liveBook · Manning

Molecular descriptor

As explained in Build AI Drug Discovery Pipelines

A molecular descriptor is a numerical or symbolic representation of a molecule's physicochemical, structural, topological, or quantum-mechanical properties.

A molecular descriptor represents a molecule through a property or structural feature that a computer can process. Descriptors can capture physicochemical, topological, structural, or quantum-mechanical information. They turn molecular information into features for filtering compounds, searching for similar molecules, or predicting properties. section 2.2.3 section 2.6

A SMILES string describes a molecule as text; a numerical descriptor gives a filter or model a quantity to work with. Different descriptors capture different aspects of a molecule, so the choice of representation affects what a model can learn. section 2.2.3

Common descriptor choices include:

  • Molecular weight, hydrogen bond donor and acceptor counts, and logP: properties useful for compound filtering. section 2.2.3
  • Topological and fragmental descriptors: representations of molecular connectivity or fragments. Fingerprints are fragmental descriptors used in similarity searching. section 2.2.3 section 2.6
  • Three-dimensional descriptors: features such as molecular shape and volume that can capture spatial information. section 2.2.3
Table 2.2 Taxonomy of molecular descriptors. Outside of molecular descriptors, there are also reaction descriptors and ensemble descriptors (mixtures or materials considering more than a single molecule).
Classification
Type
Examples
Locality
Local
Atom charge
Global
Molecular weight, lipophilicity
Dimensionality
1-D
Molecular weight, polar surface area, number of methyl groups
2-D
Topological descriptors, fragmental descriptors
3-D
Molecular shape and volume, quantum chemical descriptors
4-D
4D-fingerprints derived from conformers
Calculation Method
Physicochemical
Hydrophobicity, solubility
Topological Indices
Branching index (based on degree of each atom), chi molecular connectivity index (based on bond paths of different lengths)
Fragmental
Fingerprints (covered in next subsection)
Pharmacophores
Pharmacophore fingerprints
Spatial (Geometric)
Molecular volume, moments of inertia, molecular flexibility
Quantum-chemical
Molecular orbitals, HOMO / LUMO energy

Compare descriptors by locality, dimensionality, and calculation method. section 2.2.3

In a screening workflow, calculate descriptors after loading and checking molecules. RDKit can compute molecular weight, hydrogen bond acceptors, hydrogen bond donors, and logP for each molecule. Check for failed calculations and missing values before applying filters; the example workflow drops rows missing any of these four values. section 2.2 section 2.2.3

Listing 2.3 Compute RDKit descriptors
def calculate_ro5_descriptors(df, mol_col='mol'):
    def compute_descriptor(mol, func_name, missing_val=None):
        if mol is None: return missing_val
        try:
            func = RDKIT_DESCRIPTORS[func_name]
            return func(mol)
        except Exception as e:    #A
            print(f"Descriptor {func_name} failed: {e}")
            return missing_val    #A
    
    for desc in RO5_PROPS:
        df[desc] = df[mol_col].apply(
            lambda x: compute_descriptor(x, desc)
        )
    
    row_count_before = len(df)
    df = df.dropna(subset=RO5_PROPS)    #B
    row_count_after = len(df)
    return df
specs = calculate_ro5_descriptors(specs)

Inspect a calculation function that uses a descriptor-function mapping and a list of properties defined earlier, handles failures, and drops rows with missing values. section 2.2.3

Predicting solubility

For a solubility model, compute descriptors as input features and plot each against measured solubility. In one dataset, solubility decreased as logP increased, while topological polar surface area showed little apparent relationship. Scatterplots also exposed outliers. These observations can inform feature engineering, but a model still needs evaluation. section 4.0 section 4.1.3

Descriptors can also connect molecular structure to measured activity in a QSAR model. Such models can suggest which properties influence binding and generate optimization hypotheses. Interpretations remain bounded by the experimental data and the model’s applicability domain; some descriptors are easier to explain than others. section 6.0 section 6.3.2

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Computes descriptors, handles missing values, and inspects feature correlations, then connects descriptor choice to model interpretation.
Examples
Uses solubility scatterplots and RNA-binding QSAR models to illustrate descriptor relationships with measured properties.
What you can do after
Apply it via RDKit to compute descriptor features and check for missing values before filtering compounds.
How it connects
Connects molecular descriptors to fingerprints, similarity searching, feature engineering, QSAR, and applicability domains.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build AI Drug Discovery Pipelines. The section links point to the parts of the book it draws on.