does not require HLA matching.

Updated Phase 1 data presented at SITC 2025 highlight agenT-797 as one of the most promising entrants in next-generation, off-the-shelf cell therapy. In patients with PD-1–refractory, heavily pretreated solid tumors, the combination of agenT-797 with anti-PD-1 produced durable and meaningful clinical benefit, with a median overall survival approaching 23 months—a striking result in a population with historically poor outcomes. Several patients achieved deep, long-lasting remissions, including a complete and sustained response beyond two years in metastatic germ-cell/testicular cancer, and prolonged disease control in gastric cancer, thymoma, cholangiocarcinoma, renal cancer, and adenoid cystic carcinoma.

Presented by Ben Garmezy, at SITC 2025

Mechanistic analyses reveal why: agenT-797 functions not only through direct cytotoxicity but through immune re-orchestration, activating dendritic cells, reversing macrophage suppression, rescuing exhausted T cells, and enhancing CD8⁺ and NK-cell infiltration into tumors. The therapy maintained an exceptionally clean safety profile, with no dose-limiting toxicities, no high-grade CRS, and no neurotoxicity, underscoring its suitability for combination regimens and repeat dosing.

Read More About iNKT on OncoDaily IO

Together, these findings position agenT-797 as a first-in-class allo-iNKT cell therapy capable of restoring immune responsiveness in checkpoint-resistant solid tumors and advancing a new therapeutic paradigm for immune-cold cancers.

9. Pasritamig (JNJ-78278343): A New Bispecific Weapon in Prostate Cancer

Indication: Metastatic Castration-Resistant Prostate Cancer (mCRPC).

Mechanism of Action: A first-in-class Bispecific T-cell Engager (BiTE) targeting Human Kallikrein-related Peptidase 2 (KLK2).

Key Trial: Phase 1 First-in-Human Study (NCT04898634).

Main Results:

• PSA Response: In the Recommended Phase 2 Dose (RP2D) group, 42.4% of patients achieved a PSA50 response (≥50% reduction in PSA levels). This occurred even in patients who had failed prior potent androgen receptor inhibitors and taxane chemotherapy.
• PFS Benefit: The median Radiographic Progression-Free Survival (rPFS) was 7.9 months (95% CI 2.9–NE).
• Manageable Safety: Low CRS: Cytokine Release Syndrome (CRS) occurred in only 8.9% of patients, and importantly, all cases were Grade 1 (mild). No Neurotoxicity: No ICANS (Immune Effector Cell-Associated Neurotoxicity Syndrome) was observed.

Developer: Janssen Research & Development (a Johnson & Johnson company)

Pasritamig (JNJ-78278343) is a first-in-class bispecific T-cell–engaging antibody directed against human kallikrein-2 (KLK2), a prostate-specific serine protease highly expressed in metastatic castration-resistant prostate cancer (mCRPC). Its design enables selective T-cell redirection against KLK2-expressing tumor cells while minimizing off-tumor targeting, a long-standing limitation of T-cell therapies in prostate cancer. Based on emerging clinical activity and favorable tolerability, the FDA granted Fast Track designation in 2025, facilitating accelerated development and eligibility for priority review.

Presented by Capucine Baldini at ASCO 2025

Phase 1 results (NCT04898634) presented at ASCO and expanded at ESMO 2025 included 174 patients with mCRPC who had received ≥1 prior systemic therapies. Pasritamig displayed manageable toxicity, with 82.2% experiencing a treatment-related adverse event (TRAE), mostly low-grade. Grade ≥3 TRAEs occurred in only 9.2%, and CRS—typically a major concern with bispecifics—occurred in <10%, all Grade 1. The most frequent TRAEs at the recommended Phase 2 dose (RP2D) were infusion-related reactions (22.2%), fatigue (15.6%), and low-grade CRS (8.9%).

The RP2D regimen was established as:

• SUI 3.5 mg Day 1, SU2 18 mg Day 8, then
• 300 mg Day 15, followed by 300 mg every 6 weeks (Q6W) intravenously.

This schedule produced the most favorable balance between pharmacokinetics, immunologic engagement, and tolerability. In this RP2D cohort, the median radiographic PFS was 7.9 months (95% CI, 2.9–NE), and 42.4% of patients achieved a ≥50% PSA decline, an early marker of antitumor activity in mCRPC. Responses were observed across patients with extensive prior therapies and molecularly heterogeneous disease.

Presented by Capucine Baldini at ESMO 2025

At ESMO 2025, investigators presented a detailed translational analysis comparing dosing intervals. Weekly and Q3W dosing schedules were associated with rising PSA values and evidence of T-cell exhaustion. In contrast, Q6W dosing preserved a reprogrammable progenitor CD8⁺ T-cell compartment, characterized by lower expression of activated caspase-3 and γH2AX in peripheral blood mononuclear cells (n=186), indicating reduced activation-induced cell death (AICD). This immunologic preservation correlated with superior clinical activity: 44% of patients on the Q6W schedule achieved a complete response at any time, compared with 33% on Q3W.

Investigators emphasized that the degree of progenitor T-cell maintenance strongly correlated with PSA50 responses, independent of dose level, providing mechanistic validation that KLK2-directed T-cell redirection can produce clinically actionable immunity in prostate cancer.

Collectively, Phase 1 clinical and correlative data support pasritamig as a potentially impactful therapeutic strategy for mCRPC. It is the first agent to successfully leverage KLK2 as a T-cell–engaging target, with a safety profile compatible with outpatient delivery and enough preliminary efficacy to justify accelerated development. Upcoming Phase 2 studies are expected to further define its role within the evolving immunotherapy landscape for advanced prostate cancer.

10. Iza-bren (Izalontamab Brengitecan): An EGFR×HER3 ADC for  NSCLC, SCLC and Nasopharyngeal Cancer

Indication: EGFR-mutant NSCLC after third-generation EGFR TKI and platinum chemotherapy; recurrent/metastatic nasopharyngeal carcinoma (R/M-NPC) after ≥2l; relapsed/refractory small cell lung cancer (SCLC); NSCLC with non-classical oncogenic drivers.

Mechanism of Action: A first-in-class EGFR×HER3 bispecific antibody–drug conjugate linked to a topoisomerase-I inhibitor payload (Ed-04)

Key Trials: BL-B01D1-101/102 (Phase I/Ib); BL-B01D1-303 (Phase III NPC).

Main Results:

• NPC (Phase III): cORR 54.6% vs 27.0% with chemotherapy; median PFS 8.38 vs 4.34 months (HR 0.44); DoR 8.51 vs 4.76 months. OS immature. Grade ≥3 TRAEs 79.9%, predominantly hematologic, manageable with supportive care.
• NSCLC (Phase Ib): ORR 45.6%, cORR 35.3%, DCR 82.4%, median PFS 6.7 months across multiple oncogenic drivers. EGFR exon20ins and HER2 cohorts showed highest activity (cORR ~67%, DCR 100%). Only one low-grade ILD event.
• SCLC (Phase I): ORR 55.2%, confirmed ORR 44.8%, median PFS 4.0 months, OS 12.0 months. In patients with only one prior line of chemo-immunotherapy: ORR 80%, cORR 75%, median PFS 6.9 months, OS 15.1 months.

