Predict Drug-Likeness with AI

What you'll accomplish

By the end of this guide, you'll have drawn a molecule, asked the AI to compute its properties, and interpreted a Lipinski Rule of Five evaluation. You'll understand what the confidence indicators mean and how to use property predictions in your research workflow.

Why drug-likeness matters

Drug-likeness is a set of molecular property criteria that predict whether a compound is likely to be orally bioavailable. The most widely used filter is Lipinski's Rule of Five, which flags compounds that violate two or more of four thresholds: molecular weight > 500 Da, logP > 5, more than 5 hydrogen bond donors, or more than 10 hydrogen bond acceptors.

These rules are not absolute — many successful drugs violate them (notably antibiotics and natural products). But they are a useful first-pass filter when evaluating a compound library or planning a medicinal chemistry campaign.

Step 1: Draw or load your molecule

Draw a structure on the canvas, use a quick-start template, or import a file (SMILES, MOL, SDF, or CDXML). The AI sees whatever is on the canvas in real time — no copy-paste needed.

For this walkthrough, try loading aspirin from the quick-start templates. It's a good example because it passes the Rule of Five comfortably.

Step 2: Ask the AI for properties

In the chat panel, type: "What are the properties of this molecule?" The AI calls RDKit via a backend tool to compute six molecular descriptors.

Results appear in a two-tier card. The summary shows key values at a glance. Click to expand the detailed view with range bars showing where your molecule falls relative to typical drug-like ranges.

Step 3: Read the property card

The property card reports:

• • Molecular weight (MW) — Total mass in Daltons. Lipinski threshold: ≤ 500 Da.
• • logP (Wildman-Crippen) — Estimated octanol-water partition coefficient. Higher values mean more lipophilic. Lipinski threshold: ≤ 5. The Wildman-Crippen method is a fragment-based approach and is the standard logP method used in RDKit.
• • Hydrogen bond donors (HBD) — NH and OH groups. Lipinski threshold: ≤ 5.
• • Hydrogen bond acceptors (HBA) — N and O atoms. Lipinski threshold: ≤ 10.
• • Topological polar surface area (TPSA) — Sum of polar atom surfaces, in Ų. Not part of Lipinski, but values above 140 Ų correlate with poor oral absorption.
• • Rotatable bonds — Bonds that allow free rotation. Not part of Lipinski, but more than 10 rotatable bonds is associated with poor oral bioavailability.

Step 4: Interpret confidence indicators

Every result in ChemStitch carries a visible confidence level:

Computed (green badge) means the value was calculated deterministically by RDKit. These values are reproducible and exact for the given structure. All six molecular descriptors listed above are Computed values.

AI Suggested (yellow badge) means the AI generated the result. This applies to reaction predictions, retrosynthesis suggestions, and other AI-generated content. These are useful for exploration but should be verified experimentally or against literature.

Step 5: Evaluate Lipinski compliance

The property card includes a Lipinski Rule of Five pass/fail evaluation. A molecule passes if it violates no more than one of the four thresholds (MW, logP, HBD, HBA).

If your molecule fails, the card shows which specific rules were violated. This tells you exactly which properties to optimize if you're designing analogs.

Batch analysis for compound libraries

If you have an SDF file with multiple molecules, you can compute properties for the entire set at once. Import the SDF file, click "Analyze All (N)" in the notification bar, and ChemStitch processes up to 500 molecules. Results appear in a sortable table with a Lipinski column that flags violations with a warning icon.

See the Batch Analyze an SDF Library guide for the full walkthrough.