The origins of childhood cancer

Abstract Background Treatment success in childhood cancer has reached a plateau. Aetiological research remains at the level of molecular inventories. The first fact is that the window for mutation accumulation is measured in months. The second fact is that no recognisable multi-step precancerous sequence exists in the general paediatric population. The third fact is that tumour histology follows developmental programmes. Existing models answer where the field is, what it looks like and which lineage it comes from. These models do not provide a measurable state. That state should distinguish cells that will complete transformation from cells that will not. Methods We built a five-state latent-variable developmental model. The P state is persistent fetal developmental-sequence deviation. The G state is failure of growth-phase epigenetic correction and gate failure. The C state is a gradient-cell clone. The T state is a clinical tumour. The R state is a regressed or silent clone. The model was calibrated to age-specific incidence for five developmental-origin subtypes in SEER 2008–2012. The five subtypes were neuroblastoma, retinoblastoma, Wilms tumour/renal tumour, hepatoblastoma and embryonal central nervous system tumour. The model used autopsy silent-clone data to anchor P and R. The model used a single-branch mutation-accumulation model as a comparator. The model simulated the effect of lowering the gate-failure rate or raising the clone-regression rate within the developmental window. Results The model reproduced age-specific incidence for the five subtypes. Observed and predicted peak groups agreed in four of five subtypes. The observed peak for embryonal central nervous system tumour was < 1 year. The predicted peak for this subtype was 1–4 years. The two peaks differed by one age group. The predicted composite incidence was 107.6/79.8/16.6/5.2/2.3 per million. The observed composite incidence was 115.9/65.0/16.6/5.6/2.6. The latent-variable model had a lower AIC than the single-branch model in five of five subtypes. ΔAIC was 3.1–416.9. The clone-to-clinical ratio was 37.6–72. The model value for neuroblastoma was 0.66% clone and 0.018% clinical. Autopsy in situ lesions were 1/150–1/250. Clinical neuroblastoma was 1/7,000. The model ratio was in the same direction as the autopsy ratio. A 50% rise in kR reduced cumulative clinical incidence by 24.4%–26.7% across the five subtypes. A 50% fall in kG0 had an effect of 0%–21.1%. Conclusions The rate of mutation accumulation cannot produce the early peak of childhood cancer. The single-branch model cannot reproduce the peak groups of the five subtypes. The entry point of childhood cancer is persistent fetal developmental-sequence deviation, growth-phase correction failure and gate failure. The clinical tumour is the terminal readout of this path. The clone-regression rate within the window is an earlier intervention variable than the mutation rate.