Descriptors

Empirical Approaches vs. Simulation-Based Approaches

There are many questions in chemistry that cannot be answered by simulation, or for which simulation is poorly effective. For instance, imagine trying to predict the effect of ligand structure on palladium ligation state: which ligands will lead to L•M complexes, and which ligands will lead to L₂•M complexes? Using simulation, one could model all conformers of the two states, apply corrections for free energy and for solvent effects, compute ∆G, and compare it to the experimental value. But solvation free energies are poorly predicted by DFT, and predicting the resting state of a complex with many low-energy aggregation states remains too challenging to accurately simulate routinely.

Instead, one could imagine a different approach: rather than simulating the entire system, one can generate descriptors for each compound in question and see if statistical approaches or machine learning can find a relationship between these descriptors and the output variable. This approach has, in fact, been applied to the specific task of predicting ligation state and found to work quite well!

There are real disadvantages to this cheminformatics-based method: it requires a training set of labeled examples with which to find a relationship, making it challenging in cases where very little data is present. On the other hand, in principle any chemical relationship can be learned this way! Unlike conventional simulation, where a new computational workflow must be built and benchmarked for each property, descriptor-based methods can be used even when no workflow exists, and for problems too complex to simulate (e.g. in vivo toxicity).

Choosing Descriptors

The above discussion purposely avoids specifying the nature of the descriptors. In reality, there are a number of competing descriptor schemes out there:

• Molecular fingerprints, which turn 2D graph-based representations of molecules into bit vectors, are often used.
• Cheminformatic descriptors, like number of hydrogen bonds or fancier metrics, are also very commonly used.
• Quantum mechanics-based descriptors can be very helpful, as recent work from Green and co-workers shows, but are somewhat more expensive to compute.
• And learned molecular representations have attracted considerable attention in recent years.

Rowan's descriptor workflow automatically calculates over 1800 cheminformatic descriptors using mordred-community and supplements these with quantum mechanics-based descriptors from xTB. (xTB-based charges are similar in quality to DFT-based charges, but are several orders of magnitude faster to compute, making them appealing for high-throughput workflows.)

For a full list of all descriptors in Rowan, see the bottom of this page.

Visualizing Descriptors in Rowan

Using Rowan's descriptor workflow, it's simple to compute hundreds of descriptors for a set of molecules: just paste in the SMILES strings and select the descriptor workflow. The output data can be viewed as a table, searched by descriptor, or downloaded as a .csv file for further processing.

If multiple descriptor calculations are selected, all descriptors can be viewed for each compound in a spreadsheet-like view. Principal component analysis (PCA) can be performed with the click of a button, making it simple to visually observe which compounds are similar and which are different. This self-supervised approach is often used to visualize chemical space. Both the bulk descriptor results and the PCA coordinates can also be downloaded as .csv files for further processing.

Descriptor Index

Here's a full list of all the descriptors computed in Rowan (many of these descriptions are originally from Mordred's documentation):