SCOPE predicts oncology trial outcomes

Cancer drug development has a brutal math problem. Most oncology trials fail late, after years of work and huge spending, because the models used earlier in the pipeline never really captured what human tumors would do. A team from Gustave Roussy and its spinout Orakl Oncology says it may have a better filter now — a platform called SCOPE that tries to predict trial outcomes before the trial reads out. (medrxiv.org) The news is a new medRxiv preprint, posted in April 2026, laying out the method and the first validation numbers. ### What is SCOPE actually doing? SCOPE stands for Screening-to-Clinical Outcome Prediction Engine. (medrxiv.org) It combines two things: drug-response data from patient-derived organoids — basically mini tumor models grown from real patient samples — and a clinical score that captures how sick or treatment-resistant a trial population is likely to be. Instead of predicting one patient’s response, it aims at the trial-arm level: median progression-free survival and objective response rate. ### Why use organoids here? Organoids matter because they preserve more of a tumor’s real behavior than flat cell lines. (medrxiv.org) That has made them attractive for precision oncology for years, but mostly as a patient-level or lab-screening tool. The gap SCOPE is trying to close is different — can you take those biological readouts and say something useful about how a whole clinical trial arm will perform? That jump from bench model to population forecast is the real claim here. ### How did they build the model? The training set was small but concrete: 54 treatment lines from 52 patients with metastatic colorectal cancer or metastatic pancreatic ductal adenocarcinoma, matched to organoid drug-screen data across 9 compounds. (medrxiv.org) Then the team generated synthetic trial populations from published eligibility criteria and matched those virtual cohorts to a biobank of 81 organoid lines. That is the key trick — using real tumor models plus a simulated version of who would have been allowed into the study. ### Did it work? On the paper’s numbers, yes — at least in retrospective validation. (cell.com) Across 32 arms from 23 published trials, predicted median progression-free survival tracked observed outcomes with an R² of 0.85, a mean absolute error of 0.82 months, and a Pearson correlation of 0.92. Objective response rate prediction came in at R² 0.71 with a 7.3-point mean absolute error. The combined organoid-plus-clinical model beat either component alone, and it correctly chose the better arm in 7 of 8 head-to-head comparisons. ### What makes that interesting? Most AI-for-biotech claims lean hard on molecular data or purely statistical pattern matching. (medrxiv.org) SCOPE is more of a hybrid. It starts with a living tumor proxy, then layers machine learning and clinical context on top. Basically, it is trying to answer a question investors and drug teams care about more than elegant biology: should this regimen move forward, or is it about to burn cash in phase 2 or phase 3? ### What’s the catch? The catch is right in the paper. This is a preprint, not a peer-reviewed study, and the training base is still narrow — two tumor types, 9 compounds, and a relatively modest organoid bank. (medrxiv.org) Synthetic cohorts are clever, but they are still approximations of real enrolled patients. And organoids, while better than many older models, still miss parts of the tumor microenvironment and immune context that can decide whether a cancer drug wins or fails. ### Where could this go next? The paper includes a prospective-style test case for daraxonrasib, also called RMC-6236, in metastatic pancreatic cancer, which hints at how SCOPE could be used before full trial outcomes are known. (medrxiv.org) Orakl has been positioning its platform as a way to help pharma companies rank regimens, spot biomarkers, and avoid bad bets earlier in development. But the real proof will be prospective calls that later turn out to be right. ### Bottom line? SCOPE is not a magic oracle. But it is one of the more concrete attempts to tie organoid biology to the question that actually matters in oncology development — will this trial work? (medrxiv.org) If later validation holds up, that could change where drug programs get killed, and save some of those failures from happening at the most expensive stage.