AI transforms cancer drug development from trial-and-error to targeted selection

Overview

For decades, cancer drug trials have failed at a rate exceeding 95%—burning through $50-60 billion annually on treatments tested in patients who were never likely to respond. On April 17, 2025, researchers from AstraZeneca and Tempus AI published results in Cancer Cell showing that the Predictive Biomarker Modeling Framework (PBMF)—a machine learning system using contrastive learning—can identify, from existing clinical data, which cancer patients will survive longer on immunotherapy versus chemotherapy. Applied retrospectively to completed phase 3 trials, the system improved survival outcomes by 15% compared to traditional patient selection.

Now nearly 10 months later, this approach gains regulatory momentum with the European Society for Medical Oncology (ESMO) releasing in December 2025 the first guidance on validation requirements for AI-based biomarkers (EBAI), establishing standards for clinical implementation and building trust among clinicians and regulators. The implications extend beyond individual drugs: widespread adoption of AI-driven biomarker discovery could fundamentally alter cancer drug economics by enabling precise patient selection before trials begin, potentially rescuing failed therapies through identified responsive subgroups.

Play on this story Voices Debate Predict

Key Indicators

15%

Survival Risk Improvement

Patients identified by the PBMF biomarker showed 15% better survival outcomes than those in the original trial design

95%

Oncology Trial Failure Rate

Traditional cancer drug development fails at an exceptionally high rate, largely due to poor patient selection

$200M

AstraZeneca-Tempus Partnership Value

Combined data licensing and model development fees for building the largest multimodal oncology foundation model

8M+

Tempus Patient Records

De-identified patient records including clinical notes, genomics, imaging, and transcriptomics available for AI training

Interactive

Exploring all sides of a story is often best achieved with Play.

Ever wondered what historical figures would say about today's headlines?

People Involved

Etai Jacob

Lead researcher on PBMF development

Eric Lefkofsky

Leading expansion of AI precision medicine platform

Gustavo Arango-Argoty

AstraZeneca researcher

Organizations Involved

AstraZeneca

Pharmaceutical company

Developing AI-powered oncology drug development infrastructure

British-Swedish pharmaceutical giant investing heavily in AI integration across oncology research and development.

Tempus AI

Multimodal oncology data platform

Providing data infrastructure for AI-driven precision medicine

Chicago-based company that built the largest clinical and molecular database in oncology, now powering AI drug development.

U.S. Food and Drug Administration

Federal regulatory agency

Adapting approval pathways for AI-driven precision medicine

Federal agency responsible for approving drugs and companion diagnostics, increasingly accepting AI-derived biomarkers in regulatory submissions.

Timeline

AstraZeneca Acquires Modella AI to Scale Oncology AI
Industry

AstraZeneca acquired Boston-based Modella AI to embed multimodal foundation models and AI agents across its global oncology R&D organization.
January 13th, 2026
ESMO Releases First Guidance on AI-Based Biomarkers
Regulatory

European Society for Medical Oncology published EBAI framework defining validation requirements and clinical implementation standards for AI-generated biomarkers in oncology, addressing gaps in trust and regulatory acceptance.
December 17th, 2025
Cancer Cell Publishes PBMF Demonstration of AI-Improved Trial Outcomes
Research

Peer-reviewed publication showed the Predictive Biomarker Modeling Framework improved survival outcomes by 15% when applied retrospectively to immunotherapy trials for lung, kidney, and bladder cancers.
April 17th, 2025
Tempus AI Goes Public on Nasdaq
Industry

Tempus began trading under ticker TEM after raising $1.05 billion in private funding, providing public market validation for AI-driven precision medicine.
June 14th, 2024
AstraZeneca Commits $200M to Oncology AI Partnership
Industry

AstraZeneca, Tempus, and Pathos AI announced agreements to build the largest multimodal foundation model in oncology, trained on Tempus's database of 8 million patient records.
April 1st, 2024
PBMF Research Posted as Preprint
Research

AstraZeneca and Tempus researchers posted their Predictive Biomarker Modeling Framework study to medRxiv, showing AI could identify predictive biomarkers in immunotherapy trials.
January 31st, 2024
Watson for Oncology Revealed to Recommend Unsafe Treatments
Investigation

Internal documents showed IBM's Watson recommended treatments with black box warnings for patients who should never receive them, trained on synthetic cases rather than real patient data.
July 25th, 2018
FDA Approves First Pan-Tumor Genomic Profiling Test
Regulatory

FoundationOne CDx became the first FDA-approved comprehensive genomic profiling test for all solid tumors, analyzing 324 cancer genes.
November 30th, 2017
Tempus Founded After Founder's Personal Cancer Experience
Industry

Eric Lefkofsky founded Tempus after his wife's breast cancer diagnosis, recognizing that treatment decisions relied on inadequate data infrastructure.
January 1st, 2015
IBM Launches Watson for Oncology
Industry

IBM partnered with Memorial Sloan Kettering to develop Watson for Oncology, promising AI-powered treatment recommendations for cancer patients.
January 1st, 2012
FDA Approves First Biomarker-Targeted Cancer Drug
Regulatory

The FDA approved trastuzumab (Herceptin) for HER2-positive breast cancer alongside the HercepTest companion diagnostic, establishing the template for precision oncology.
September 25th, 1998

Scenarios

Predict which scenario wins. Contrarian picks score more — points lock in when the scenario resolves.