Safety: Class-typical topo-I ADC profile with predominantly hematologic toxicity

Developer: SystImmune and Bristol Myers Squibb (BMS)

Iza-bren (izalontamab brengitecan) is emerging as one of the most versatile next-generation ADCs in solid tumors, not because it targets a single niche mutation, but because it leverages a dual-receptor strategy that is relevant across multiple epithelial cancers. EGFR and HER3 form a powerful signaling pair: EGFR drives proliferation, while HER3 acts as an amplifier of ligand-mediated PI3K/AKT signaling and a key partner in resistance to targeted therapy. By co-targeting both and delivering a topoisomerase-I payload, iza-bren is designed to hit the signaling architecture and the DNA simultaneously.

The first indication where iza-bren looks truly practice-changing is recurrent/metastatic nasopharyngeal carcinoma. In BL-B01D1-303, a heavily pretreated R/M-NPC population (≥2 prior chemo lines, prior PD-(L)1) was randomized to iza-bren or physician’s-choice capecitabine, gemcitabine, or docetaxel. The interim analysis, presented as a late-breaking oral at ESMO 2025 and published in The Lancet, showed that iza-bren more than doubled confirmed ORR (54.6% vs 27.0%) and nearly doubled median PFS (8.38 vs 4.34 months; HR 0.44), with responses lasting almost twice as long (median DoR 8.51 vs 4.76 months). OS is not yet mature, but the magnitude of PFS and response benefit alone strongly suggests that this ADC can outperform late-line chemotherapy in a space with historically modest options.

Presented by Huaqiang Zhou at ESMO 2025

Toxicity in BL-B01D1-303 was dominated by myelosuppression, reflecting the potent topo-I payload, with grade ≥3 TRAEs in about 80% of patients versus 62% with chemotherapy. However, treatment discontinuation rates and non-hematologic toxicities were acceptable, and no new safety signals emerged compared with earlier studies.

Beyond NPC, iza-bren has generated a surprisingly broad signal in lung cancer. In the phase Ib NSCLC expansion presented at ASCO 2025, patients were enrolled by genotype rather than by histology—EGFR exon 20 insertions, atypical EGFR mutations, HER2, ALK/ROS1/RET fusions, KRAS (including G12C), BRAF, MET exon 14, NTRK, and SMARCA4. Most had progressed on standard targeted therapy when available and had received ≤1 prior line of chemotherapy. Shorten with AI

At the RP2D of 2.5 mg/kg D1D8 Q3W, the overall ORR was ~46%, with a cORR of 35% and DCR above 80%; median PFS was 6.7 months. The most striking activity was seen in EGFR exon 20 insertion and HER2-mutant disease, where confirmed response rates exceeded 50–65% and disease control was near-universal, with median PFS not reached in some cohorts at the time of reporting.Shorten with AI

Presented by Yan Huang at ASCO 2025

In SCLC, where meaningful progress has been rare, the phase I BL-B01D1-101 results are striking. Among 58 previously treated patients, iza-bren achieved an ORR of 55.2% (confirmed 44.8%), with median PFS 4.0 months and OS 12.0 months. Activity was even stronger in patients treated after only one prior line of PD-(L)1 plus platinum: ORR reached 80%, confirmed ORR 75%, and median PFS and OS were 6.9 and 15.1 months. These data justify the ongoing phase III trial in SCLC after one prior PD-(L)1 + platinum regimen (NCT06500026)

Why “Not Yet Approved” Still Matters Now

In oncology, “not yet approved” is not a technical footnote – it is the whole context. These medicines are still investigational: their safety and efficacy are being defined in controlled trials, not yet in everyday practice. Early response rates and PFS curves can look extraordinary, but until regulators review mature data in larger, more diverse populations, these drugs remain accessible mainly through clinical trials or carefully selected compassionate use. For patients and clinicians, that means hope, but not yet a standard of care.

The therapies highlighted here were selected by the OncoDaily Research & Intelligence Editorial Team based solely on the strength of their clinical data to date – agents showing meaningful activity in difficult cancers, innovative mechanisms, and the potential to shift treatment expectations if pivotal trials confirm their promise.

Together, this list of 10 most promising cancer drugs not yet approved, reflect the direction in which cancer medicine is moving—and remind us that tomorrow’s breakthroughs are already being written today.

Written by OncoDaily Research & Intelligence Editorial Team

The settlement ends an investigation opened in 2011 after Santander flagged potential wrongdoings at its branch in Paris, the prosecutor Laure Beccuau said in the statement, which confirmed an earlier report by BFM TV station.

Sign up here.

A judicial case started two years later over potential tax fraud, embezzlement and other offenses between 2003 and 2010, she added.

A Santander spokesperson said the bank had identified the issues 15 years ago and reported them back then.

The bank had provisioned for the settlement so it will not have an impact on its bottom line, he said.

“Santander remains committed to complying with the highest anti-money-laundering industry standards and regulations,” he said.

($1 = 0.8595 euros)

Our Standards: The Thomson Reuters Trust Principles., opens new tab

To investigate how neural activation for correctly recalled memories varied across different time delays, we examined the contrast of remote >recent correct trials during object presentation at retrieval (Figure 4, ‘Retrieval fMRI’).

We first tested whether the remote > recent contrast significantly differed from zero in each age group and session (day 1 and day 14), as an indicator of differential engagement during memory retrieval. FDR-adjusted results showed no significant results in the anterior and posterior HC (Figure 4A), anterior PHG (Figure 4B), and RSC (Figure 4G) across sessions and age groups (all p>0.054; see Supplementary file 3 for details). To rule out the possibility that these nonsignificant differences reflect an overall absence of retrieval-related activation, we tested whether mean activation for recent and remote items – each relative to the implicit baseline – was significantly above zero. FDR-adjusted results revealed that activation in these ROIs was significantly greater than zero (all p<0.031), except in the recent day 1 condition in children for the posterior HC (p>0.141) and the precuneus (p>0.056, see Supplementary file 4 and Figure 4—figure supplement 1 for details). These findings indicate that the anterior and posterior HC, anterior PHG, and RSC are similarly engaged during successful retrieval of both recent and remote memories, regardless of delay or age group. (As a control analysis, we tested whether the anterior and posterior HC, anterior PHG, and RSC were similarly engaged during retrieval of recent and remote items over time using the LME models. These models included mean activation relative to the implicit baseline, a Session × Delay × Group interaction, and Subject as a random intercept. The results were consistent with the earlier findings, showing no significant main effect of Delay [all p>0.106], Group [all p>0.060], or Session × Delay interaction [all p>0.340], indicating comparable engagement of these ROIs across delays and age groups [see Supplementary file 6 for full statistical details].) Other ROIs showed more differentiated patterns, which are discussed below. (In contrast, the vlPFC, CE, posterior PHG and LOC, precuneus, and mPFC showed a significant main effect of Delay [all p<0.009, see Supplementary file 5 for details], indicating time-related changes in the remote > recent contrast. These effects are examined in more detail below. Notably, these findings are consistent with results from the whole-brain analyses; Supplementary file 7.)

