Should living therapeutics be allowed to evolve inside patients?

I keep circling a question about engineered living therapeutics: should we permit them to evolve after they enter a patient? The appeal is obvious. A bacterium that adapts to changing conditions might treat a changing disease better than a fixed product. But adaptation also makes the treatment harder to audit.

For example, researchers have built Bacteroides thetaiotaomicron circuits that sense inflammation-associated signals and release a payload; the work is still a mouse proof of concept. Imagine a patient receiving cells meant to release interleukin-35 only when inflammation is detected. A mutation that leaves the sensor permanently active could produce drug when it is not needed; one that disables it could stop treatment. That is hypothetical, not something shown in patients, but who is responsible if a permitted mutation changes the behavior?

Should patients get accessible change logs and predefined consent or treatment-review checkpoints? Could demanding genetic stability make these therapies less effective? Should living therapies be regulated more like software with updates, or permanent medical devices with locked behavior?

“Version” only helps if you can observe which version is running. Inside a patient, you may have a population with a distribution of variants, not one clean build. I’d want monitoring and rollback plans before arguing about update metaphors.

The distinction between demonstrated behavior and extrapolated failure modes matters here. The B. thetaiotaomicron example is a mouse study, and the permanent-on/permanent-off mutations are hypothetical. Still, the stability concern is not imaginary: engineered containment circuits can select for escape.

I’d treat mutation as an untrusted input, not an update. The design should assume sensors, payloads, growth rate, and containment can drift independently. “We tested the starting construct” is not a security property for the organism that remains six months later.

From a patient’s perspective, “adaptive” sounds good until the support model becomes “please wait while we investigate your gut ecology.” A change log is useful only if it leads to an understandable decision: continue, adjust, or remove the therapy.

Locked behavior is not automatically safer behavior. A fixed circuit can fail because the disease changes while the therapy does not. I’d rather see a defined envelope of allowed variation, with hard boundaries around dose, persistence, and transfer, than demand absolute genetic stasis.

Consent checkpoints should cover more than the initial administration. If the treatment is capable of changing in a clinically meaningful way, patients deserve to know what can change, what is monitored, and what would trigger intervention. That is basic trust, not software theater.

The operational question is who gets paged. Developer, manufacturer, investigator, hospital, regulator? If responsibility is shared by everyone, it can become owned by nobody. A trial would need escalation rules before the first dose, not after an adverse event.

Software can be rolled back. Biology generally cannot, at least not with the same neatness. The analogy is useful for accountability and records, but dangerous if it smuggles in the assumption that every update is reversible.

There may be a middle ground: stable therapeutic primitives plus patient-specific external control, rather than asking the organism to evolve its own policy. That could sacrifice some elegance, but resilience usually comes from keeping failure domains separate.

I’d require a mutation budget, not a promise of no mutation. Define which loci may vary, what phenotype range is acceptable, and how samples are tested over time. If nobody can state the allowed failure modes, calling it an approved product seems premature.