Building a dataset for modelling
A model learns from two things in your dataset: the structure of each molecule and the value in the column you train on. Everything else in the row, including which assay or source a value came from, is invisible to it. This page covers how to set up the columns you train on and how to decide which measurements to put in them.
Property types and units
When you import a file, each column is given a property type (for example LogD, Caco-2 or Other potency/activity) and, for most numeric types, a unit. We suggest both from the column header, and you can change them in Import and rename columns or later in Edit columns.
They describe your data; they don't change the model. The property type and unit label the column and decide which units the app offers for it. Models are trained on the numbers in the column as they are: changing the type or unit does not convert any values, or change how a model is trained or scored.
If nothing fits, choose the closest type and set the unit to Other. For a measurement without its own entry, such as a pMIC, pick the closest property type (Other potency/activity, or Cell-based Assay for a whole-cell readout) and set the unit to Other. The model is unaffected; the app just won't show a unit label for that column.
How values are read
Each cell is read as one number. Text around the number is dropped, so 12 nM is read as 12. Qualifiers are dropped too: >64 is read as 64 and <0.1 as 0.1, and the model treats them as exact measurements. If many of your values sit at an assay limit (inactive compounds reported as >64, say), consider removing them, or recording each compound as 1 (active) or 0 (inactive) and training a classification model instead.
Use one scale per column. Values in a column are compared directly, so they must share a unit and a scale. Don't mix µM with nM, or IC50 with pIC50. Convert everything before import, and if you convert from mass units such as µg/mL, use the same method for every compound.
Log quantities are modelled as they are. pIC50, pMIC, LogD and other values already on a log scale are used unchanged, and MAE is reported in those units. Raw concentrations that span more than two orders of magnitude are modelled on a log scale; see How models are evaluated.
One consistent assay, or more data?
A common decision is whether to train on a smaller set of molecules all measured in one assay, or a larger set pooled from several sources: different labs, publications, protocols, cell lines or strains.
The model can't tell sources apart. It only sees structure and value, so a difference between two assays looks to it like a difference between molecules. If each source also covers different chemistry, the model can learn an assay offset as though it were a structure–activity relationship.
Each option trades one problem for another.
| One consistent assay | Pooled from several sources | |
|---|---|---|
| Values | Directly comparable | The same compound can differ several-fold between labs |
| Chemistry covered | Narrower | Broader |
| Most reliable on | Close analogues, measured in that assay | A wider range of chemistry, with more noise |
| How it is scored | Cross-validated if it has fewer than 100 measurements | More likely to get temporal validation |
Try both, and compare them on the same molecules. Which works better depends on your data and on what you will use the model for, so there is rarely a way to know in advance. The scores shown in the app can't settle it. Each is measured on different molecules, and often by a different method: a cross-validated score on 90 molecules and a temporal score on 300 are not comparable. Instead, set aside the same molecules from both datasets before training, ideally recent compounds measured in the assay you care about and close to the chemistry you plan to make. Then predict them with each model and compare the predictions with the measured values.
Check the overlap before pooling. If some molecules were measured in more than one source, compare their values. Agreement within the assay's usual reproducibility suggests pooling is safe. A consistent offset or a wide scatter means the pooled model will inherit it.