How to create a database
CSV file
A CSV file, known as a "comma-separated values" file, enables you to store data in a table-structured format. Data records are divided by line breaks (i.e., each data record starts a new line).
Note
Creating CSV files is most straightforward when using a spreadsheet application, such as Microsoft Excel.
Database generation using Excel
Step 1: Open a new spreadsheet in Microsoft Excel. Any spreadsheet software can be employed to convert data to CSV files.
Step 2: Specify the headers (one for the names of each data point, one for each descriptor, and one for the target values) in row 1 at the top of the spreadsheet. More information regarding the selection of appropriate descriptors is available in the following section.
Step 3: Enter your data into the spreadsheet under each column as needed. For example, write the compound name in cell A2, the descriptor values in cells B2, C2 and D2, and the target values in cell E2.
Step 4: Go to the "File" tab and select "Save As". Select "CSV" under the "Save as type" dropdown menu and input a title for your CSV file. Then, click the "Save" button to generate your CSV file.
Selection of descriptors
In machine learning, a "descriptor" refers to a specific feature or property of the data used to describe or represent an entity. In the context of chemistry, descriptors are characteristics of molecules or complexes that are used to define their structures, behaviors, or other relevant attributes. There are no strict limitations in defining descriptors, but ideally, descriptors should encompass relevant properties that assist machine learning models in differentiating between compounds within the database. Descriptors can be specified using experimental measurements (i.e., density, melting point, etc.) and theoretical methods (i.e., DFT properties, NBO charges, etc.).
When creating a database, users have two approaches:
1. User-Defined Descriptors
Users can utilize their own descriptors, whether obtained experimentally or through computational methods.
2. Automated: SMILES to Descriptors
If users don't have pre-calculated descriptors, they can use the automatic generation of descriptors with the AQME module.
User-defined descriptors
If users want to use descriptors that have been measured or calculated before, they have to enter them manually into the database. Researchers can use experimental or theoretical descriptors, or a mix of both to define their problem. For example, in the article J. Am. Chem. Soc. 2022, 144, 9586, the authors wanted to correlate the equilibrium energies (ΔG) obtained with 20 different solvents using a machine learning model.
In this case, the authors needed to differentiate the solvents using properties/descriptors (defining the problem). To achieve this, they constructed a set of 17 molecular descriptors that captured electronic and structural differences between the solvents. These descriptors included the dielectric constant (ε), the second COSMO σ-moment, which characterizes a molecule's overall electrostatic polarity (Sig2), and the McGowan molar volume (V). This is how the database was structured:
Automated: SMILES to descriptors
When users don't have previously measured/calculated descriptors, they can employ AQME to generate molecular and atomic descriptors. Let's consider the article J. Chem. Inf. Comput. Sci. 2004, 44, 1000, which includes a database with SMILES strings and aqueous solubility values in its electronic supporting information. A solubility predictor can be developed using this database, as it includes columns with the names of the molecules, SMILES strings of the molecules, and the property to be predicted (aqueous solubility in this case). The titles of the first two columns must be adjusted to "code_name" and "SMILES", respectively. This is how the database would be structured:
ROBERT will then be executed with one command line (see the From SMILES to predictors section and the Full workflow from SMILES example), generating results like these:
