Figure 3.

The clinical predictive model and mechanistic model were constructed separately for atomoxetine and citalopram, against the principal benchmark of 30% behavioral improvement. (A) The model parameters were optimized using a “grid‐search” algorithm, which searches across exponentially growing sequences of C and γ to maximize the cross‐validation accuracy of a given model. The illustrated example used data from the clinical model of atomoxetine response. Colors indicate cross‐validation accuracy values. (B) Cross‐validation accuracy of the models that were optimized for the principal benchmark and for alternative benchmarks (e.g., 10–50% behavioral improvement, see Table 2 and Supporting Information, Tables S3 and S4 for details). (C) Robustness of the optimal models for the principal benchmark was measured as cross‐validation accuracy of the models when tested against alternative benchmarks.