The Ebola drugs work. The timing cliff starts at 48 hours.

The PALM trial — 681 patients, six treatment arms, conducted during the 2018–2020 Nord Kivu Ebola outbreak — established that antibody therapies REGN-EB3 and mAb114 are effective treatments for Ebola virus disease. What it did not establish is how that effectiveness depends on when treatment begins. The trial was designed to determine which drug works best, not how the time-from-onset to first dose shapes the survival curve.

We built a coupled model of Ebola virus disease — twenty-two differential equations covering viral replication, antibody response, immune activation, coagulation failure, and progressive organ failure — and calibrated it against all six PALM arms and the observational control group. The model reproduces the reported case-fatality rates within the published confidence intervals across the trial. We then ran the same six arms with the dosing time shifted, holding everything else constant.

The result is a sharp cliff.

The numbers, by treatment-start day

Using mAb114 as a representative arm:

• No treatment. Viral load peaks at 7.1 log₁₀, mean arterial pressure drops to 39 mmHg, case-fatality rate approximately 47%.
• Immediate treatment (day 0). Viral load peaks at 4.8 log₁₀, MAP at 66 mmHg, CFR approximately 23%.
• Day 1. Viral load 4.9, MAP nadir 59 mmHg, CFR approximately 30%.
• Day 2. Viral load 4.9, MAP drops to 51 mmHg, CFR approximately 31%.
• Day 5 or later. Viral load 5.7–7.0, MAP settles at 40 mmHg, CFR 32–35%.

The pattern holds across every antibody arm we simulated. The mechanism is the same: blood-pressure collapse driven by capillary leak is the proximate cause of death, with 60 mmHg representing the shock threshold below which organ failure becomes self-sustaining. Antibody therapy started early enough keeps the patient above that threshold. Started late, it does not.

Why the trial did not surface this

PALM enrolled patients on arrival at treatment centres. Time from symptom onset to enrolment was neither randomised nor controlled, which makes it impossible to isolate timing as an independent variable from the trial data alone. Patients reaching care on day 1 systematically differ from patients arriving on day 5 — in disease severity, in health-seeking behaviour, in geography, in age. That confounding is intrinsic to the design. Our model treats time as a free input, which the trial could not.

The remdesivir question

PALM reported remdesivir with a 53% case-fatality rate versus 48% in the observational control arm, which led several commentators to conclude the drug was harmful. Our model suggests a quieter explanation: remdesivir is therapeutically neutral. Its antiviral effect is real but insufficient to suppress residual viral load enough to prevent the immune cascade from engaging. The 53% versus 48% reflects allocation bias — patients assigned to remdesivir arrived later and were sicker — rather than drug-induced harm. The model reproduces both numbers when the time-to-enrolment distribution is matched to the trial’s actual allocation pattern.

The clinical implication is not pharmacological

A 48-hour window is, in principle, an enormous opportunity. Rapid post-exposure prophylaxis with mAb114 or REGN-EB3 would, on this model, push case-fatality into the low 20s — a level closer to the seasonal-influenza ward than to the historical Ebola outbreak literature.

It is also, in practice, an enormous bottleneck. Ebola outbreaks occur in geographies where reaching a hospital within 48 hours is logistically difficult and frequently impossible. The model says the pharmacology is solved. The implication is that the next gains are infrastructure, not chemistry — decentralised treatment, community case detection, pre-positioned mAb at health posts, and post-exposure prophylaxis for contacts of confirmed cases.

Limitations

We are not virologists. The compartmental form of the model treats the patient as a well-mixed system, which is wrong at the tissue level — Ebola has spatial heterogeneity in liver, spleen, and endothelium that the eight-organ structure abstracts. We do not model stochastic dynamics at low viral counts (relevant for clearance), and we do not capture genetic diversity across outbreak lineages (relevant for cross-outbreak generalisation). The timing numbers above are model outputs, not clinical estimates; they are calibrated to PALM but extrapolate beyond it. Expert critique on each of these is welcome.

What the model does say — confidently, because all six arms calibrate consistently — is that the time-from-onset axis is steeper than the trial’s enrolment design could reveal. The cliff is real. It starts at hour 48.