Why Treatment Order Can Change the Outcome

Consider two instructions: apply a primer, then paint; or paint, then apply a primer. The materials have not changed. The quantities may be identical. But the second step meets a different surface, so the finished result is different.

Treatment sequences can behave in the same way.

If treatment A is given before treatment B, B does not act on the original biological condition. It acts on the condition left behind by A. Reverse the order and A inherits the changes produced by B instead. The two plans contain the same components, but they are not necessarily equivalent.

Researchers sometimes describe this as a noncommutative treatment effect. In plain language, it means that swapping the order can change the result:

A followed by B may produce a different outcome from B followed by A.

This is not a mathematical curiosity. It changes how combinations should be screened, how evidence should be interpreted and how treatment strategies should be designed.

The first treatment changes the problem faced by the second

Biological systems do not remain fixed while a treatment is applied. An intervention can alter signalling activity, target availability, metabolism, immune behaviour, cell-cycle position or the composition of a mixed cell population. Some changes are brief. Others persist after the first treatment is withdrawn.

The important point is not that every treatment produces every kind of change. It is that a treatment can leave a biologically meaningful history.

That history may make the next intervention:

• more effective because a relevant pathway or population has become more vulnerable;
• less effective because the system has adapted, compensated or selected for resistance;
• more toxic because the first exposure altered clearance or reduced the capacity to recover; or
• simply different because the second treatment now acts in another biological context.

Order, timing and recovery are therefore not administrative details around a treatment. In some settings, they are part of the treatment itself.

Why a combination label can hide the difference

Many datasets describe exposure using a compact label: treatment A, treatment B or combination A+B. That is useful when order is irrelevant. It is incomplete when the system carries memory.

Imagine that one group receives A followed by B and another receives B followed by A. If both groups are recorded only as having received "A+B," a sequence-specific benefit in one group and a sequence-specific harm in the other can be averaged together. The combined result may look small even when order has a large effect.

This is where a fixed average treatment effect can miss the important structure. The statistical estimate is not necessarily wrong; it may be answering a treatment question that was defined too broadly. If A followed by B and B followed by A are treated as distinct strategies, established causal and trial methods can compare them. If treatment history is collapsed, no downstream analysis can recover information that was never represented.

The issue is therefore upstream of model choice:

What, exactly, counts as the treatment?

For a history-dependent system, "received A and B" may not be a sufficient answer.

Public research shows that sequence effects are real—and context-dependent

Sequence dependence has been observed across several areas of biomedical research. The evidence does not imply that every combination needs a special schedule, or that a result in one model transfers directly to patients. It does show why order should be tested rather than assumed away.

In a 2012 preclinical cancer study, researchers screened targeted inhibitors and DNA-damaging drugs across different doses, timings and orders. In a subset of triple-negative breast-cancer cell models, giving an EGFR inhibitor before a genotoxic drug produced substantially more cell death than simultaneous administration. The study linked the result to treatment-induced changes in apoptotic signalling. This was a laboratory finding, not proof of clinical benefit, but it demonstrated the central principle clearly: the first exposure changed the response to the second. Lee et al., Cell (2012)

Experimental antibiotic research provides another example. A 2015 study tested 136 sequential regimens in an E. coli laboratory model. Five sequences cleared the bacterial populations at the tested sublethal doses across all replicates, while the equivalent two-drug combination treatment was ineffective. The direction of the switch mattered because adaptation to one antibiotic could change sensitivity to the next. These were in-vitro results and cannot be translated directly into prescribing decisions, but they show how treatment history can reshape the target population. Fuentes-Hernandez et al., PLOS Biology (2015)

Clinical evidence is more difficult—and appropriately more cautious. In the phase 3 OBELICS trial, 230 people with metastatic colorectal cancer were randomised to receive bevacizumab either on the same day as oxaliplatin-based chemotherapy or four days before it. The sequential schedule did not improve the primary endpoint of objective response rate. It was associated with longer overall survival and some better safety and quality-of-life outcomes, but the authors treated those secondary findings as a basis for further study rather than definitive proof. The trial is valuable precisely because it shows both sides of the problem: schedule is a legitimate experimental variable, but plausible sequence effects still require endpoint-specific clinical validation. Avallone et al., JAMA Network Open (2021)

The missing capability is a history-aware view of treatment

The standard response to combination complexity is often to test more pairs, add more covariates or fit a more flexible predictor. Those steps can help, but they do not solve a treatment definition that ignores history.

