GeinQSAR Studio
Home › The library dashboard

The library dashboard

The designed molecules as structures, as a scatter of activity against a property, as Pareto fronts, and as the ranked table behind all three.

The dashboard opens when a design finishes, and again whenever the workflow is run. It has four tabs, Library, Chart, Pareto and Table, and one Export report button at the right of the tab row. Every view reads the same ranked table, so the structure grid, the charts, the exported files and the report cannot disagree.

The status bar counts the molecules, how many lie inside the applicability domain and how many were selected for diversity. The inspector on the right names the scaffold, the predicted response, whether the interval is calibrated or not calibrated, the two domain counts, the number of Pareto fronts and the number of diverse picks.

Library

The Library tab, best first. The heading of the prediction column names the model column and the coverage, and each cell gives the predicted value with its calibrated 80 % interval. Read the three together: the ranking separates these analogues by about a tenth of a log unit, while each interval is more than two log units wide, so the order is a suggestion of what to make first and not a claim that the first is the best.
The Library tab, best first. The heading of the prediction column names the model column and the coverage, and each cell gives the predicted value with its calibrated 80 % interval. Read the three together: the ranking separates these analogues by about a tenth of a log unit, while each interval is more than two log units wide, so the order is a suggestion of what to make first and not a claim that the first is the best.

The structure grid shows the molecules in the order of the ranking, with the rank, the identifier, the prediction and the domain verdict under each structure. The heading of the prediction column is the column the model writes, followed by the coverage of the interval when one was calibrated, and each cell then reads as the value with its interval, for example 7.12 (6.01 to 8.23). The domain column reads In domain or Outside domain.

The same tab for a design run without a calibration set. The heading carries no coverage, the cells no interval, and the line above the grid says why and what to supply. The domain verdict is the column after the prediction. A bare predicted value is the less honest of the two displays, which is why the Interval page exists.
The same tab for a design run without a calibration set. The heading carries no coverage, the cells no interval, and the line above the grid says why and what to supply. The domain verdict is the column after the prediction. A bare predicted value is the less honest of the two displays, which is why the Interval page exists.

Without a calibration set the tab states it: No calibrated prediction interval: supply a calibration set of measured compounds the model was not fitted on. A design that produced no molecule says that too, rather than showing an empty grid.

Chart

The Chart tab: predicted activity against molecular weight. Green points are inside the applicability domain, red ones outside it. Look at where the red points sit: the two highest predictions of this library are both extrapolations, which is the usual shape of the problem.
The Chart tab: predicted activity against molecular weight. Green points are inside the applicability domain, red ones outside it. Look at where the red points sit: the two highest predictions of this library are both extrapolations, which is the usual shape of the problem.

Property chooses what the predicted activity is plotted against, from every computed property of the library. The two series are Inside domain in green and Outside domain in red, the colour convention used throughout the program: green for favourable, red for unfavourable. A selection in the chart is a selection in the other tabs.

A design ranked on computed properties alone has no predicted activity to plot, and the tab says so rather than drawing an empty pair of axes.

Pareto

The Pareto tab of the same library ranked by fronts, predicted potency against molecular weight. The colour bar is the front number, green for front 1 and red for the last: a molecule on front 1 can only be improved in one objective by giving up the other. The subtitle under the chart title says so, so the figure can stand alone.
The Pareto tab of the same library ranked by fronts, predicted potency against molecular weight. The colour bar is the front number, green for front 1 and red for the last: a molecule on front 1 can only be improved in one objective by giving up the other. The subtitle under the chart title says so, so the figure can stand alone.

The tab is drawn when the ranking was by Pareto fronts and at least two objectives vary. The colour bar is the front number, and the subtitle says what a front is. A library ranked by desirability, or with one objective only, says so instead: the fronts of a single objective are its sort order.

Which of the two rankings to use. Desirability gives one list, which is what a synthesis plan needs, but the weights that produce it are a judgement you have to defend. Pareto fronts make no such judgement and leave the choice among the non-dominated molecules to the chemist. Reporting both is cheap: the same run produces either, and the ranked table carries the front number and the desirability side by side.

Table

The Table tab with the whole ranked library, and Export CSV and Export SD above it. Every value the run computed is a column here, which is what makes a ranking checkable. The R1 column names the group that was put at the first attachment point, so a row can be read as a chemical proposal rather than as a number.
The Table tab with the whole ranked library, and Export CSV and Export SD above it. Every value the run computed is a column here, which is what makes a ranking checkable. The R1 column names the group that was put at the first attachment point, so a row can be read as a chemical proposal rather than as a number.

The table holds, in this order: the rank, the identifier and the SMILES; one column per attachment point, headed R1, R2 and so on, naming the substituent that was put there; for each model the predicted value, the two ends of its prediction interval, its standard error, whether the compound is in the domain and the reasons when it is not; the computed properties QED, synthetic accessibility, molecular weight, log P, polar surface area, rotatable bonds, donors, acceptors, heavy atoms, the fraction of sp3 carbon and the similarity to the training set; the PAINS and unwanted-group verdicts; the desirability; the Pareto front and the crowding distance when the ranking used fronts; whether the molecule is in the domain overall; and the diverse pick when a selection was made.

Export CSV and Export SD write the table and the structures. The CSV holds every column at round-trip precision; the SD file carries the results as data fields.

The report

Export report writes the report of the library as PDF, Word or Excel, through the dialog described in Reports and export. Besides the usual sections it carries a Designed library chapter with the whole ranked table, the verdict of every filter on every compound, the two charts, and a structure grid of the designed compounds in the order of the ranking, each with its rank, its predicted value and its interval, coloured green inside the applicability domain and red outside it. The grid shows the first 120 compounds and says so; the Excel workbook and the ranked table hold them all.

What the report does not do is turn a prediction into a measurement. The honest reading of a designed library is in Accuracy and validation: a model that predicts a measured test set to within a few tenths of a log unit predicts a designed analogue no better than that, and often worse, because the analogue is further from the training set than any test compound was.