A 28% reduction in sum of diameters is clinically meaningful. To an investor it is a miss.
Where the Threshold Came From
RECIST draws partial response at a 30% decrease in the sum of longest diameters, and progressive disease at a 20% increase from the smallest sum recorded. The thresholds exist so that trials, readers and institutions can be compared with one another. Before standardized criteria, a response in one trial was not the same object as a response in another.
Judged on that purpose, the criteria work. They are not arbitrary, they are not badly chosen, and nothing that follows is an argument for abandoning them.
What They Were Not Designed to Do
They were not designed to carry a financing decision for a single-arm early-phase study.
For an emerging biotech, the objective response rate from a small early trial is often the entire efficacy argument presented to investors, to a partner, and to a board deciding whether to fund the next study. A threshold built for cross-trial comparability has acquired an economic function nobody intended.
Two points either side of the line can separate an asset that raises a round from one that does not. The biology does not change across the boundary. The reporting convention does.
What Falls Off Either Side
When trained readers measure the same lesions independently, they do not produce identical numbers, and the variation is not uniform. A lesion with a blurred edge and an irregular shape is measured far less consistently than a sharp, round one, by anyone.
That variability sits comfortably inside the 30% response boundary. It sits less comfortably inside the 20% progression boundary, and the asymmetry deserves more attention than it gets.
A borderline patient called progressive on measurement noise comes off a therapy that may have been working and enters the survival curve as an event. The response threshold costs a sponsor a claim. The progression threshold can cost a patient a therapy.
The Case the Threshold Handles Worst
The near miss is a story about bad luck at the margin. A larger version of the same problem is not about luck.
FDA investigators published a retrospective analysis of 22 randomized trials in metastatic non-small cell lung cancer, finding tumor growth rate inversely associated with overall survival and progression-free survival. The authors note that an endpoint such as objective response rate may not fully capture a drug's effect on tumor growth rate, particularly where a therapy is not expected to produce a high response rate — a later-line setting, or an agent with a primarily cytostatic effect.
The finding is not confined to lung. In two phase 3 breast cancer trials, growth rate predicted survival better than response rate (He et al., The Oncologist 2025).
If your asset holds disease stable rather than shrinking it, the endpoint understates it at every readout, by design. A drug that holds disease steady for eleven months and a drug that does nothing can produce the same line in the response table.
What a Continuous Measure Can Add, and What It Cannot
The measurements needed to say more already exist. The reader measured every target lesion at every time point and summed them. The category is what discarded the rest — how far the disease moved, how fast, and in which lesions.
Four caveats belong with any continuous measure of response, and a vendor who omits them is selling.
- None of it is new or proprietary. The regression-growth model was developed at the National Cancer Institute and is implemented in a public software package.
- Growth rate is not an endpoint. The threshold for a clinically meaningful change has not been defined, and a working group convened by Friends of Cancer Research this year left open whether continuous scores need strict cut points at all. It complements RECIST, and regulators still expect the RECIST call.
- It needs time. A growth rate needs baseline plus two follow-ups — three time points, roughly twelve weeks on a six-week schedule. Three time points make a patient eligible for a fit, not certain of one.
- Some patients produce no growth rate at all, because their disease only shrank. How many depends on the drug; in published analyses, the better the drug, the larger that group. It is a favorable result, and a design that treats it as missing data has misread the output.
Where It Should Arrive
Since the method belongs to nobody, the question that matters to a sponsor is delivery.
A growth rate computed months later in a separate analysis can explain a result. Delivered inside the signed read, alongside the RECIST call, in the same audit trail, at the time point where it would change a decision, it is something a sponsor can act on while the study is still running.
The threshold stays. It does the job it was built for. The information underneath it should stop being discarded on the way to the report.
GenPhase.ai is an imaging core lab for oncology sponsors, built on ONIX AI. Response analytics — continuous depth of response, distance to the partial response threshold, and lesion-level trajectories — is in development and is being built into the read output alongside the RECIST call.
Sources: Malinou JN, Fan J, Cheng J, Gong Y, Shen Y-L, Larkins E. An FDA analysis of the association of tumor growth rate, overall survival, and progression-free survival in patients with metastatic non-small cell lung cancer. The Oncologist 2026;31(3):oyag009. He P, et al. Correlation between tumor growth rate and survival in patients with metastatic breast cancer treated with trastuzumab deruxtecan. The Oncologist 2025;30:oyaf057. Goldmacher G. Tumor growth rate as an intermediate trial endpoint. The Oncologist 2026;31:oyag149. Friends of Cancer Research. Leveraging AI-Enabled Tumor Assessment Tools on Radiological Images to Evaluate Treatment Effect and Support Clinical Trial Endpoints in Solid Tumors. White paper, 2026.