More evidence from every patient you enroll.

Renmara builds patient-level models from historical oncology trials to help sponsors design and analyze randomized trials and choose which doses to take forward.

Bring us a completed or planned trial, and we will show what our models would add to it.

Trial analysis

More power, fewer events or fewer patients.

Patients enter a trial with very different prognoses, and that spread blurs the comparison between arms. Renmara builds a prognostic score from the control arms of earlier trials, locked before unblinding, that your analysis adjusts for as one prespecified baseline covariate.

FDA’s 2023 guidance on covariate adjustment notes that adjusting for prognostic baseline covariates generally gives narrower confidence intervals and more powerful tests, and that prognostic indices can come from earlier studies. Read the guidance.

What could greater trial precision change?

Choose a lever to see what changes. Each uses the same gain in precision.

20%40%60%80%100%0%0.600.700.800.901.00True hazard ratio (further left, larger benefit)Planned effect, 0.72Smaller effect, 0.8094.8%90.0%68.4%59.6%Power
20%40%60%80%100%0%0.600.700.800.901.00True hazard ratio0.720.8094.8%90.0%68.4%59.6%Power
Power at the planned effect (hazard ratio 0.72)
94.8%from 90.0% without adjustment
Power if the true effect is smaller (hazard ratio 0.80)
68.4%from 59.6%

Age, sex, ECOG and stage alone reach 92.0% and 62.8%.

1002003004000Enrollment0612182430Months from first patientExpected deaths390 deathsmonth 26.0319 deathsmonth 21.5
1002003004000Enrollment0612182430Months from first patientExpected deaths390 deaths26.0319 deaths21.5
Deaths needed for 90% power
319from 390, 18% fewer
Expected event-driven readout
Month 21.5from month 26.0, 4.5 months sooner

Age, sex, ECOG and stage alone need 362 deaths, reached at month 24.0.

0200400600Patients enrolledNo adjustment600Age, sex, ECOG, stage558Renmara prognostic index492Index, smaller control arm only508
0200400600Patients enrolledNo adjustment600Age, sex, ECOG, stage558Renmara prognostic index492Index, smaller control arm only508
Patients for the same power and readout date
492from 600, 18% fewer
Control patients if only the control arm shrinks
208from 300, 31% fewer

Age, sex, ECOG and stage alone need 558 patients.

Variance removed is measured on held-out trials in our preprint. The illustrative trial follows the preprint’s design example, with 600 patients enrolled over 18 months and a control median survival of 9 months, powered at 90% for a hazard ratio of 0.72. The levers are alternative uses of the same gain.

How the numbers are calculated

On each held-out control arm, ρ² is the share of covariate-adjusted log-rank variance an analysis removed (preprint Table 1). Adjustment multiplies the variance of the test statistic by 1 − ρ². With deaths fixed, power rises accordingly. For the same power, the deaths required fall by the share ρ² (Schoenfeld’s formula, rounded up). With the accrual period and readout month fixed, enrollment falls in the same proportion, because expected deaths scale with enrollment. Shrinking only the control arm follows the preprint’s Equation 9.

These translations assume the index explains a similar share of variance in both arms of a new trial, which is why one of your completed studies is the right place to estimate it.

The gain depends on how prognostic the score is in your indication, and one of your completed studies can show it.

Get in touch

Dose optimization

Which doses deserve further development?

Project Optimus and FDA’s 2024 dosage-optimization guidance ask sponsors to justify a dosage on activity, safety, tolerability and exposure-response, not the maximum tolerated dose alone. That brings costly choices forward, including which dosages to carry into a randomized comparison and when a regimen no longer justifies investment.

Renmara models how hematologic toxicity unfolds in each patient, across dosages and schedules.

  1. 1Candidate dosagesTwo or more dosages from dose-finding studies
  2. 2Longitudinal tolerabilityDepth, timing and recovery of toxicity, by patient and dosageRenmara
  3. 3Benefit-risk assessmentCombined with activity, pharmacokinetics and exposure-response
  4. 4Dose decisionWhich dosages to randomize, advance or drop

Simulated neutrophil course through Cycle 1 at three candidate dosages

Higher dosages deepen the nadir and push more patients below grade 3. How many cross it, and whether they recover before Cycle 2, is what a dose comparison needs.

