cheminformatics

Cheminformatics

Part of Agent Skills™ by googleadsagent.ai™

Description

Cheminformatics provides computational chemistry workflows using RDKit for molecular property prediction, virtual screening, ADMET analysis, molecular docking preparation, and chemical space exploration. The agent generates reproducible cheminformatics pipelines that transform molecular structures (SMILES, SDF) into actionable predictions about drug-likeness, toxicity, and binding affinity.

Drug discovery generates vast chemical libraries that cannot all be synthesized and tested. Cheminformatics narrows the search space computationally: filtering by Lipinski's Rule of Five, predicting ADMET properties (Absorption, Distribution, Metabolism, Excretion, Toxicity), scoring docking poses, and clustering chemical space to identify diverse lead candidates. Each step eliminates compounds that would fail in later, more expensive stages.

This skill covers the molecular informatics workflow from SMILES parsing through descriptor calculation, fingerprint generation, similarity searching, property prediction, and visualization. It integrates with databases like PubChem and ChEMBL for compound retrieval and benchmarking against known actives and inactives.

Use When

Calculating molecular properties and descriptors
Screening compound libraries for drug-likeness
Predicting ADMET properties for lead compounds
Performing molecular similarity searches
Preparing structures for molecular docking
Visualizing chemical space and structure-activity relationships

How It Works

graph TD
    A[Molecular Input: SMILES/SDF] --> B[Parse + Validate Structures]
    B --> C[Calculate Descriptors]
    C --> D[Drug-likeness Filters]
    D --> E{Passes Lipinski?}
    E -->|No| F[Flag as Non-Drug-like]
    E -->|Yes| G[ADMET Prediction]
    G --> H[Virtual Screening Score]
    H --> I[Docking Preparation]
    I --> J[Ranked Candidate List]
    F --> K[Report with Flags]
    J --> K

Compounds flow through increasingly selective filters. Drug-likeness removes obviously non-viable candidates, ADMET prediction flags absorption and toxicity risks, and virtual screening ranks the survivors by predicted activity.

Implementation

from rdkit import Chem
from rdkit.Chem import Descriptors, AllChem, Draw, Lipinski, DataStructs
from rdkit.Chem import rdMolDescriptors
import pandas as pd

def molecular_properties(smiles: str) -> dict:
    mol = Chem.MolFromSmiles(smiles)
    if mol is None:
        raise ValueError(f"Invalid SMILES: {smiles}")
    return {
        "smiles": smiles,
        "mw": Descriptors.MolWt(mol),
        "logp": Descriptors.MolLogP(mol),
        "hbd": Descriptors.NumHDonors(mol),
        "hba": Descriptors.NumHAcceptors(mol),
        "tpsa": Descriptors.TPSA(mol),
        "rotatable_bonds": Descriptors.NumRotatableBonds(mol),
        "rings": Descriptors.RingCount(mol),
        "lipinski_violations": sum([
            Descriptors.MolWt(mol) > 500,
            Descriptors.MolLogP(mol) > 5,
            Descriptors.NumHDonors(mol) > 5,
            Descriptors.NumHAcceptors(mol) > 10,
        ]),
    }

def lipinski_filter(df: pd.DataFrame) -> pd.DataFrame:
    return df[df["lipinski_violations"] <= 1].copy()

def similarity_search(query_smiles: str, library: list[str], threshold: float = 0.7) -> list[dict]:
    query_mol = Chem.MolFromSmiles(query_smiles)
    query_fp = AllChem.GetMorganFingerprintAsBitVect(query_mol, radius=2, nBits=2048)

    results = []
    for smi in library:
        mol = Chem.MolFromSmiles(smi)
        if mol is None:
            continue
        fp = AllChem.GetMorganFingerprintAsBitVect(mol, radius=2, nBits=2048)
        tanimoto = DataStructs.TanimotoSimilarity(query_fp, fp)
        if tanimoto >= threshold:
            results.append({"smiles": smi, "tanimoto": tanimoto})

    return sorted(results, key=lambda x: -x["tanimoto"])

def admet_flags(props: dict) -> list[str]:
    flags = []
    if props["logp"] > 5:
        flags.append("High lipophilicity: poor aqueous solubility risk")
    if props["tpsa"] > 140:
        flags.append("High TPSA: poor membrane permeability risk")
    if props["mw"] > 500:
        flags.append("High MW: poor oral absorption risk")
    if props["rotatable_bonds"] > 10:
        flags.append("High flexibility: poor oral bioavailability risk")
    return flags

Best Practices

Always validate SMILES parsing before computing descriptors—invalid structures produce silent errors
Use Morgan fingerprints (radius=2, 2048 bits) as the default for similarity calculations
Apply Lipinski's Rule of Five as a first-pass filter, not an absolute cutoff
Report Tanimoto similarity thresholds used in all similarity searches
Standardize molecules (desalt, neutralize, canonicalize) before comparison
Visualize chemical space with t-SNE or UMAP on fingerprint representations

Platform Compatibility

Platform	Support	Notes
Cursor	Full	Python + RDKit environment
VS Code	Full	Jupyter + molecular viz
Windsurf	Full	Scientific Python
Claude Code	Full	Pipeline generation
Cline	Full	Cheminformatics workflows
aider	Partial	Code generation only

Related Skills

Keywords

cheminformatics rdkit molecular-properties virtual-screening admet lipinski drug-discovery molecular-similarity