To further explore the more differentiated patterns observed in other ROIs, we examined changes in the remote >recent contrast across age groups and sessions (day 1 and day 14) using LME models, controlling for sex, handedness, general intelligence, and mean reaction time. All main and interaction effects were FDR-adjusted, and all post hoc tests were Sidak-corrected (see Supplementary file 5 for details).

For the posterior PHG (Figure 4B), a significant Session × Group interaction, F_(1,83) = 9.54, p=0.020, ω²=0.09, indicated a more pronounced increase in remote >recent mean signal difference over time in young adults compared to children, b=0.11, t₍₈₃₎ = 3.09, p=0.003.

Similarly, also for the cerebellum (Figure 4C), a significant Session × Group interaction, F_(1,161) = 7.68, p=0.020, ω²=0.04, indicated a stronger increase in remote > recent mean signal difference over time in young adults compared to children, b=0.09, t₍₁₆₀₎ = 2.77, p=0.006.

For the mPFC (Figure 4D), a significant main effect of Group, F_(1,86) = 7.61, p=0.023, ω²=0.07, denoted that the overall remote > recent mean signal difference in children was higher than in young adults, b=–0.10, t₍₈₆₎ = –2.76, p=0.007.

For the vlPFC (Figure 4E), a significant main effect of Group, F_(1,82) = 31.35, p=<0.001, ω²=0.13, indicated an overall lower remote > recent mean signal difference in children compared to young adults, b=–0.125, t₍₁₀₈₎ = –3.91, p<0.001. In addition, a significant main effect of Session, F_(1,99)=10.68, p=0.005, ω²=0.09, pointed out overall higher remote > recent mean signal difference on day 14 compared to day 1, b=0.08, t₍₉₉₎ = 3.27, p=0.001.

For the precuneus (Figure 4F), a significant main effect of Group, F_(1,161) = 5.09, p=0.027, ω²=0.02, indicated an overall lower remote > recent mean signal difference in adults compared to children, b=–0.05, t₍₁₆₀₎ = –2.26, p=0.037. In addition, a significant main effect of Session, F_(1,161) = 6.50, p=0.036, ω²=0.03, denoted an overall lower remote > recent contrast for day 14 compared to day 1, b=–0.05, t₍₁₆₀₎ = –2.55, p=0.012. Although the remote > recent contrasts were mostly negative, the mean activation for recent and remote items – each relative to the implicit baseline – was significantly greater than zero for all delays and group (all p<0.023), except for children’s recent items on day 1 (p=0.056).

For the LOC (Figure 4H), a significant main effect of Group, F_(1,82) = 9.12, p=0.015, ω²=0.09, indicated a higher remote > recent mean signal difference in young adults compared to children, b=0.07, t₍₈₂₎ = 3.02, p=0.003. Additionally, a significant main effect of Session, F_(1,97) = 16.76, p=<0.001, ω²=0.14, showed an overall increase in remote > recent mean signal difference on day 14 compared to day 1, b=0.07, t₍₉₇₎ = 4.10, p=<0.001. Furthermore, a significant Session × Group interaction, F_(1,81) = 6.42, p=0.032, ω²=0.06, demonstrated higher increase in remote > recent mean signal difference over time in adults compared to children, b=0.09, t(81) = 2.53, p=0.013.

Of note, we conducted an additional univariate analysis using a subsample that included only participants who needed two learning cycles to reach the learning criteria (see Supplementary file 8 for details). The subsampled results fully replicated the findings from the full sample and demonstrated that the amount of re-exposure to stimuli during encoding did not affect consolidation-related changes in memory retrieval at the neural level.

In summary, our findings revealed distinct consolidation-related neural upregulation for remote memory between children and adults. From day 1 to day 14, adults showed a higher increase in remote > recent signal difference for remembered items in the posterior PHG, LOC, and cerebellum than children. Adults showed overall higher remote > recent difference in the vlPFC than children, while children showed overall higher remote > recent difference in the mPFC than adults. Furthermore, we observed a constant activation of anterior and posterior HC, anterior PHG, and RSC in memory retrieval across age groups irrespective of memory type or delay.

Story Continues

While the company is growing quickly, its first-quarter earnings report for fiscal 2026 was a bad miss. The company posted a loss of $0.34 per share, while analysts had expected a gain of $0.14 per share.

However, it’s just a paper loss and not operational. Iren posted a $665 million unrealized loss tied to the revaluation of financial instruments connected to its convertible notes. The loss reflects mark-to-market swings and not a change in Iren’s business performance.

Revenue of $240.3 million was up 355% from a year ago, and net income of $384.56 million—before taking into account the $665 million charge—was an improvement over the company’s net loss of $51.7 million in the first quarter of fiscal 2025. Bitcoin mining revenue brought in $232.9 million, while AI cloud services revenue was $7.3 million.

“Iren continues to execute with discipline, delivering record results this quarter and meaningful progress in our AI cloud expansion,” Co-CEO Daniel Roberts said. “We secured several new multi-year contracts, including a landmark partnership with Microsoft, which solidifies Iren’s position as a leading AI cloud service provider and expands our reach into new hyperscale customer segments.”

We already know that Morgan Stanley is bullish. The consensus rating from other analysts isn’t quite as strong, with some growing concern in recent weeks.

Nine of 14 analysts surveyed by Barchart have Iren stock as a “Strong Buy,” with two suggesting investors hold the stock and three others calling it a “Strong Sell.” But three months ago, there were no “sell” ratings at all—the trend of cautious sentiment is slowly building.

The mean price target of $85.45 is a whopping 86% increase from today’s price, and the most bullish call of $136 forecasts a possible 200% jump. The low price target indicates a possible fall of 15%.

On the date of publication, Patrick Sanders had a position in: NVDA, NBIS. All information and data in this article is solely for informational purposes. This article was originally published on Barchart.com

The New York Times sued an embattled artificial intelligence startup on Friday, accusing the firm of illegal copying of millions of articles. The newspaper alleged Perplexity AI had distributed and displayed journalists’ work without permission en masse.

The Times said that Perplexity AI is also violating its trademarks under the Lanham Act, claiming the startup’s generative AI products create fabricated content, or “hallucinations”, and falsely attribute them to the newspaper by displaying them alongside its registered trademarks.

The newspaper said that Perplexity’s business model relies on scraping and copying content, including paywalled material, to power its generative AI products. Other publishers have made similar allegations.

The lawsuit is the latest salvo in a bitter, ongoing battle between publishers and tech companies over the use of copyrighted content without authorization to build and operate their AI systems.

