EGFR tyrosine kinase inhibitors (TKIs) have altered the natural history of EGFR-mutant non-small-cell lung cancer, yet acquired resistance remains almost inevitable.[1] Persistent therapeutic pressure reshapes tumour ecology: sensitive clones are suppressed, resistant clones undergo competitive release, and the latter gradually become the dominant population.[1] This process bears a subtle resemblance to a classic problem in artificial-intelligence training. Reinforcement learning can enable a model to maximise a specified reward without ensuring that it has learned the objective humans actually intended.[2] Similarly, anticancer treatment can minimise tumour burden without preventing the tumour from maximising its own fitness within the newly imposed selective environment.

The analogy may appear abstract, but it exposes a question in cancer treatment that has long been underestimated: are we treating the tumour, or training it? Conventional treatment emphasises maximal tumour-cell killing, aiming for rapid radiographic shrinkage, prompt declines in tumour markers, and early symptom relief. From an evolutionary perspective, however, treatment is also a selective pressure. Every exposure changes the relative survival prospects of competing clones within the tumour. The more effective the treatment, the more rapidly it may eliminate sensitive clones; the cells left behind may be precisely those best adapted to the current drug environment.

If a tumour is an evolving ecosystem, the objective of treatment should extend beyond reducing tumour burden to reshaping the tumour fitness landscape. Adaptive therapy is the clearest clinical expression of this idea.[3] Rather than treating eradication as the sole objective, it seeks—while tumour burden remains controllable—to preserve a population of drug-sensitive clones that can continue to compete with resistant clones.[4] In an adaptive abiraterone study in metastatic castration-resistant prostate cancer, for example, treatment was started and stopped dynamically according to changes in prostate-specific antigen, with the aim of delaying expansion of resistant clones by retaining sensitive competitors.[4]

At a simplified level, this interaction can be considered as a Nash-like equilibrium between sensitive and resistant clones. Under some treatment conditions, the system need not move inexorably towards complete replacement of one population by the other; a controllable zone of coexistence might instead be possible. Within this zone, sensitive clones remain sufficiently abundant to constrain resistant-clone expansion, while the total tumour burden remains clinically controlled. The therapeutic task is then neither simply to intensify pressure nor to remove it completely, but to recalibrate it—through dose reduction, intermittent treatment, combination therapy, altered sequencing, local ablation of dominant resistant lesions, or closer dynamic monitoring.

The more difficult problem, however, is not imagining this equilibrium but detecting it. Imaging captures the resulting tumour burden; serum tumour markers provide only a relatively coarse account of its trajectory. Circulating tumour DNA might offer earlier clonal signals: whether resistance mutations are emerging, whether the original driver clone remains dominant, and whether the relative abundance of different clones is changing directionally.[5,6] Yet this measure remains imperfect. Low tumour burden, limited DNA shedding, isolated brain metastases, and localised progression can all reduce its sensitivity.[3] The future priority is therefore not only to develop more potent drugs, but also to establish monitoring systems capable of reading the tumour ecosystem dynamically.