Extended Data Fig. 8: Reverse inference classification with neural networks.

Neural network classifiers were trained to perform reverse inference, using brain activation coordinates to predict occurrences of mental function terms grouped by domains shown in Extended Data Fig. 4. Classification models comprised 8 fully connected (FC) layers, all with ReLU activation functions except the output layer which was activated by a sigmoid. The optimal learning rate, weight decay, number of neurons per layer, and dropout probability were determined for each framework through a randomized grid search. ROC curves are shown for the test set performance of classifiers with mental function features defined by b, the data-driven framework, c, RDoC, and d, the DSM. e-g, For each domain, the significance of the test set ROC-AUC was determined by a one-sided permutation test comparing the observed value to a null distribution, and the p value was FDR-corrected for multiple comparisons (* FDR < 0.001). Observed values in the test set are shown with solid lines. Null distributions (gray) were computed by shuffling true labels for term list scores over 1,000 iterations; the 99.9% CI is shaded, and distribution means are shown with dashed lines. Bootstrap distributions of ROC-AUC (colored) were computed by resampling articles in the test set with replacement over 1,000 iterations. h, Differences in bootstrap distribution means were assessed for each framework pair. While there were no differences between frameworks at the 99.9% confidence level, the data-driven framework had higher ROC-AUC than RDoC at the 99% confidence level (99% CI of the difference = [0.007, 0.050]), and higher ROC-AUC than the DSM at the 95% confidence level (95% CI of the difference = [0.0003, 0.049]). Solid lines denote bootstrap distribution means.