Perplexity in particular has become a target of multiple legal disputes and faces similar accusations from a number of publishers as it tries to aggressively build market share in a hyper-competitive market for generative AI tools. Cloudflare, one of the world’s most prominent digital infrastructure companies, accused Perplexity earlier this year of hiding its web-crawling activities and scraping websites without permission – a serious accusation with potential copyright implications. Perplexity denied the allegations.

Perplexity has raised around $1.5bn in the past three years through multiple funding rounds, most recently closing a $200m round in September that valued the company at $2obn. It has attracted a variety of big-name investors, including Nvidia and Jeff Bezos, as money has flooded the AI industry.

San Francisco-based Perplexity AI also faced a lawsuit from media baron Rupert Murdoch’s Dow Jones and the New York Post.

Multiple news outlets, including Forbes and Wired, have accused Perplexity of plagiarizing their content, in one case allegedly copying a Wired article about Perplexity’s own plagiarism issues. The Chicago Tribune, Merriam-Webster Dictionary and Encyclopedia Britannica have all additionally filed lawsuits against Perplexity in recent months, accusing the company of copyright infringement.

In October, social media company Reddit also sued Perplexity in New York federal court, accusing it and three other companies of unlawfully scraping its data to train Perplexity’s AI-based search engine.

Perplexity faces legal challenges from its fellow tech companies as well. Amazon last month filed a lawsuit against Perplexity over the search engine’s AI agent shopping feature. The suit alleged that Perplexity was covertly accessing Amazon users’ accounts and masking its AI browsing activities, which Perplexity has denied while accusing Amazon of bullying and attempting to stifle competitors.

Perplexity did not immediately respond to a Reuters request for comment.

Mayer Brown has been recognized in five categories in the 2026 edition of Chambers FinTech Legal USA, including earning a Band One ranking in the Payments & Lending category. Additionally, partner David Beam was ranked in Band One for Payments & Lending.

Chambers FinTech Legal USA offers expert legal insights on critical issues for businesses, highlighting key developments in five practice areas: FinTech Nationwide; Payments & Lending; Blockchain & Cryptocurrencies; Data Protection & Cyber Security; and Corporate, Securities & Financing.

Review the complete list of Mayer Brown’s rankings.

###

Clinical Topics:
Arrhythmias and Clinical EP, Implantable Devices, SCD/Ventricular Arrhythmias, Atrial Fibrillation/Supraventricular Arrhythmias, Cardiovascular Care Team, Prevention

Keywords:
CardiaCast, Arrhythmias, Cardiac, Wearable Electronic Devices

Introduction

Diabetes represents one of the greatest public health challenges of the 21st century, contributing to morbidity, mortality, and economic loss worldwide. Currently, 589 million adults (aged 20-79 years) worldwide have diabetes (approximately 1 in 9 adults), and this number was projected to rise to 853 million by 2025 [,]. India accounts for a rapidly growing disease burden, particularly among its rural and underserved populations []. Escalating rates of type 2 diabetes and prediabetes are closely linked to shifts in lifestyle, dietary patterns, and limited access to preventive health care []. Rural communities are often disproportionately impacted due to limited health care access, low health literacy, and barriers to effective behavioral risk reduction [].

Mobile health (mHealth) interventions have emerged as promising tools for addressing such challenges by delivering targeted health education and behavior change messages directly through mobile device. A growing body of work also demonstrates that mHealth interventions, including SMS- and app-based educational programs, can improve diabetes-related health behaviors and clinical outcomes in low-resource settings []. Recent systematic reviews and trials show significant benefits for glycemic control, adherence, and lifestyle change in both urban and rural Indian populations [,]. At the same time, studies from 2024-2025 emphasize the emerging utility of artificial intelligence (AI)–powered platforms that dynamically tailor content using real-time user data, potentially setting a new standard in digital health communication []. Several contemporary randomized trials and reviews report that both traditional and AI-enabled mHealth modalities are feasible and beneficial, with limited but rapidly increasing direct comparisons in rural and underserved populations [].

Despite this progress, most previous interventions deploy generic, “one-size-fits-all” messages that do not consider individual behavioral patterns, motivational states, or changing needs over time []. This lack of personalization may lead to low engagement and suboptimal outcomes, and few published studies have directly contrasted static mHealth with AI-personalized approaches in rural Indian settings [].

Key uncertainties persist regarding the comparative effectiveness, user acceptability, and scalability of AI-driven personalized navigation versus simpler, static messaging strategies in real-world rural communities with varying digital access and literacy []. There is a pressing need to evaluate which digital strategies offer the maximum benefit for the least cost and the highest reach.

In this study, we rigorously compared, in a large rural cohort, an adaptive AI-driven mHealth intervention (dynamic) with a standard, static messaging system for diabetes prevention. To bridge the gap between the two types of interventions, this study leveraged reinforcement learning algorithms to optimize message delivery by sending customized messages based on an individual’s preferences (healthy food intake, unhealthy dietary habits, physical activity, diabetes symptom knowledge, and awareness of complications), aimed at promoting a healthy lifestyle based on the transtheoretical model of behavior change []. This individualized, data-driven communication framework represents an innovative strategy for enhancing engagement, contextual relevance, and behavioral adoption in diabetes prevention among rural populations. Our study will offer a model for future user-centered digital health interventions that can be customized for diverse behavioral domains beyond diabetes prevention. Therefore, this formative evaluation aimed to assess the effectiveness of an AI-enabled, personalized mHealth messaging intervention compared to traditional, nonpersonalized mHealth messaging in promoting engagement with diabetes risk reduction behaviors among adults in the rural district of Gulbarga, Karnataka, South India, while also examining how the intervention functions in real-world settings, what the participant engagement patterns are, and what the potential areas for improvement are [].

Methods

Study Design, Setting, and Population

The study used a quasi-experimental pre-post design with a control group and was conducted from January to November 2022, with recruitment and baseline data collection occurring between January and March 2022 and follow-up data collection completed from September to November 2022, among the rural population in the district of Gulbarga, Karnataka, South India. Gulbarga (also known as Kalaburagi), is situated in northeastern Karnataka, part of the Deccan Plateau, and is characterized by a semiarid climate and a predominantly agrarian economy. With a population of approximately 2.5 million (based on the 2011 Census, India), the district has a literacy rate of 64.8% []. Mobile phone usage among the rural population has experienced significant growth in recent years, with increasing adoption for communication, agricultural information access, and digital payments, although challenges in network connectivity persist in remote areas. The study population consisted of adults without diabetes aged 18-60 years.

Sample Size

Sample size calculations were based on a two-sided significance level of 95% and a statistical power of 80%, assuming a 10% difference in the primary outcome (physical activity) between the AI-enabled mHealth (intervention) group and the traditional mHealth (control) group. Based on these parameters, the calculated sample size was 415 participants for each group. To account for an anticipated 20% nonresponse rate, the target sample size was adjusted to 520 participants per group, resulting in a total target enrolment of 1040 participants.

