Multi-drug attribution in combination oncology regimens

Multi-drug attribution is the assessment of which drug in a combination regimen caused an adverse event, made separately for each suspect drug instead of once for the whole regimen. Each drug receives its own relatedness category from a named method, such as the WHO-UMC criteria or the Kramer algorithm, weighed against the other drugs, the underlying disease and the timing of exposure. In oncology, where combination therapy is the norm, that per-drug call decides whether an event belongs to the backbone or to the investigational agent.

Why multi-drug attribution shapes what the label says

Multi-drug attribution shapes which drug's label an adverse reaction lands on, which is why it matters first to pharma and biotech sponsors. FDA's labeling rule limits the Adverse Reactions section to events where there is "some basis to believe there is a causal relationship" with the drug (21 CFR 201.57(c)(7)). IND safety reporting turns on whether there is "a reasonable possibility that the drug caused" an event (21 CFR 312.32). For the full path from trial record to label section, see how adverse event attribution shapes the drug label.

A combination trial adds one question to every event: which drug. A wrong answer goes one of three ways:

• The investigational agent's event is recorded against the backbone. The new drug's label says less about its own adverse reactions than the trial saw.
• The backbone's event is recorded against the investigational agent. The new drug can inherit a warning or monitoring schedule that belongs to a drug it was given with.
• The attribution is corrected after approval. It reaches the label as a revision, sometimes with a safety communication, instead of while the trial could act on it.

Why are combination regimens the hard case for attribution?

In combination therapy, adverse event attribution weighs several candidate drugs, several exposure timelines and, often, no clean dechallenge. Single-agent attribution weighs one of each.

NCI's reporting guidelines for its own INDs name two features of this case (NCI CTEP, effective August 30, 2024):

1. Under routine reporting, when a commercial agent is used on the same arm as the investigational agent, "the entire combination (arm) is then considered an investigational intervention."
2. An investigational agent "might exacerbate the expected SAEs associated with a commercial agent," so an expected serious event for the approved drug that arrives at a higher severity requires expedited reporting.

So the routine reporting unit is the combination arm, but expedited reporting still asks about each drug. The same section notes that NCI's reporting system "is programmed to automatically submit the reports for grade 1-4 AEs due to commercial agent(s) with a possible, probable, or definite attribution" to FDA through MedWatch. Item 2 is what separates a backbone's expected event from the investigational agent's signal. The event that looks like the backbone's familiar profile may be the new drug's. Telling the two apart is the work.

Even one candidate is hard: in placebo arms, where no active study drug is given, a recurring adverse event's attribution changes between visits for the same patient in 31 to 36% of cases (Hillman et al., JCO 2010; reviewed in adverse event attribution and the drug label). A combination arm puts the same judgment across three or four real candidates.

Which causality scales handle more than one suspect drug?

Three methods come up in oncology trials: NCI CTEP attribution, the WHO-UMC system and the Kramer algorithm. WHO-UMC and Kramer weigh the other drugs in a regimen explicitly; NCI CTEP rates the event against the investigational agent or intervention.

Method	Categories	What it weighs	How other drugs enter
NCI CTEP attribution	Unrelated, Unlikely, Possible, Probable, Definite	Investigator judgment of the event against the intervention	Rated against the investigational agent or intervention
WHO-UMC system	Certain, Probable/Likely, Possible, Unlikely, Conditional/Unclassified, Unassessable/Unclassifiable	Timing, alternative causes, withdrawal, rechallenge, documentation	Named in the criteria; Probable needs other drugs to be unlikely
Kramer algorithm	Definite, Probable, Possible, Unlikely	Six scored axes, summed into a category	Other drugs score as alternative etiologic candidates

Sources: NCI CTEP adverse event reporting requirements; The use of the WHO-UMC system for standardised case causality assessment; Kramer et al., JAMA 1979.

The wording matters. Under WHO-UMC, Certain requires that the event "cannot be explained by disease or other drugs," Probable that it is "unlikely to be attributed to disease or other drugs," and Possible allows that it "could also be explained by disease or other drugs." For a drug interaction, WHO-UMC assesses the actor drug, the one that changes the kinetics or dynamics of the other.

These are classifications, not shares of blame. The WHO-UMC document lists quantifying "the contribution of a drug to the development of an adverse event" among the things causality assessment cannot do, so two drugs in one regimen can both sit at Possible.

Named methods also narrow disagreement between assessors without removing judgment. In the original Kramer validation, pair-wise agreement between two clinical pharmacologists rose from 47% without the algorithm to 63% with it (Hutchinson et al., JAMA 1979). For the methods side by side, including the Naranjo scale, see Naranjo, WHO-UMC and Kramer compared.

How per-drug attribution works, step by step

A defensible per-drug assessment answers the same questions for every suspect drug, in the same order:

1. Expand the regimen. Name every drug received, with its known adverse event profile: the label for approved drugs, the investigator's brochure for the investigational agent.
2. Fix the timeline for each drug. First exposure, most recent dose, cycle and day, and time to onset of the event.
3. Set the baseline. Was the finding present before first dose, and at what grade? That decides whether the event is new, worse than baseline, or neither.
4. List the alternative causes. Disease progression, comorbidities, concomitant medications and interactions between the drugs in the regimen.
5. Read dechallenge and rechallenge per drug. When one drug is interrupted and the others continue or resume, the course of the event afterwards is evidence about that drug specifically.
6. Score each drug separately. Apply WHO-UMC and Kramer to each suspect, with the other drugs treated as alternative candidates.
7. Record the reasoning and the source. Each category carries the evidence for and against it and the lines in the record it rests on, so a reviewer can disagree with a specific step.