60% of MTD8% below grade 30714210246880% of MTD17% below grade 3071421MTD27% below grade 3Grade 3071421Day of cycle 1ANC, 109/L 60% of MTD8% below grade 3Grade 30246880% of MTD17% below grade 302468MTD27% below grade 307142102468Day of cycle 1ANC, 109/L

Teaching simulation with textbook model parameters, not patient data. Line: median simulated patient; shading: middle 50% and 90%; maroon: counts below grade 3 (ANC 1.0 × 109/L).

In our preprint, updating the marrow model with Week-1 blood counts improved forecasts of new severe neutropenia in Cycle 1, from an AUROC of 0.791 to 0.816 across 3,465 patients.

Research

PreprintNovember 2026

Zero-Shot Prognostic Covariate Adjustment and Mechanistic Week-1 Myelosuppression Updating Across 9,500 Oncology Patients Under FDA Guidance

The preprint tests both models on about 9,500 patients: control arms from 14 randomized trials and three AACR Project GENIE cohorts. On three held-out trials scored once after locking, the index removed 11% to 18% of overall-survival variance.

18.3%
of overall-survival log-rank variance removed on VITAL, a held-out lung cancer trial whose regimen no training trial used. Age, sex, performance status and stage removed 7.2%.
0.791 → 0.816
out-of-fold AUROC for new severe neutropenia in Cycle 1 once Week-1 blood counts update the marrow model, across 3,465 patients.
Results by validation tier

Survival variance removed, by validation tier

Redrawn from Figure 1 and Table 1 of the paper

Share of overall-survival log-rank variance removed (rho squared) by the baseline prognostic index and by a base model, for each cohort and validation tier.
Cohort Base Index (80% band) Chart
Out-of-fold, pooled control arms
NSCLC, 4 trialsn = 1,536 0.048 0.177 (0.154, 0.199)
Pancreatic, 5 trialsn = 1,151 0.041 0.186 (0.164, 0.211)
Colorectal, 5 trialsn = 2,387 0.080 0.180 (0.163, 0.197)
Leave-one-trial-out
9 development trialsrange across held-out trials 0.065 to 0.198
Held-out trials, locked before scoring
VITAL · 2L NSCLCdocetaxel · 346 deaths 0.072 0.183 (0.149, 0.222)
LEAP · 1L pancreaticgemcitabine · 158 deaths 0.044 0.177 (0.119, 0.235)
HORIZON III · 1L colorectalFOLFOX + bevacizumab · 241 deaths 0.031 0.110 (0.080, 0.141)
AACR GENIE BPC, index refit out of fold
NSCLC, first linen = 518 0.024 0.082 (0.052, 0.111)
Colorectal, first linen = 858 0.057 0.127 (0.102, 0.154)
Pancreatic, first linen = 662 0.019 0.073 (0.046, 0.102)
NSCLC, later line229 deaths 0.050 0.122 (0.079, 0.176)
Colorectal, later linen = 703 0.030 0.218 (0.183, 0.253)
Pancreatic, later linen = 510 0.076 0.219 (0.170, 0.264)
ρ² is the share of covariate-adjusted log-rank variance removed; it sets how much precision adjustment buys. The base model uses age, sex, ECOG performance status and stage; in GENIE it uses age, sex, stage, prior lines and regimen class. Bands are 80% intervals from a 500-draw patient bootstrap. GENIE rows use an index refit within GENIE without trial screening labs, so they are not a transport of the trial index.

Put Renmara to work on your trial.

Tell us about a completed or planned study and the decision in front of you, whether it concerns trial design and analysis or dose selection. We start with a retrospective evaluation on one completed study, run inside your own environment, so you can see what our models add before you use them in a prospective trial.

Tell us about your trial

Area of interest

Please do not include patient-level data, protected health information or confidential results. See our privacy notice.