Recruitment of Participants

Frontline workers (FLWs) facilitated the recruitment through community mobilization events in Gulbarga, aimed at raising awareness about the benefits of the AI-enabled, personalized mHealth messaging intervention used in this study. All participants received comprehensive information about the study’s purpose, objectives, potential risks, and benefits before they provided written consent for participation based on the eligibility criteria. Participants’ eligibility was determined based on self-reported or absence of physician-diagnosed diabetes (convenience sampling). Individuals with a known diagnosis of diabetes mellitus or that confirmed by medical records or current use of antidiabetic medication were excluded. This study followed a quasi-experimental design, without randomization. Eligible participants were divided into two groups, intervention and control, with 541 participants in each group.

Following recruitment, participants received an opt-in link, and FLWs guided them through the WhatsApp opt-in process. Upon successful opt-in, participants began receiving diabetes prevention messages starting the next day. The first follow-up was conducted 6 months after baseline data collection, using in-person interviews to fill out the survey at follow-up, facilitated by the FLWs. Blinding was not possible due to the nature of the intervention.

Intervention Design and Implementation

Arogya World’s mDiabetes program forms the foundation of the mHealth initiative evaluated in this study []. The mDiabetes program delivers diabetes prevention and control information directly to individuals’ mobile phones regardless of their risk status. Developed in 2011 in collaboration with the Rollins School of Public Health at Emory University, the standard program consists of 57 messages crafted using the transtheoretical model of behavior change. The system was built on a reinforcement learning framework using a Deep Q-Learning (DQN) algorithm, adapted for real-world mHealth deployment without prior training data, as described by Kinsey et al []. At baseline, participants completed a questionnaire assessing five behavioral domains: healthy food intake, unhealthy dietary habits, physical activity, diabetes symptom knowledge, and awareness of complications. Their responses were used in a warm-up phase to initialize state scores, ensuring that participants with lower baseline scores in a given domain received more relevant messages. Each week, the reinforcement learning agent delivered two customized health messages and two follow-up questions. Participant responses were used to update their individual state scores. A reward signal was generated when a participant’s score improved, and this was used to update the replay buffer.

The DQN was optimized weekly using minibatch gradient descent and a target network, with an ϵ-greedy policy applied to balance exploration and exploitation. Over time, this iterative process enabled the system to adaptively tailor content to participant needs (eg, used in the intervention arm: “You can reduce your risk of diabetes by walking briskly for 30 minutes daily; try walking to the temple or shops.”). These messages, available in 12 languages, were validated by Arogya World’s Behaviour Change Task Force comprising experts in diabetes, public health, and behavior change from both national and international spheres. Messages in the mDiabetes program are typically distributed as SMS texts, automated voice calls, or WhatsApp messages, sent twice weekly over a 6-month period. The program has reached approximately 2 million people to date, with this study using WhatsApp as the delivery platform.

A detailed description of building a customized messaging system for health intervention in underprivileged regions using reinforcement learning has been provided elsewhere () [].

The mDiabetes program includes an AI-based system to enhance traditional mHealth (diabetes) interventions by developing dynamic, customized text messages to improve adherence to diabetes prevention behaviors:

• AI-enabled mHealth: This intervention was different from the traditional mHealth program due to the addition of an AI system to develop a dynamic and customized text messaging intervention based on end-user feedback. Participants in the intervention group received two customized health-related messages on WhatsApp (containing information about diabetes complications and the impact of nutrition and physical activity on diabetes prevention), coupled with two questions probing their risk profile/behavior. The subsequent week’s messages for each participant in the intervention group were based on their responses to the two lifestyle-related questions from the previous week.
• Traditional mHealth: A total of 57 static mHealth messages were delivered twice a week via WhatsApp as per a standard scheduler for a period of 6 months, focusing on improving knowledge, attitudes, and practices related to diabetes prevention behaviors, including physical activity and dietary habits [].

Data Collection

We assessed 2096 individuals for eligibility, of whom 1014 (48.4%) were excluded (n=598, 59%, did not meet the inclusion criteria and 416, 41%, declined participation). A total of 1082 (51.6%) participants were divided into two groups with an equal number of participants, an intervention group (n=541, 50%) and a control group (n=541, 50%). In the control group, 34 (6.3%) participants did not complete the WhatsApp opt-in process and therefore never received the intervention messages; these individuals did not contribute intervention or outcome data and were excluded from the analysis. Data for this study were collected from primary sources using structured questionnaires, direct interviews, and anthropometric measurements conducted by trained FLWs. Participant demographics, such as age, sex, education, and employment status, were recorded.

Physical activity was assessed through self-report questions on frequency and duration (≥30 minutes/day considered active), adapted from the World Health Organization guidelines on physical activity and sedentary behavior for adults [].

Dietary habits were measured by frequency of fruit and vegetable intake per day or per week and avoidance of high-fat foods. Secondary outcomes included knowledge of diabetes symptoms, complications, and preventive behaviors. These questionnaire domains underwent expert review by public health and behavioral science specialists and were pretested in a pilot sample for contextual relevance, clarity, and cultural appropriateness before field implementation. Anthropometric measurements, including height and weight, were obtained using calibrated instruments (a SECA 213 portable stadiometer and a SECA 803 digital scale, respectively). In addition, the BMI was calculated (kg/m²), and participants were categorized as underweight (<18.5 kg/m²), normal (18.5-24.9 kg/m²), and overweight/obese (≥25 kg/m²) []. All responses were self-reported except for anthropometric measurements. Awareness-related items (eg, “Are you aware of diabetes?”) and knowledge questions (eg, causes, complications) were scored dichotomously (yes/no). Lifestyle questions on diet and physical activity used frequency scales (eg, daily, 3-4 times/week, rarely) The same questionnaire was administered at baseline and endline to capture change over time (see ). Engagement and response data from WhatsApp-delivered messages were automatically logged by the mHealth platform.

Statistical Analysis

All statistical analyses were performed using StataMP 64 software (version 17.0). Descriptive statistics summarized the baseline characteristics and outcome variables for both groups. Categorical variables were reported as frequencies and percentages, and continuous variables were summarized using means (SDs).

Normality of continuous variables was assessed using the Shapiro-Wilk test. For normally distributed continuous variables (eg, BMI), independent-sample t tests were performed to compare group means. For categorical variables (eg, physical activity, dietary behaviors), chi-square tests were performed to examine group differences at both baseline and endline.

To evaluate the effect of the intervention, separate multivariable logistic regression models were constructed for each binary outcome variable (≥30 minutes of daily physical activity, daily fruit intake). Each model included the intervention group (AI-enabled vs traditional mHealth), the baseline value of the respective outcome, and additional covariates identified through univariate analysis (variables with P<.20 were considered for inclusion). Key demographic factors, such as age, sex, and employment status, were also retained, where appropriate. Both unadjusted odds ratios (ORs) and adjusted odds ratios (aORs) with corresponding 95% CIs were reported. For continuous outcomes, such as the BMI, ANCOVA was performed, with endline BMI as the dependent variable and baseline BMI included as a covariate to control for initial differences.