Predictions leaderboard

AI Biomarker Discovery Becomes Standard for Oncology Drug Approval

FDA and other regulators formally incorporate AI-derived biomarkers into companion diagnostic requirements, making retrospective biomarker optimization a standard step before phase 3 trials. Drug development programs that integrate AI patient selection achieve the documented 2x improvement in approval likelihood, fundamentally changing oncology economics.

Discussed by: CAS scientific trends analysis, BioSpace precision oncology coverage, ARK Invest research

Consensus —

Pharmaceutical Industry Consolidates Around Data-Rich AI Platforms

The demonstrated value of large clinical datasets for AI training creates a winner-take-most dynamic. Companies like Tempus with 8+ million patient records become essential infrastructure partners. Smaller biotechs without data access struggle to compete in precision oncology, leading to increased partnerships or acquisitions.

Discussed by: Pharmaceutical Technology, Fierce Biotech, industry analysts covering AstraZeneca-Tempus partnership

Consensus —

AI Biomarker Systems Face Watson-Like Scrutiny Over Real-World Performance

As AI-derived biomarkers move from retrospective validation to prospective use, gaps emerge between research performance and clinical reality. Regulators or researchers identify cases where AI recommendations underperform physician judgment or fail to generalize across patient populations, triggering reassessment of the technology's readiness.

Discussed by: STAT News, oncology clinical practice commentators, AI ethics researchers

Consensus —

AI Rescues Failed Drugs by Identifying Responsive Subpopulations

Pharmaceutical companies systematically apply AI biomarker frameworks to previously failed compounds, identifying patient subgroups where the drugs actually work. Several therapeutics return to clinical development with AI-optimized patient selection, creating a new category of 'rescued' precision medicines.

Discussed by: Frontiers in Oncology, Cancer Cell editorial commentary, drug development strategists

Consensus —

Historical Context

Herceptin and HER2 Testing (1998)

September 1998

What Happened

The FDA approved trastuzumab (Herceptin) for metastatic breast cancer alongside the HercepTest to identify patients whose tumors overexpress the HER2 protein. This was the first cancer drug approved with a required companion diagnostic, establishing the template for precision oncology that matches treatments to molecular targets rather than tumor location.

Outcome

Short Term

Herceptin transformed outcomes for the approximately 20% of breast cancer patients with HER2-positive tumors, who previously faced the worst prognoses.

Long Term

The drug-diagnostic co-development model became standard practice. By 2023, approximately 50% of new oncology drug approvals included companion diagnostic or biomarker requirements.

Why It's Relevant Today

The PBMF research extends this paradigm from single-gene testing to AI-discovered multi-factor biomarkers. Where HER2 testing identifies one known protein, PBMF discovers novel biomarker signatures from complex clinical and genomic data—potentially expanding precision medicine to cancers without obvious molecular targets.

IBM Watson for Oncology Failure (2012-2023)

2012-2023

What Happened

IBM invested an estimated $4 billion developing Watson for Oncology, promising AI-powered treatment recommendations. The system was trained on synthetic patient cases and the opinions of specialists at Memorial Sloan Kettering rather than real-world outcomes data. In 2018, internal documents revealed Watson had recommended unsafe treatments, including drugs with black box warnings for patients who should never receive them.

Outcome

Short Term

Only a few dozen hospitals adopted the system. Foreign physicians complained recommendations were biased toward American treatment practices and failed to align with local guidelines.

Long Term

IBM sold Watson Health assets in 2022, effectively ending its AI healthcare ambitions. The failure became a cautionary tale about AI systems trained on expert opinion rather than outcome data.

Why It's Relevant Today

The PBMF approach directly addresses Watson's core failure. Rather than encoding physician opinions, PBMF uses contrastive learning on real patient outcomes to discover which biomarkers actually predict treatment response. The 15% survival improvement came from retrospective validation on completed trials—exactly the outcome-based evidence Watson lacked.

FoundationOne CDx FDA Approval (2017)

November 2017

What Happened

Foundation Medicine's FoundationOne CDx became the first FDA-approved comprehensive genomic profiling test for all solid tumors, analyzing 324 cancer genes in a single test. The approval came through a novel parallel review process with FDA and the Centers for Medicare and Medicaid Services, establishing national coverage for genomic testing in cancer.

Outcome

Short Term

Comprehensive genomic profiling became accessible to cancer patients nationwide with insurance coverage, enabling identification of rare mutations and trial eligibility.

Long Term

By December 2025, Foundation Medicine achieved 100 approved companion diagnostic indications, demonstrating regulatory acceptance of multi-gene panels as standard of care.

Why It's Relevant Today

FoundationOne showed regulators would accept complex genomic data for treatment decisions. PBMF builds on this foundation by using AI to extract predictive patterns from even larger datasets combining genomics, clinical notes, imaging, and treatment outcomes—moving from 'what mutations does this tumor have' to 'which patients will respond to this drug.'