Ivy Biosciences Docs/Tools/Generative & de novo/Genmol/GenMol quickstart

GenMol quickstart

Step-by-step guide to generating your first molecules with GenMol on the platform.

Open in Claude Open in ChatGPT5 min read

You want to generate novel drug-like molecules with GenMol; this walks you through your first run in five steps — select a mode, set the parameters, submit, and review the output. Start in de novo mode to see the full pipeline without needing any input molecules.

Quick molecule generation with .bloom* platform

GenMol uses NVIDIA's masked discrete diffusion model on SAFE molecular representations to generate novel drug-like molecules. This guide walks through your first generation run, from selecting a mode to interpreting the output molecules.

Step-by-step generation workflow

Access the GenMol tool

Navigate to GenMol in the .bloom* tool selector, or ask Ivy directly:

"Generate 100 drug-like molecules with GenMol"
"Run GenMol in de novo mode"

Select a generation mode

GenMol supports nine run modes. Choose the one that matches your workflow:

Mode	When to use	Required input
de novo	Explore novel chemical space without constraints	None
lead optimization	Create analogs of a known hit	Seed SMILES + similarity threshold
scaffold decoration	Attach fragments to a core scaffold	Scaffold SMILES with attachment points
linker design	Connect two fragment hits	Two fragment SMILES
motif extension	Grow a pharmacophore motif into a full molecule	Motif SMILES
PMO	Optimize toward a scoring oracle	Oracle name (qed, gsk3b, etc.)
fragment constrained	Fix specific fragments, generate the rest	SAFE-encoded fixed fragments

For your first run, start with de novo mode to see the full pipeline without needing input molecules.

Configure generation parameters

Basic settings (recommended for first run):

Mode: de novo
Number of molecules: 100 (good balance of speed and diversity)
Quality filter: Enabled (filters out invalid and non-drug-like molecules)
Job name: Enter a descriptive name or leave blank for auto-generation

Advanced settings (via gear icon):

num_steps: Diffusion steps per molecule (default is fine for most runs)
temperature: Controls diversity vs. quality tradeoff (default: 1.0; lower = more conservative, higher = more diverse)
seed: Set for reproducible results

Submit and monitor the job

Click Generate Molecules to start the job
.bloom* shows real-time progress with live ETA
Typical completion: 30 seconds to 2 minutes for 100 molecules
Job status updates appear in the job panel and as notifications

Review and analyze results

Canvas visualization:

A result card appears on the canvas showing a 2D molecule grid
Each molecule displays its structure, QED score, and SA score
Property histograms show the distribution of molecular weight, logP, and TPSA across all generated molecules
SAFE fragment decomposition view shows how molecules break down into reusable fragments

Quality metrics panel:

Validity rate: percentage of chemically valid molecules generated. Invalid ones are removed from the returned set unless you turn enforce_validity off
Uniqueness rate: percentage of unique (non-duplicate) molecules
QED mean: average RDKit drug-likeness estimate across the set
Diversity index: mean pairwise Tanimoto distance across the generated set

None of these is a measurement. They describe the structures, not any property anyone observed — see how GenMol results compare to the bench.

Download options:

CSV: SMILES, SAFE notation, and all computed property scores
SDF: 3D coordinates for downstream docking or visualization
JSON: Full structured results with metadata

Common generation scenarios

De novo exploration

Select "de novo" mode with 500-1,000 molecules
Enable drug-likeness filter for a cleaner output set
Sort results by QED or diversity score
Pick top candidates for ADMET screening or docking

Lead optimization

Enter the SMILES of your lead compound (e.g., CC(=O)Oc1ccccc1C(=O)O for aspirin)
Set similarity threshold to 0.6-0.8 (lower = more diverse analogs)
Generate 200-500 analogs
Compare analogs side-by-side in the canvas grid
Feed top hits to ADMET-AI for property profiling

Scaffold decoration

Enter your scaffold SMILES with SAFE attachment points
Generate 100-300 decorated molecules
Review the R-group variations in the molecule grid
Export decorated scaffolds for retrosynthetic analysis with AiZynthFinder

Fragment linking

Provide two fragment SMILES from your fragment screening campaign
Generate 100-200 linked molecules
Filter for drug-likeness and synthetic accessibility
Dock top linkers against the target protein with DiffDock

Quality control checklist

Before submission

Correct generation mode selected for your use case
Seed SMILES is valid (check with PubChem or RDKit)
Similarity threshold is reasonable (0.5-0.8 for lead optimization)
Molecule count is appropriate (100 for quick exploration, 1,000+ for campaigns)
Quality filters enabled if you want drug-like output

After generation

Validity rate is >90% (if lower, check seed SMILES or constraints)
Uniqueness rate is >80% (low uniqueness may indicate constrained generation)
You have read the quality_filters block to see how many molecules the filters removed
Generated molecules are structurally reasonable (spot-check 2D structures; this catches what no metric reports)
Property distributions match your target profile
You are treating the set as a hypothesis pool, not as candidates that have been shown to be makeable or active

Next steps after generation

Property screening:

Run ADMET-AI or ADMETLab3 on generated molecules for absorption, distribution, metabolism, excretion, and toxicity predictions
Use ProTox-3 for toxicity endpoint estimation
Apply DeepChem for custom property prediction

Synthetic feasibility:

Run AiZynthFinder on top candidates for retrosynthetic analysis — this is the only makeability check available here
Check ZINC for purchasable building blocks
Have a chemist read the structures before anything is ordered. Treat sa_score as a within-series ranking hint, not a makeability verdict

Structural validation:

Dock generated molecules against target proteins using DiffDock
Minimize docked poses with MM-GBSA, which reports complex energies rather than a binding free energy
Build a Compound Passport for comprehensive drug-likeness assessment

Iterative optimization:

Use top-scoring molecules as seeds for a second round of lead optimization
Extract promising fragments and use them as constraints for fragment-constrained generation
Run PMO mode targeting specific property improvements