PTRS is built, decision by decision
A perspective on how multimodal real-world data shapes the primary nine trial-design decisions that determine oncology probability of technical and regulatory success
The historical probability of technical and regulatory success (PTRS) for an oncology candidate entering Phase 1 is approximately 5%, the lowest among major therapeutic areas.¹ Across years of customer engagements, Tempus has identified nine trial-design decisions that consistently shape oncology PTRS. In our experience, treating these nine decisions as a system and informing them with multimodal real-world data at clinical scale provides the most consistent impact on PTRS.
The PTRS problem in oncology
Oncology has the lowest clinical development success rate of any major therapeutic area, with some analyses estimating the likelihood an oncology candidate entering Phase 1 receives regulatory approval at just 5%.1
Improving the probability of technical and regulatory success (PTRS) is an incremental process dictated by every decision, long before the clinical readout is determined a success or failure. Biology, indication, biomarker stratification, eligibility, comparator, and endpoint choices each represent a critical juncture. Each decision is an opportunity to derisk development programs and optimize PTRS with the right call, or erode it with the wrong one.
A framework: Nine trial-design decisions that shape oncology PTRS
You make hundreds of decisions between candidate selection and approval. Across years of customer engagements in oncology, we have found that nine decisions consistently influence a program’s final PTRS:
Biological rationale and mechanism: Establishing the scientific basis for biomarker and target choices.
Indication and patient selection: Choosing the disease, line of therapy, and trial setting.
Molecular biomarker identification: Defining and stratifying the eligible population with molecular biomarkers.
Clinical stratification identification: Controlling clinical prognostic variability through non-molecular stratification factors.
I/E criteria optimization: Defining who is eligible for the trial.
Trial-eligible cohort identification: Determining if real patients matching the protocol exist today.
Control arm benchmarking: Projecting what outcome the comparator arm will realistically deliver.
Endpoint strategy: Selecting which readouts will declare success fastest and most reliably.
Operational and assay readiness: Ensuring the assays, algorithms, and operations can execute the trial as designed.
These decisions compound. A sharper biological rationale supports more refined biomarker-driven stratification, which supports a more defensible eligibility specification. This, in turn, enables more accurate cohort identification and calibrated comparator and endpoint assumptions, which sets up an operational and assay readiness plan.
Multimodal real-world data and the foundation models trained on it
Tempus has built one of the largest harmonized, multimodal oncology databases in the industry. It links de-identified clinical data (diagnoses, treatments, and longitudinal outcomes), molecular profiling (DNA, RNA, IHC, and ctDNA), imaging, and claims data. The integration is at the patient level, spanning more than 45 million de-identified patient journeys, including over 400,000 with full genomic, transcriptomic, imaging, and clinical data.² This scale supports in-depth research into rare patient subgroups while maintaining statistical confidence and methodological rigor.
This combination makes three kinds of PTRS-impactful work possible at the data layer:
- Quantifying populations: Understand biomarker prevalence in a given indication, co-occurrence patterns across markers, and real-world outcomes for the candidate comparator population.
- Validating mechanistic and predictive hypotheses: Determine whether a candidate predictive biomarker stratifies outcomes in a population that resembles the eventual trial population and whether the biology is mechanistically grounded.
- Pressure-testing trial design: Assess if eligibility criteria map cleanly onto how patients actually present, if the proposed stratification factor is the strongest predictor available, and if comparator-arm assumptions hold against real-world standard-of-care outcomes.
Layered on top of this data, Tempus’ foundation models function as a pre-trained biological engine. Instead of redefining biology from scratch for every new trial or research question, analyses start from a pre-trained representation of tumor biology and the patient journey. This yields higher model performance, faster development, and predictive signals that surface in indications and biomarkers where labeled data is otherwise scarce.
These models support critical decisions in ways data alone cannot. For example, Paige Predict can predict IHC and molecular biomarker status directly from H&E-stained pathology slides, unlocking molecular signals from tissue already in hand. Tempus' pathology foundation models, Virchow and PRISM, can identify morphological features associated with treatment responses that are not visible to the human eye or captured by any single marker. The multimodal foundation model surfaces candidate predictive biomarkers across modalities that single-marker analyses may miss.
In May 2026, Tempus reported initial results from its multimodal patient trajectory model, a transformer-based foundation model trained on roughly 2.5 million longitudinal patient records, including more than 250 million pages of clinical notes, 450,000 digitized medical images, and 500,000 genomic and transcriptomic sequences.2 Designed to address thousands of prediction objectives anchored in overall survival and progression-free survival without additional fine-tuning, it was applied zero-shot to a cohort of more than 1.2 million de-identified records. In an EGFR-mutant NSCLC cohort treated with osimertinib, it reached a C-index of 0.802 for overall survival (p < 0.001), and in cohorts mirroring the KEYNOTE-189, FLAURA-2, and DESTINY trials it outperformed standard approaches such as Cox proportional-hazards modeling.2 For PTRS work, a model designed to forecast survival and progression across trial-relevant populations before a trial begins is a way to pressure-test design assumptions before the protocol is locked.
Real-world data does not replace randomized evidence; it informs the decisions that surround the trial. Foundation models, applied to that data, extend what those decisions can be informed by. Tempus offers support along both axes—whether scientific teams need the data, the AI built on top of it, or both.
The impact across the development ecosystem
The nine-decision framework changes the day-to-day work of every team on an oncology program. Here is what it unlocks for each:
Program and portfolio leaders | Identify which of the nine decisions carry unresolved risk for a given program and which analyses can address them before the trial activates. | |||||
Biomarker and translational scientists | Access multimodal real-world cohorts at the scale needed to validate mechanistic hypotheses, refine eligibility criteria, and produce regulator-credible evidence. | |||||
Clinical operations and regulatory teams | Use defensible, patient-level population sizing for IND-enabling meetings and FDA pre-submission conversations, real-time patient identification for biomarker-defined trials, and external or synthetic comparator support where conventional control arms are impractical. | |||||
Patients | Benefit from trials that are better designed for the populations they enroll, with eligibility criteria, stratification factors, and endpoints grounded in how patients actually present and respond. |
Explore our interactive resource page to see this framework in action with real-world case studies and learn how Tempus’ solutions can help you increase PTRS at every stage of trial design.
Sources:
- Thomas D, Wessel C, et al. Clinical Development Success Rates and Contributing Factors 2011–2020. BIO Industry Analysis, 2021. Oncology Phase 1 → approval likelihood of approval (LOA) reported at approximately 5%, the lowest among major therapeutic areas.
- Tempus AI, Inc. Tempus Announces Initial Results from its Multimodal Foundation Model Efforts for Novel and Scalable Insight Generation in Oncology. Press release. May 29, 2026.
Article