Where multiple comparisons were made across lifestyle behavior outcomes, Bonferroni correction was applied to control for type I error. The adjusted significance threshold was set at α/N, where α=0.05 and N is the number of comparisons. With eight outcomes, this yielded an adjusted threshold of P<.006.

Model diagnostics were conducted to assess multicollinearity and model fit (eg, pseudo R², Akaike Information Criterion). All statistical tests were two-sided, and P<.05 was considered statistically significant unless otherwise corrected via Bonferroni adjustment. Missing data were handled using complete-case analysis, and follow-up rates were reported to account for potential attrition bias.

Ethical Considerations

The study protocol was reviewed and approved by the Institutional Ethics Committee of Anusandhan Trust, Mumbai (reference number IEC26/2021) prior to implementation. The research was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki and relevant national guidelines for biomedical and health research involving human participants. Informed consent was obtained from all participants before enrolment. Participants were informed about the study objectives, procedures, and potential risks and benefits and were assured that their participation was voluntary, with the right to withdraw at any time without affecting regular standard care. To ensure privacy and confidentiality, no personal identifiers were collected. The dataset used for analysis was anonymized and securely stored in password-protected files (server at Arogya World) and will be retained for 3 years, accessible only to the study investigators. This study was reported in accordance with the Transparent Reporting of Evaluations with Nonrandomized Designs (TREND) statement []. No monetary or material compensation was provided to participants for their participation. Additionally, no identifiable participant images or other personal visual materials were collected or included in the manuscript or appendices. In addition, regulatory guidelines of the Telecom Regulatory Authority of India (TRAI) were followed in sending text messages []. Our study was a quasi-experimental design and was carried out as a formative evaluation to understand how the intervention works in real settings, how participants engage with it, and how it could be improved, and there was no randomization procedure; hence, this study did not require mandatory clinical trials registration-India (CTRI) registration.

Results

Sociodemographic Characteristics of Participants

shows the participant selection flow diagram. Of the 1082 participants enrolled, 1048 (96.9%) completed the 6-month follow-up and were included in the analysis (n=541, 50%, in the AI-based mHealth [intervention] group and n=507, 48.1%, in the traditional mHealth [control] group), with 387 (36.9%) males and 661 (63.1%) females. The participants were largely within the age range of 26-50 years (n=723, 69%). Educational attainment varied across the sample, with nearly half of the participants (n=517, 49.3%) having completed some level of schooling and only 73 (6.9%) holding postgraduate or higher qualifications. Employment status showed that 76% (n=796) of the participants were employed, with a slightly higher proportion of employed individuals in the intervention group (n=426, 78.6%) compared to the control group (n=371, 73.1%). details the sociodemographic characteristics of both groups.

‎

Adherence to Different Components of Diabetes Prevention in Intervention and Control Groups

Physical Activity

The most notable intervention effects were observed in physical activity patterns, particularly in the intervention group. The percentage of participants engaging in regular physical activity increased from 66.7% (n=361) at baseline to 72.8% (n=394) at endline (P=.02) compared to the control group (n=333, 65.7%, at baseline to n=353, 69.6%, at endline; P=.18). Additionally, the duration of physical activity demonstrated substantial improvements in the intervention group, reporting a 15.4% increase in participants engaging in 30 minutes or more of daily exercise (P<.001) compared to a 14.8% increase in the control group (P<.001). details the adherence of both groups to different components of diabetes prevention.

Intake of Fruits and Vegetables

Both groups exhibited high baseline adherence to daily fruit and vegetable intake, with minimal changes observed postintervention. In the intervention group, 94.8% (n=513) of the participants continued to meet the recommended servings of fruits and 99.8% (n=540) for green vegetables, showing a slight improvement from baseline for vegetables (0.4%; P=.22). Similarly, the control group maintained comparable levels of dietary adherence, with only a negligible decrease in fruit intake (–1.7%; P<.001) and no change in vegetable intake ().

Behavioral Changes

Among participants receiving the AI-enabled mHealth intervention, the proportion who preferred walking short distances for daily chores rose from 96.7% (n=523) to 98.2% (n=531). This represented an increase of 1.5% (95% CI –0.4 to 3.4, P=.11). The use of stairs declined from 91.3% (n=494) to 87.6% (n=474), a change of –3.7% (95% CI –7.3 to 0.1, P=.48). In the control group, walking short distances fell from 97.8% (n=493) to 96.5% (n=489), a change of –1.3% (95% CI –2.9 to 1.4, P=.37), and stair use declined from 91.0% (n=461) to 88.1% (n=447), a change of –2.9% (95% CI –6.7 to 0.9, P=.07).

Participation in household chores increased significantly in both groups (). In the intervention group, the proportion of participants engaging in household chores rose from 93.5% (n=506) to 98% (530), an absolute gain of 4.5% (95% CI 2.1-6.9, P=.009). The control group showed a similar improvement (from n=474, 93.4%, to n=497, 98%), which also reached significance. Agricultural work participation increased from 75.7% (n=410) to 83.4% (n=451) in the intervention group (7.7%, 95% CI 2.8-12.4, P=.002) and from 73.5% (n=346) to 80.2% (n=407) in the control group (6.7%, 95% CI 1.5-11.9, P=.014).

Primary Outcomes

Physical Activity

shows the logistic regression results for achieving at least 30 minutes of daily physical activity at endline. After adjusting for baseline status and other covariates, there was no significant difference between the two groups. Participants in the intervention group had similar odds of meeting the 30-minute daily physical activity goal compared to those in the control group (aOR 1.0, 95% CI 0.7-1.3, P=.74). Baseline physical activity was a strong independent predictor of endline physical activity (aOR 2.1, 95% CI 1.5-3.1, P<.001). Older age was associated with greater odds of regular physical activity (aOR 3.8, 95% CI 1.6-9.3 for >50 years vs 18-25 years, P=.003), while being employed was associated with lower odds of daily physical activity (aOR 0.2, 95% CI 0.1-0.3, P<.001). These findings suggest that participant characteristics, rather than intervention type, are more influential in determining physical activity outcomes.

Daily Fruit Intake

As shown in , there was no significant difference in daily fruit consumption between the two groups at endline. After adjusting for baseline fruit intake and covariates, the odds of consuming fruit daily were modestly higher in the intervention group (aOR 1.4, 95% CI 0.8-2.3), though not statistically significant (P=.24). Baseline fruit intake was strongly predictive of endline fruit intake (P<.001). Age and employment were also associated with daily fruit intake: participants aged 26-35 years had higher odds of eating fruit daily than those aged 18-25 years (aOR 4.7, 95% CI 1.9-11.8, P=.001), while being employed was linked to lower odds of daily fruit intake (aOR 0.3, 95% CI 0.1-0.8, P=.02).

Body Mass Index

ANCOVA () showed no significant difference in the mean BMI between the two groups at endline. After adjusting for baseline BMI, the mean difference was essentially 0 (–0.0 kg/m², 95% CI –0.6 to 0.5, P=.95). Baseline BMI was a strong determinant of endline BMI (P<.001), but no intervention effect was detected.