A worked example: one ALT rise, three suspect drugs

Take an illustrative first-line trial in nonsquamous lung cancer: an investigational checkpoint inhibitor, called Agent X here, added to carboplatin and pemetrexed, all three given on day 1 of each 21-day cycle.

• Steps 1 to 3. The investigator's brochure for Agent X lists immune-mediated hepatitis as an identified risk; the carboplatin label reports abnormal liver function tests (total bilirubin, SGOT and alkaline phosphatase), and the pemetrexed label lists increased ALT among its adverse reactions. Baseline ALT was normal. Six weeks in, before cycle 3, ALT is 7 times the upper limit of normal with normal bilirubin: grade 3 Alanine aminotransferase increased in both CTCAE v5.0 and v6.0, which put grade 3 at ">5.0 - 20.0 x ULN" when baseline was normal.
• Step 4. Liver imaging shows no new metastases, viral hepatitis serology is negative, and no concomitant medication started in the window.
• Step 5. Per protocol, all three drugs are held and systemic corticosteroids are started. ALT falls to grade 1 within two weeks. Carboplatin and pemetrexed resume at a delayed cycle 3 without Agent X, and ALT stays at grade 1 or below through cycle 4.

Steps 6 and 7 then read differently for each drug:

Drug	Timing	Course when held	Rechallenge	WHO-UMC category, and the deciding criterion	Kramer axes that move	NCI attribution
Agent X	Reasonable	Improved	Not given	Probable/Likely: unlikely to be attributed to disease or other drugs; rechallenge not required	Alternative candidates, dechallenge, previous experience	Probable
Carboplatin	Reasonable	Improved	No recurrence	Unlikely: disease or other drugs provide plausible explanations	Rechallenge, alternative candidates	Unlikely
Pemetrexed	Reasonable	Improved	No recurrence	Unlikely: disease or other drugs provide plausible explanations	Rechallenge, alternative candidates	Unlikely

Timing fits both backbone drugs, so Unlikely rests on the other criterion: Agent X explains the event, and resuming both drugs without a recurrence argues against them. WHO-UMC's explanatory text allows Unlikely when "another explanation is more likely," but an assessor who weighs timing first would keep both at Possible. On the NCI scale, only Agent X counts as related. A response to corticosteroids is consistent with an immune-mediated cause, and it means the improvement after the hold is not a clean dechallenge for any single drug.

Agent X does not reach Certain: WHO-UMC reserves that category for an event that is "definitive pharmacologically or phenomenologically," which a transaminase rise is not. Had all three drugs resumed together with ALT staying down, nothing would separate them, and each would likely stay at Possible. The order in which drugs come back is evidence. The treatment decisions belong to the protocol and the investigator; the example shows only what each one leaves behind for attribution.

How Burna AI scores attribution per drug

Burna AI scores each adverse event for attribution against every drug in the regimen, using WHO-UMC and Kramer, with the reasoning shown. The engine expands a named regimen into its component drugs from pre-built oncology regimen profiles, and interactions between the drugs in the regimen are weighed in the assessment. A protocol on file is read into structured form, so grading runs against what the protocol actually names. See per-suspect attribution with the working shown.

Every grade traces to the line in the medical record that produced it, and a grade whose citation set is incomplete is flagged before it is signed. The work runs through twelve specialized agents in a cascading constraint pipeline, the architecture on which two patents have been filed.

The investigator accepts or overrides each score. A grade becomes a record when a qualified investigator signs it with a 21 CFR Part 11 electronic signature, and the signature carries what was accepted, what was overridden and the citation set as signed. Each of these is listed in the public capability record. AI suggests, clinicians decide. Human-in-the-loop, always.

A sponsor reads graded safety data for its own trials, in a view separate from the site's. Per-drug attribution recorded at the encounter, with its reasoning, helps elucidate the safety profile while the trial can still act on it.

Common questions about combination therapy attribution

Who attributes an event to each drug in a combination trial?

The investigator. Under 21 CFR 312.64(b), an investigator reporting a serious adverse event must include "an assessment of whether there is a reasonable possibility that the drug caused the event." NCI's guidelines place primary responsibility for attribution with the clinical investigators and ultimately the protocol's principal investigator. Where the case report form asks for it, the investigator records a separate attribution for each study drug, and the sponsor makes its own determination for IND safety reporting.

Can more than one drug be related to the same adverse event?

Yes. Adverse event relatedness is assessed for each suspect drug on its own evidence, and none of these methods divides an event between drugs. On the NCI scale, any drug rated Possible or above counts as related. In an individual case safety report, ICH E2B(R3) carries relatedness for each drug and reaction pair in its drug-reaction matrix (ICH E2B(R3) implementation guide).

How is an event attributed to a first-in-human agent with no adverse event history?

An event under a first-in-human agent is attributed by elimination, using the investigator's brochure and the protocol. What the established drugs' known profiles and timing explain is set aside, and what remains is weighed against the new agent's pharmacology and nonclinical findings. Under WHO-UMC, Certain requires an event that is definitive pharmacologically or phenomenologically and, where needed, a satisfactory rechallenge; the WHO-UMC document states that "in practice few adverse reactions are 'certain' or 'unlikely'." A novel agent therefore usually lands at Possible or Probable.

What if WHO-UMC and Kramer disagree for the same drug?

Record both results and the criterion that separates them. The categories are not designed to be averaged or converted into each other, and the investigator's signed call is what the record carries. See WHO-UMC and Kramer, explained.

If your combination trial needs per-drug attribution an investigator can read, check and sign, book a 15-minute conversation.