A gradient-boosting ensemble beats Jev 1.13 on five CYP assays

This topic was created 8 days ago, and the information it contains may have evolved or changed since then.

On 12,563 molecules, an Insilico Medicine chart puts a gradient-boosting ensemble above Jev 1.13 on five CYP assays, and Jev's probabilities stop at 0.69.

Mikołaj Mizera posted a chart on October 2, 2026, and tagged Insilico Medicine. The footer reads “MMAI Gym · Insilico Medicine” and “12,563 molecules.” The subtitle says an uncalibrated specialist trained on MMAI Gym data outcalibrates Jev 1.13.

Five CYP assays

The assays are 1A2, 2C9, 2C19, 2D6, and 3A4. AUROC for the gradient-boosting ensemble is 0.94, 0.89, 0.89, 0.89, and 0.90. Jev 1.13 is 0.76, 0.73, 0.60, 0.78, and 0.67. A dashed line marks random chance at 0.50.

Balanced accuracy, in the same order, is 0.87, 0.79, 0.83, 0.74, and 0.81 for the ensemble, and 0.65, 0.55, 0.49, 0.67, and 0.57 for Jev. The post’s range, 0.74 to 0.87 against 0.49 to 0.67, is that second span. Jev’s lowest balanced accuracy on the chart is 0.49, on 2C19, just under the chance line.

Where Jev’s probabilities stop

The left panel draws one calibration curve per assay and pools the histogram. Jev’s curves are annotated “Jev tops out at 0.69.” The ensemble is the green series, labeled as trained on MMAI Gym data with no calibration step.

Its curves keep rising toward the top right. The chart does not print a p-value or an interval around either series.

The post has no paper link and no repository. The image is the measurement.

This site's reading

Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.

Verify

Mikołaj Mizera, @AjMizera, October 2, 2026, 09:12 UTC, status 2105949080166019486. The post tags @InSilicoMeds. The attached chart is the card saved for this page. The footer reads MMAI Gym, Insilico Medicine, and 12,563 molecules. The model line on the chart is Jev 1.13.

AUROC for the gradient-boosting ensemble is 0.94, 0.89, 0.89, 0.89, and 0.90 on 1A2, 2C9, 2C19, 2D6, and 3A4. Jev is 0.76, 0.73, 0.60, 0.78, and 0.67. Balanced accuracy is 0.87, 0.79, 0.83, 0.74, and 0.81 against 0.65, 0.55, 0.49, 0.67, and 0.57. The calibration panel marks "Jev tops out at 0.69."

The post does not link a paper or a repository. We did not rerun the assays.

Compare

This is a five-assay chart on molecules. The public 231 and the v1.5.4 composite are different tests. The chart's model line is Jev 1.13.

DimABSA also fits a frozen Jev and then reports error. That paper prints an RMSE. This chart prints AUROC and balanced accuracy, and it does not print a significance test.

Terms

MMAI Gym: The training set named on the chart. The footer says 12,563 molecules. The post does not link a download or a paper.
0.69: The predicted-probability ceiling marked on Jev's calibration curves. The ensemble curves continue toward 1. The chart says the ensemble had no calibration step.