GenMol

What is GenMol?

GenMol is NVIDIA's discrete diffusion model for generating drug-like small molecules. It is designed for the common medicinal chemistry cases where the desired output is not one molecule, but a set of plausible analogs to rank, filter, dock, and refine.

The model works with SAFE, a fragment-based molecular representation. Instead of treating a molecule as a left-to-right SMILES string, SAFE represents molecules as fragment blocks with attachment points. GenMol can therefore generate a molecule from scratch, connect two fragments with a linker, grow from a motif, decorate a scaffold, or build a larger superstructure from an existing fragment using the same model family.

GenMol is useful early in discovery, before expensive structure-based or experimental filtering. Its outputs are candidate molecular graphs, not proof of binding, potency, selectivity, or synthesizability.

How to use GenMol online

Run GenMol online by choosing de novo generation or a fragment-constrained task, setting the number of molecules and sampling parameters, then submitting the job. ProteinIQ returns generated SMILES, ranked molecule properties, and downloadable SDF files for follow-up analysis, filtering, visualization, or docking.

Inputs

Input	Required	Description
Fragment	No for De novo; yes for Fragment-constrained	SMILES or SMI text with * attachment points. Linker design uses two fragments separated by ..
Job name	No	Optional label for identifying the run in job history.

Fragment input examples

Task	Example input	What it asks GenMol to do
Linker design (one-step)	CC(*)c1ccccc1.*c1ccc(F)cc1	Connect two fragments with a newly generated linker.
Motif extension	c1ccc(*)cc1	Grow substituents from a known motif.
Scaffold decoration	c1cc(*)cc(*)c1	Add groups at multiple scaffold attachment sites.
Superstructure generation	c1ccc(*)cc1	Generate larger molecules that retain the supplied fragment context.

Attachment points matter. A chemically unreasonable * position can force GenMol into poor local chemistry even when the parent fragment looks drug-like. For fragment work, prepare the fragment at the bond where synthetic elaboration or medicinal chemistry exploration is intended.

Fragment SMILES requirements

GenMol fragment-constrained mode does not accept an ordinary complete SMILES as a starting molecule. The input must be a fragment SMILES with one or more * attachment markers that tell GenMol where new chemistry can be generated.

Requirement	Valid pattern	Invalid pattern	Why it matters
De novo mode has no fragment input	Empty fragment field	CCO in the fragment field	De novo generation starts from the learned chemical distribution, not from a supplied molecule.
Motif extension has one growth point	c1ccc(*)cc1	c1ccccc1	The * marks the atom where GenMol should grow new chemistry.
Scaffold decoration can mark several positions	c1cc(*)cc(*)c1	c1cc(*)cc(*)c1.*	Multiple * markers belong on the same molecular scaffold, not as a separate bare fragment.
Linker design has two molecular fragments	CC(*)c1ccccc1.*c1ccc(F)cc1	CC(*)c1ccccc1.*	A standalone * is only an attachment marker. It is not a second fragment for GenMol to connect.
Each linker side has one attachment point	warhead*.*recruiter	warhead*.**recruiter*	Linker design expects one open bond on each side of the linker. Extra markers make the connection ambiguous.

GenMol is trained for drug-like organic small molecules. Fragment inputs should use ordinary medicinal-chemistry atoms such as C, N, O, S, P, F, Cl, Br, and I. Inorganic cages, organosilicon scaffolds such as [Si], salts, disconnected non-fragment species, and highly unusual query-style structures are outside the intended input space and may fail before molecule generation begins.

In practical terms, a good fragment input is a chemically meaningful partial ligand with one clear open attachment point:

• c1ccc(*)cc1: phenyl motif with one growth point.
• CC(*)c1ccccc1.*c1ccc(F)cc1: two ligand fragments prepared for linker design.
• c1cc(*)cc(*)c1: scaffold with two decoration positions.

These are not equivalent:

• CCO: a complete molecule with no generation site.
• *: an attachment marker without a molecular fragment.
• CC(*)c1ccccc1.*: one valid fragment plus one missing fragment.
• *[H][Si](C)(C)O*: an organosilicon fragment outside GenMol's drug-like organic fragment scope.

Settings

Setting	Range or values	Default	Description
Generation mode	De novo, Fragment-constrained	De novo	De novo generates molecules without a starting structure. Fragment-constrained conditions generation on the supplied fragment input.
Fragment task	Linker design (one-step), Linker design (two-step), Motif extension, Scaffold decoration, Superstructure generation	Linker design (one-step)	Selects the fragment-constrained generation behavior. This setting only affects Fragment-constrained runs.
Number of molecules	10 to 200	50	Number of molecules to generate. Larger runs sample more chemical diversity and take longer.
Softmax temperature	0.5 to 2.0	1.0	Controls token-level diversity during sampling. Lower values are more conservative; higher values produce broader chemistry with more risk of low-quality outputs.
Randomness	0.1 to 5.0	0.3	Controls stochasticity during generation. GenMol V2 de novo examples commonly use 0.3; fragment tasks often need higher values such as 1.0 to 3.0.
Molecular context guidance (gamma)	0.0 to 1.0	0.3	Controls how strongly fragment-constrained generation follows the supplied context. Higher values keep closer to the input fragment; lower values allow more exploration.

Choosing a generation mode

Goal	Best mode	Practical note
Explore broad ligand-like chemical space	De novo	Start with 50 molecules, then increase to 100 or 200 if the first run has promising property ranges.
Design a PROTAC or bivalent compound linker	Linker design (one-step)	Use two fragments with one attachment point each. Follow with length, polarity, and flexibility filtering.
Build a more conservative linker	Linker design (two-step)	Useful when one-step sampling produces linkers that look too abrupt or chemically strained.
Elaborate a known hit fragment	Motif extension	Best for one attachment point where the fragment should remain recognizable.
Explore SAR around a core	Scaffold decoration	Mark every intended substitution position with *; avoid marking positions that should remain fixed.
Search around a fragment in a larger chemical context	Superstructure generation	More exploratory than simple motif extension. Expect more diversity and a wider property spread.

Understanding the results

GenMol returns a ranked table of generated molecules, SMILES strings, molecular property columns, and SDF files. The viewer tab prioritizes generated molecules by drug-likeness, but the right candidate depends on the next assay or computational filter.

Result	Description
Rank	ProteinIQ ranking of generated molecules, primarily useful for scanning the output table.
QED	Quantitative Estimate of Drug-likeness, from 0 to 1. Higher values usually indicate a property profile closer to known oral drugs.
SA	Synthetic accessibility score. Lower values are generally easier to synthesize.
MW (Da)	Molecular weight in Daltons. Many oral small molecules sit below roughly 500 Da, although PROTACs and macrocycles often exceed that range.
LogP	Predicted octanol-water partition coefficient. Higher values usually mean more hydrophobic molecules.
SMILES	Canonical SMILES for downstream filtering, property prediction, or docking preparation.
Download	Downloadable SDF for each generated molecule.

Reading QED and SA together

QED should not be used as a single pass or fail decision. It rewards molecular property profiles similar to approved drugs, which is useful for triage, but it does not know the target, binding pocket, assay format, liabilities, or project-specific design constraints.

Pattern	Interpretation
High QED, low SA	Often the easiest candidates to inspect first. These molecules combine drug-like property balance with comparatively accessible chemistry.
High QED, high SA	Attractive on paper but potentially difficult to make. Check ring systems, stereochemistry, and unusual functional groups before prioritizing.
Low QED, low SA	Chemically accessible molecules that may need property optimization. These can still be useful for fragment or tool-compound exploration.
Low QED, high SA	Usually lower priority unless the structure fits a specific project hypothesis.

The quality summary used in GenMol examples combines QED of at least 0.6 with SA no greater than 4. That is a useful first filter for many drug-like campaigns, not a universal medicinal chemistry rule.

Summary metrics

Metric	Meaning
Validity	Fraction of requested molecules that were valid molecular structures.
Uniqueness	Fraction of valid generated SMILES that are non-duplicates.
Diversity	Average pairwise fingerprint distance across unique molecules. Higher values indicate a broader chemical set.
Quality	Fraction of requested molecules meeting the run's quality criteria, commonly based on QED and SA.

High diversity is useful for hit discovery. Lower diversity can be acceptable for scaffold decoration or linker design when the goal is a focused analog series around a fixed starting point.

SDF files and 3D coordinates

The SDF files are convenient for visualization and downstream tools, but they should not be interpreted as binding poses. The coordinates are generated molecular conformers, not docked orientations in a protein pocket. For target-specific pose assessment, use DiffDock, AutoDock Vina, or GNINA after preparing a receptor structure.

How GenMol works

GenMol combines masked discrete diffusion with a transformer model trained on SAFE molecular strings. The diffusion process starts from masked molecular tokens and iteratively predicts replacements, refining a full molecule over repeated denoising steps.

This differs from autoregressive SMILES generation. Autoregressive models generate one token after another, so every token depends heavily on the order chosen so far. GenMol predicts masked regions with bidirectional context, which lets the model use information from both sides of a fragment arrangement and sample multiple positions in parallel.

SAFE representation

SAFE stands for Sequential Attachment-based Fragment Embedding. It decomposes molecules into fragment blocks, often using BRICS-style chemistry rules, and records how fragments attach. A linker design input such as CC(*)c1ccccc1.*c1ccc(F)cc1 contains two fragment contexts and attachment markers for the region that should be generated.

Fragment order is less important in SAFE than in ordinary SMILES. That property fits the way GenMol uses bidirectional context: the model can reason over the fragment set instead of only the next token in a string.

Fragment remasking

GenMol's exploration strategy can mask whole fragments rather than only individual tokens. During generation, uncertain or replaceable fragments can be masked again and regenerated in the context of the surrounding molecule. This makes the search operate at a chemically meaningful level, closer to replacing a substituent or linker block than editing a single atom at a time.

Molecular context guidance

Molecular context guidance, controlled by Molecular context guidance (gamma), increases the influence of supplied fragments during constrained generation. Higher gamma values usually produce molecules that preserve the intended fragment context more strongly. Lower values give the model more room to depart from the input and may increase novelty.

Practical workflows

De novo hit discovery

• Run De novo generation with 50 molecules for a first pass.
• Increase to 100 or 200 molecules if the first run returns valid, diverse chemistry.
• Filter the table by QED, SA, MW (Da), and LogP.
• Check rule-based drug-likeness with Lipinski's Rule of 5 and Veber's rule.
• Screen liabilities with PAINS filter, Brenk filter, and Toxicity prediction.
• Predict broader absorption and safety properties with ADMET-AI.
• Dock promising ligands against a prepared receptor with DiffDock, AutoDock Vina, or GNINA.

Fragment-based lead optimization

• Start from a validated fragment or hit-like motif, not a random substructure.
• Mark the intended elaboration atom with *.
• Use Motif extension for one growth point or Scaffold decoration for several substitution positions.
• Keep Softmax temperature near 1.0 to 1.2 for conservative analog generation.
• Raise Randomness only when repeated runs return near-duplicates or overfit to the starting fragment.

Linker design

• Prepare two fragments with one attachment point each.
• Use dot-separated input such as warhead*.*recruiter.
• Start with Linker design (one-step) for broad exploration.
• Try Linker design (two-step) if the first run produces linkers that are too strained, too hydrophobic, or too similar.
• Filter linkers by atom count, rotatable bonds, polarity, and project-specific spatial constraints before docking or synthesis planning.

When to use GenMol vs alternatives

GenMol is a ligand generation tool. It proposes candidate molecules or analogs, then other tools test whether those candidates satisfy project constraints.

Need	Better starting point
Generate new ligand-like molecules from scratch	GenMol De novo
Grow or decorate a known fragment	GenMol Motif extension or Scaffold decoration
Connect two ligand fragments	GenMol Linker design
Generate ligands inside a known protein pocket	PocketFlow or PocketxMol
Predict binding poses for generated molecules	DiffDock, AutoDock Vina, or GNINA
Convert generated SMILES into structure files	SMILES to SDF, SMILES to PDB, or SMILES to MOL2
Filter property and liability risks	ADMET-AI, Lipinski's Rule of 5, PAINS filter, or Brenk filter

Limitations

GenMol generates molecular candidates, not validated drugs. A high-ranked molecule still needs target-aware modeling, liability screening, retrosynthetic review, and experimental validation.

Fragment-constrained runs depend strongly on input quality. Poor attachment points, unstable fragments, salts, disconnected chemistry outside the intended format, or very unusual motifs can reduce validity and usefulness.

The SDF output is not a substitute for docking or conformational analysis. It is a convenient structure file for downstream work, while binding geometry must be assessed in a receptor context.