Two agents, two different arithmetics
Safety and efficacy do not combine the same way, and mixing them breaks both predictions. Atlas Bio's published framework uses additive monotherapy risk plus a population fragility term for safety, and a max-of-arms times indication-specific synergy rule for efficacy.
The reasoning is mechanistic. Immune-mediated adverse events and vasculature-mediated events act through different targets, so a patient can experience both and joint probability follows independent-event arithmetic. The exception is an organ both agents injure, which the framework models as supra-additive with a dual-hit coefficient. Efficacy is different because responses overlap on the same population: the patient who would have responded to one agent also responds to the combination.
Population fragility dominates
The published finding Atlas Bio puts first is that the same regimen at the same dose produces very different fatal event rates in different populations, and that population matters more than the regimen. Its framework carries an explicit population fragility coefficient driven by age, performance status, prior cytotoxic exposure and organ substrate.
This is why a 'new safety signal' in a later trial is so often not a new signal. A regimen moved from a fit first-line population into a heavily pretreated one will show more high-grade events with no change in drug behaviour at all. Before escalating a signal, normalise for the population you are treating.
- Baseline performance status and age distribution
- Prior cytotoxic therapy and prior radiation to the at-risk organ
- Organ reserve — hepatic, renal, cardiac — at enrolment
- Dose actually received over time, not the protocol starting dose
Dose trajectory, not starting dose
Combination regimens with a tyrosine kinase inhibitor titrate. Atlas Bio's published observation is that constitutional adverse events are sub-additive because by month three a large majority of patients are on a reduced dose rather than the starting dose, so predicting from the starting dose over-estimates fatigue, diarrhoea and weight loss.
Signal detection should therefore use exposure over the window in which the event occurred. A toxicity rate that looks alarming against the protocol dose may be unremarkable against the dose patients were actually taking.
Checking the model against a negative trial
A framework that only reproduces successful trials is untested. The LEAP-002 phase 3 trial reported median overall survival of 21.2 months with lenvatinib plus pembrolizumab versus 19.0 months with lenvatinib plus placebo (HR 0.84, 95% CI 0.71–1.00), did not meet its prespecified significance thresholds, and its authors concluded the findings do not support a change in clinical practice (Llovet et al., Lancet Oncology 2023).
Atlas Bio publishes that trial as the hardest test of its framework and as a documented miss: an earlier version predicted a positive result by transferring a synergy coefficient across indications, the trial was negative, and the recalibrated version reproduces the negative outcome. The lesson stated is that synergy reflects tissue-specific immune-vascular biology rather than a portable drug-class property.
Post-market signals need different tools
Once a combination is approved, spontaneous reporting takes over. FDA's adverse event reporting system collects adverse event reports, medication error reports and product quality complaints submitted to the agency, and supports post-marketing safety surveillance for drugs and therapeutic biologics.
Those data support disproportionality analysis, not incidence estimates, so treat them as hypothesis-generating and pair them with a denominator you trust before changing a clinical recommendation.
Atlas Bio's Combo Safety and Efficacy platform predicts combination toxicity and response from a single parameter set, with fifteen clinical-pharmacology agents spanning first-in-human dose selection through pharmacovigilance and polypharmacy risk.
Why model combination safety as additive rather than taking the worse arm?
Because the mechanisms are largely independent — immune-mediated and vasculature-mediated toxicity act on different targets, so a patient can have both. Taking the worse arm systematically under-predicts the combined event rate.
When is combination toxicity supra-additive?
When both agents damage the same organ. Overlapping target-organ risk is modelled with a dual-hit coefficient rather than independent arithmetic, because the second insult lands on tissue the first has already compromised.
Can you compare combinations without a head-to-head trial?
Indirectly, and with caveats. Safety comparisons are more defensible than efficacy comparisons because the toxicity mechanisms are class-conserved, while efficacy comparisons across non-randomised cohorts remain exploratory.
- Llovet JM et al. Lenvatinib plus pembrolizumab versus lenvatinib plus placebo for advanced hepatocellular carcinoma (LEAP-002): a randomised, double-blind, phase 3 trial. Lancet Oncology 2023 (PMID 38039993)
- FDA. Adverse Event Monitoring System (formerly FAERS) — post-marketing safety surveillance
- FDA. Codevelopment of Two or More New Investigational Drugs for Use in Combination — guidance for industry
Talk to Atlas Bio.
Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.
Contact Us →