Exploratory Behavioral Outcomes

Other incidental physical activity measures also showed no significant differences between the two groups. The aORs (95% CI) for the intervention versus the control group were as follows: stair use (aOR 0.9, 95% CI 0.7-1.4, P=.79), walking for chores (aOR 2.4, 95% CI 1.0-6.1, P=.06), helping with household chores (aOR 1.0, 95% CI 0.4-2.3, P=.94), and farm work (aOR 1.3, 95% CI 0.9-1.8, P=.19). See Tables S1-S4 in .

Message Delivery Rate, Responses to Feedback, and Engagement Over 6 months

Across both study arms, 85,000 WhatsApp messages were scheduled. About 82,600 (97.2%) of these were successfully delivered (95% CI 96.9-97.1), and approximately 77,000 (93.2%) of these were marked as read (95% CI 92.8-93.2) .At baseline, 65.1% (352) of intervention group participants reported making changes in response to messages; at endline, this proportion was 66.4% (n=359 participants). In comparison, 64.7% (n=328) of participants in the control group at both baseline and endline reported making changes in response to messages; see Table S5 in . In addition, engagement remained relatively stable over the 6‐month period ().

Discussion

Principal Findings

This study evaluated the effectiveness of an AI-enabled mHealth intervention (mDiabetes) versus traditional mHealth for promoting diabetes prevention behaviors in rural India. Our primary finding was that there were no significant differences between the AI-enabled and traditional mHealth groups for the primary outcomes of physical activity and dietary behaviors after 6 months of intervention. Both interventions demonstrated effectiveness in maintaining and promoting physical activity behaviors, with baseline activity being the strongest predictor (aOR 2.1, 95% CI 1.5-3.1, P<.001). This suggests that any form of consistent mHealth messaging may be beneficial for diabetes prevention regardless of AI customization. The finding that age>50 years is associated with higher odds of physical activity (aOR 3.8, 95% CI 1.6-9.3, P=.003) is particularly encouraging, as this demographic faces higher diabetes risk. Conversely, employment being associated with lower physical activity odds (aOR 0.2, 95% CI 0.1-0.3, P<.001) highlights the real-world barriers that working adults face in rural settings. Our finding are consistent with recent reviews that show that although mHealth interventions generally improve physical activity within groups, the incremental benefit of AI- or app-based enhancements over active comparators is often modest [,].

Our findings are in contrast with those of existing studies, which generally report declining physical activity with advancing age due to reduced mobility, chronic conditions, and competing health limitations []. However, some studies have reported that older adults may engage more consistently in routine or incidental activities, such as walking, household chores, or agricultural work, particularly in rural areas [].

For daily fruit intake, although the intervention effect was not significant, the age-specific patterns are noteworthy. Participants aged 26-35 years showed higher odds of daily fruit intake (aOR 4.7, 95% CI 1.9-11.8, P=.001), suggesting this age group may be more receptive to dietary behavior change. The negative association with employment (aOR 0.3, 95% CI 0.1-0.8, P=.022) again underscores socioeconomic barriers to healthy behaviors. Similar findings have been reported in other mHealth interventions in India, where improvements in fruit intake were observed but often without strong between-group differences. For instance, the SMART Eating trial in Chandigarh documented a significant rise in fruit consumption using IT-enabled strategies compared to standard education []. These findings suggest that AI-enabled approaches may promote healthier choices, such as fruit consumption [].

To the best of our knowledge, this study is the first in India to rigorously, directly compare, in a large rural cohort, an AI-enabled mHealth intervention (dynamic) with a traditional static mHealth intervention for diabetes prevention. Key novel features include its real-world rural setting with minimal exclusions, the innovative use of reinforcement learning AI to customize messages based on individual feedback, and its focus on adults without diabetes for prevention rather than diabetes management []. Additionally, the study achieved a high retention rate of 97%, demonstrating both feasibility and acceptability of mHealth in rural populations.

ANCOVA revealed no significant difference in the mean BMI between the two groups at endline after adjusting for the baseline BMI and other covariates (mean difference –0.0 kg/m², 95% CI –0.6 to 0.5, P=.95). Previous research suggests that even noncustomized mHealth interventions can yield positive outcomes, particularly in contexts where strong community support structures exist. Such interventions, regardless of personalization, have the potential to influence health behaviors; however, their effectiveness may be enhanced when integrated with complementary support mechanisms [,]. A recently published study [] showed that combined mHealth and community health education intervention improves diabetes awareness and healthy habits in rural areas, indicating potential for lasting outcomes and guiding future public health efforts in rural settings [].

The implications of these findings for public health practice are substantial. The demonstrated effectiveness of the AI-based mHealth intervention in increasing physical activity and maintaining healthy behaviors suggests that such tools could be crucial in diabetes prevention programs, especially in rural and underserved areas where health care resources are limited. Moreover, the ability of AI-driven interventions to provide customized guidance and real-time feedback makes them particularly suited for scalable, population-level health initiatives. These mHealth interventions can bridge significant gaps in health care delivery, particularly in resource-constrained settings []. However, the findings also suggest that a “one-size-fits-all” approach may not be sufficient. Integrating AI-driven mHealth interventions with existing health care systems, including community health workers and primary care providers, could enhance their effectiveness and sustainability.

Additionally, the existing literature also emphasizes the importance of physical activity and active lifestyles in managing health outcomes, particularly in mHealth interventions. Several studies have demonstrated that regular physical activity, such as engaging in 30 minutes or more of exercise daily, significantly reduces the risk of chronic conditions, including diabetes and cardiovascular diseases [,]. Moreover, nonexercise activities, such as household chores and farm work, which showed significant associations with better health outcomes in this study, are well recognized as beneficial contributors to physical and metabolic health. Evidence suggests that active engagement in household tasks and manual labor can improve cardiovascular health and reduce the risk of complications in chronic diseases like diabetes [,].

The significance of using stairs and walking for daily chores, particularly in the mHealth AI group, mirrors findings from other studies that promote incidental physical activity as a valuable component of overall health management []. Activities like stair climbing, which are simple to incorporate into daily routines, have been shown to improve cardiovascular function and aid in glucose regulation, both critical factors in diabetes prevention and management []. These findings underscore the importance of integrating physical activity, both structured and unstructured, into health interventions to enhance their effectiveness, particularly in AI-driven mHealth programs.

Strengths and Limitations

This study marks the first national attempt to use WhatsApp-based text messaging on mobile phones to support educational interventions aimed at preventing diabetes. The strengths of this study include a large sample size (N=1048), which offers adequate statistical power to detect intervention effects, and a high retention rate of 97% that minimizes selection bias and enhances the reliability of findings. Conducting the study in a real-world rural community setting further strengthened external validity. Additionally, rigorous statistical approaches were used, including appropriate adjustments for multiple comparisons, and a comprehensive set of outcome measures covered both behavioral changes and knowledge gains []. Moreover, in-person data collection by trained FLWs helped ensure data quality and reduced the potential for response bias, and the development of the intervention was guided by the transtheoretical model of behavior change, ensuring a solid theoretical basis.

