Background: The autism spectrum disorders (ASD) are common neuropsychiatric conditions of childhood for which the vast proportion of population risk is attributable to inheritance, and for which there exist few if any replicated biomarkers. Main body: This commentary summarizes a set of recent studies involving identical (monozygotic, MZ) twins which, taken together, have significant implications for the search for biomarkers of inherited susceptibility to autism. A first is that variation-in-severity of the condition (above the threshold for clinical diagnosis) appears more strongly influenced by stochastic/non-shared environmental influences than by heredity. Second is that there exist disparate early behavioral predictors of the familial recurrence of autism, which are themselves strongly genetically influenced but largely independent from one another. The nature of these postnatal predictors is that they are trait-like, continuously distributed in the general population, and largely independent from variation in general cognition, thereby reflecting a developmental substructure for familial autism. A corollary of these findings is that autism may arise as a developmental consequence of an allostatic load of earlier-occurring liabilities, indexed by early behavioral endophenotypes, in varying permutations and combinations. The clinical threshold can be viewed as a “tipping point” at which stochastic influences and/or other non-shared environmental influences assert much stronger influence on variation-in-severity (a) than do the genetic factors which contributed to the condition in the first place, and (b) than is observed in typical development. Conclusion: Biomarkers identified on the basis of association with clinical symptom severity in ASD may reflect effects rather than causes of autism. The search for biomarkers of pathogenesis may benefit from a greater focus on traits that predict autism recurrence, among both clinical and general populations. In case–control studies, salient developmental liabilities should be systematically measured in both cases and controls, to avoid the erosion in statistical power (i.e., to detect differences) that can occur if control subjects carry sub-clinical aggregations of the same unmeasured traits that exert causal influences on the development of autism.