A history-aware approach asks a different set of questions:

• What condition does the first intervention leave behind?
• Which parts of that change are likely to persist until the next intervention?
• Does the second treatment encounter sensitisation, adaptation or recovery?
• At what interval does a useful ordering effect disappear or reverse?
• Which outcomes—efficacy, toxicity, resistance or recovery—are affected?

These questions turn a combination from a bag of components into a path through a changing system.

That path can become large very quickly. With two treatments there are already two possible orders, before considering dose, interval and repetition. Add a third treatment and the number of plausible schedules grows sharply. Exhaustively testing every schedule in every biological context is rarely practical.

This is where mechanistic software can be useful: not as a substitute for experiments or clinical trials, but as a way to organise the search. A model that represents treatment history explicitly can help identify which sequences are meaningfully different, which assumptions drive that difference and which experiments would be most informative next.

The Mondren perspective

At Mondren, we start from a simple principle: when actions change the conditions inherited by later actions, order must be represented as part of the system.

For biotechnology, that means developing mechanistic software that can reason about treatment history, interaction and change over time at a level appropriate to the available evidence. The public claim is deliberately limited. Such a system can structure hypotheses, compare candidate strategies and support experimental design; it does not establish that a sequence is safe or effective in patients.

The distinction matters. A useful research system should make the assumptions and limits of a sequence hypothesis easier to inspect. The final evidence must still come from suitable laboratory studies, external validation and, where clinical use is intended, properly designed trials.

What this changes for development

Treating order as a first-class variable can change several parts of the development process.

Combination screening becomes sequence screening. Candidate therapies can be tested not only together, but in both orders and across biologically meaningful intervals.

Trial arms become treatment strategies. A followed by B and B followed by A can be specified as different interventions, with outcomes selected to capture efficacy, toxicity and durability rather than a single short-term response.

Data capture includes treatment history. Dates, intervals, prior exposures, washout and recovery become essential context rather than optional metadata.

Decision support becomes state-aware. The relevant question is no longer only "Which treatment performs best on average?" It is also "Given what has already happened, what should be considered next?"

Existing therapies may be used more intelligently. In some cases, value may come not from discovering another component, but from identifying a better order, interval or stopping rule for components already available.

What remains difficult

Sequence-aware modelling does not make biological uncertainty disappear.

An effect observed in a cell line may not survive in an animal model. An effect observed in one cohort may not transfer to another. Patient history may be measured incompletely. The interval between treatments may matter as much as the order. Benefits on one endpoint may coexist with harms on another. Observational records can also confound treatment order with disease severity, clinician judgement and access to care.

Most importantly, order dependence should not be assumed merely because it is possible. A followed by B may, in fact, be equivalent to B followed by A for a given context and outcome. That is an empirical question.

The goal is not to declare every treatment sequence unique. It is to avoid declaring sequences equivalent before the evidence justifies it.

Treatment is a path, not a list

Biology carries history. Each intervention can change what the next intervention encounters, and those changes can alter efficacy, toxicity, resistance and recovery.

When that happens, the components alone do not fully define the treatment. The path matters.

For research and development teams, the practical implication is straightforward: record the order, test the order and model the order. A treatment strategy should describe not only what is given, but what comes first, what comes next and what condition is likely to exist in between.

That is the difference between analysing a list of interventions and understanding a changing system.

---

This article discusses research and development principles. It does not provide medical advice or recommend any clinical treatment sequence.