However, the study has certain limitations. First, the reliance on self-reported data for physical activity and dietary habits (primary outcomes) is subject to recall and social desirability biases, potentially overestimating the true effects of the interventions. Second, the relatively short intervention period of 6 months limits the assessment of long-term sustainability of behavior changes. Third, the recruitment process involved an opt-in procedure, which could introduce selection bias, as participants who chose to participate may be more motivated to adopt healthy behaviors than the general population. Finally, biochemical markers were not objectively assessed to evaluate the clinical outcomes due to a lack of financial resources, which would have provided more detailed insights into the biological effects of the intervention. To address these limitations, further studies could incorporate objective measures of primary outcomes, such as accelerometers or pedometers for physical activity and validated dietary assessment tools, including food frequency questionnaires or 24-hour dietary recalls. Extending the intervention and follow-up periods would allow for evaluation of the sustainability of behavior changes over time. Recruitment strategies that ensure a more representative sample of the target population could help minimize selection bias. Furthermore, the inclusion of biochemical markers or other clinical endpoints would provide more robust evidence of the physiological and metabolic impacts of the intervention, enhancing the translational relevance of the findings. Additionally, reliance on WhatsApp messaging may have excluded individuals without smartphone access, limiting generalizability to economically disadvantaged populations. Collectively, these limitations likely bias the results toward the null hypothesis, suggesting that the true effects of the interventions may be underestimated rather than exaggerated.

Conclusion

This study revealed that engaging, well-designed static messages can be just as effective as complex AI-personalized approaches in diabetes prevention, challenging prevailing assumptions and pointing to cost-effective, scalable options for program managers and policymakers.

This study demonstrates that traditional mHealth interventions are as effective as AI-enabled approaches for promoting diabetes prevention behaviors in rural India. Although this finding challenges assumptions about the superior effectiveness of AI-powered health interventions, it provides valuable evidence for scalable, cost-effective diabetes prevention strategies. The high acceptability and retention rates of both AI-driven and traditional interventions suggest that consistent health messaging through accessible platforms like WhatsApp can effectively support diabetes prevention efforts in rural populations.

Rather than viewing the lack of AI superiority as a negative finding, this result should be interpreted as evidence for the democratization of effective health interventions. Simple, well-designed mHealth programs can achieve meaningful health behavior changes without requiring sophisticated technological infrastructure, making diabetes prevention more accessible to underserved rural populations.

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

When Dario Amodei left OpenAI in 2020, chief executive Sam Altman wished him well, anticipating that the project Amodei, his sister Daniela and other departees were planning would probably focus “less on product development and more on research”.

“We look forward to a collaborative relationship with them for years to come,” Altman wrote in a blog.

Instead, Amodei has become the sharpest thorn in Altman’s side. Anthropic, the artificial intelligence start-up he co-founded, will end the year with around $10bn in annualised revenue and is growing fast. It is in talks to raise funds at a valuation of over $300bn and is now laying the groundwork for a blockbuster initial public offering.

Amodei’s decision to leave OpenAI, and the work he has done since, is underpinned by two convictions, according to multiple people who know him. One, that he is more capable of building all-powerful AI than his former boss; and two, that the world would be safer if he does.

“He has a strong view on where he’s going. Dario understands you have to have a good business to pursue the mission,” says Matt Murphy, a partner at Silicon Valley investment firm Menlo Ventures, which led a funding round for Anthropic last year.

Investors in the company describe Amodei as someone whose evangelism for “safe AI” is wedded to keen commercial instincts. “Founders are either technical, good at product or sales. Dario is one of the few CEOs I’ve met in my life who does all three,” says Divesh Makan, founder of Iconiq Capital, which led Anthropic’s last funding round.

Amodei was born and raised in San Francisco by a mother who renovated libraries and a leathersmith father. The 42-year-old studied physics at Stanford and gained a PhD in biophysics before embarking on a career as an AI researcher. He joined Chinese internet giant Baidu in 2014 before a brief stint at Google Brain.

In 2016, he was among the earliest employees of OpenAI, founded by Altman, Elon Musk and nine others as a place to pursue AI research without the commercial pressures of a corporate tech parent. Amodei was instrumental in developing the large language models behind chatbot ChatGPT.

But after five years, Amodei left following disagreements with Altman over OpenAI’s direction and with concerns about AI’s potential for harm if appropriate guardrails were not put in place. In 2021, he co-founded Anthropic with his sister.

“What he thought was important was developing a company where these things could be deployed safely and transparently in the world,” says Ravi Mhatre, co-founder of VC firm Lightspeed Venture Partners, which invested over $1bn in Anthropic this year. “He felt he needed a clean slate.”

But this focus on safe AI development has earned him criticism in both Washington and Silicon Valley. David Sacks, Trump’s AI tsar, claimed in October that Anthropic was running a “sophisticated regulatory capture strategy based on fear-mongering”. Investor Marc Andreessen argues that extra AI regulation will impede US start-ups.

Critics cast Amodei as a “doomer” influenced by the effective altruism movement, which believes AI poses an existential threat to humanity. The Amodei siblings deny they are effective altruists or doomers. But the company’s early funding came from investors with ties to the movement, including Facebook co-founder Dustin Moskovitz and FTX co-founder Sam Bankman-Fried, who was later convicted of fraud.

There are also concerns from some in the AI sector that Anthropic’s rapid growth is now testing Amodei’s ability to balance the pursuit of “safe” AI with the needs of his shareholders.

“At the early stage of Anthropic, they very much said ‘we don’t want to fuel the AI race, we want to be just behind the frontier and do AI safety research.’ That’s clearly not the case now. Some people in the AI safety community are pretty unhappy with that,” said a person who works in AI safety.

Anthropic has received the backing of Google, Amazon, Microsoft and Nvidia, and earlier this year changed its stance on accepting funding from the Middle East. “Unfortunately, I think ‘no bad person should ever benefit from our success’ is a pretty difficult principle to run a business on,” Amodei told staff.

To articulate his strategic shifts and view of the world, Amodei likes to write lengthy public essays (the last one exceeded 13,000 words). During a five-hour podcast interview last year, he also took a brief diversion from discussing programming languages to hold forth on the meaning of life.

His earnest messages have been well received by the public. And his ebullience about his company’s mission is popular among employees, helping Anthropic retain top researchers in a competitive market. “He has cult leader status,” says the person in AI safety.

The company is now in the earliest stages of preparing for a public listing. There is strong demand from investors to own a slice of what Amodei has built.

“I can’t imagine the company without Dario. He is the person who spearheads the key technical challenges and motivates everyone,” says Lightspeed’s Mhatre. “What’s Apple without Steve Jobs or Microsoft without Bill Gates?”

george.hammond@ft.com

Additional reporting by Cristina Criddle and Tabby Kinder