QSAR – Descriptor Generation, Classification Models, and Diverse Selection

[Andrew Orry]

Q: Can ICM generate numerical descriptors for QSAR modeling?

A: Yes. Go to Chemistry → Calculate Properties. There are several categories of descriptors:

• Individual (e.g., MolLogP, MolWeight, etc.)

• Grouped descriptors like Atom Counts, Bond Counts, and Topological Descriptors

These generate multiple columns in the table. You can mix numerical descriptors with fingerprints when building a model. For example, the MolSoft solubility model (MolLogS) was built using MolLogP + binary ECFP4 fingerprints.

Documentation: https://molsoft.com/gui/calculate-properties.html

---

Q: Does ICM support classification models (e.g., active/inactive)?

A: Yes. Use Random Forest for classification.
Ensure your training column contains integers (e.g., 1 = active, 0 = inactive). The "Learn" dialog will allow you to train a classifier using the selected descriptors.
ICM also supports neural network training via command-line in a special Linux package.

Documentation: https://molsoft.com/gui/learning.html

---

Q: How can I select a diverse set of compounds based on activity?

A: Use the clustering tool in the following way:

1. Select the activity column.

2. Choose Table → Clustering and check Keep Distance Matrix.

3. Use the slider to define the number of clusters.

4. Right-click on the table and choose Select Centers by…

5. In the dialog, select From Clustering Distance.
   This will select representative compounds from each cluster for a diverse output selection.

Documentation: