Generated molecules, quality metrics, SAFE fragments, and downloadable artifacts from GenMol.

GenMol returns each generated molecule as SMILES and SAFE notation with RDKit-computed properties (QED, molecular weight, logP, TPSA and counts) plus run-level quality metrics. Use QED and your own filters to decide which molecules move on — the candidates you keep feed into ADMET-AI for property screening, DiffDock for a pose hypothesis, and AiZynthFinder for synthetic feasibility.

What these numbers are not

Every value below is computed from the structure. None of them is a measurement, and none of them is evidence that a molecule is active or makeable. sa_score is a heuristic estimate from the structure, not a synthesis route. For a makeability judgement use a retrosynthesis plan and a chemist. See how GenMol results compare to the bench.

Output artifacts

Each completed GenMol job returns structured results containing generated molecules, computed properties, and quality metrics.

Primary output fields

Field	Type	Description
smiles	string	Canonical SMILES for each generated molecule
safe	string	SAFE notation with fragment decomposition
qed	float (0-1)	Quantitative Estimate of Drug-likeness, from RDKit
sa_score	float (1-10)	Synthetic accessibility from RDKit's SA_Score, 1 (easy) to 10 (hard). Absent, never 0.0, when the scorer could not run, and the result's caveats then say so
molecular_weight	float	Molecular weight in Daltons
logp	float	Calculated octanol-water partition coefficient
tpsa	float	Topological polar surface area in square angstroms
hbd	integer	Hydrogen bond donor count
hba	integer	Hydrogen bond acceptor count
rotatable_bonds	integer	Number of rotatable bonds
validity	boolean	Whether the molecule passed chemical validity checks
unique	boolean	Whether the molecule is unique within the generated set

Quality summary metrics

The results panel displays aggregate statistics for the entire generation run:

Validity rate: percentage of generated molecules that are chemically valid. With enforce_validity on (the default), invalid molecules are removed from the returned set; with it off they stay in with is_valid: false
Uniqueness rate: percentage of non-duplicate molecules among the valid ones. With enforce_uniqueness on (the default), repeated SMILES are removed from the returned set
Diversity index: mean pairwise Tanimoto distance across the valid molecules (higher = more diverse). Returns null when fewer than two molecules are valid
QED mean and standard deviation: distribution of the RDKit drug-likeness estimate across the set

The validity and uniqueness rates describe everything the model generated. The property means (including sa_score_mean and sa_score_std) and the diversity index describe the molecules you receive after filtering. A quality_filters block in the result reports how many molecules went in, how many came out, and how many each filter removed.

Mode-specific outputs

De novo generation

Full property profile for each molecule
No similarity score (no reference molecule)

Lead optimization

tanimoto_similarity: Similarity to the seed molecule (0-1)
Molecules sorted by similarity descending
Seed molecule included for comparison

Scaffold decoration

r_groups: Enumeration of attached R-group fragments
Core scaffold highlighted in the molecule visualization
Attachment points annotated in the SAFE notation

Linker design

linker_smiles: SMILES of the generated linker substructure
linker_length: Number of heavy atoms in the linker
Full molecule with both fragments and linker shown

PMO (goal-directed)

oracle_score: Score from the target oracle function
Molecules sorted by oracle score descending
Optimization trajectory showing score improvement over iterations

SAFE fragment auto-ingestion

Generated molecules are automatically decomposed into SAFE fragments and ingested into the project fragment library:

Fragment extraction: Each molecule is broken into its constituent SAFE fragments
Deduplication: Duplicate fragments across molecules are merged
Metadata: Each fragment records its source job, parent molecule, and attachment point topology
Reuse: Ingested fragments appear in the project fragment library for use in subsequent generation runs

Downloading artifacts

CSV export: SMILES, SAFE notation, and all computed properties in tabular format
SDF file: 3D coordinates for all generated molecules (conformer generated automatically)
JSON export: Full structured results with metadata, parameters, and quality metrics
Fragment library export: SAFE fragments as a separate CSV for fragment-level analysis

Canvas visualization

When a GenMol job completes, a result card appears on the canvas.

Visualization: 2D molecule grid with property histograms and SAFE fragment decomposition

Tier 2 metrics

Molecules generated count
Drug-like percentage
Diversity index
Unique scaffold count

Interpreting the molecule grid

The canvas card shows a 2D grid of generated molecules with:

Structure thumbnails: 2D depictions of each molecule rendered via RDKit
QED badges: Color-coded drug-likeness scores (green >0.6, yellow 0.3-0.6, red <0.3)
Sort controls: sort by QED, molecular weight, or similarity
SAFE/SMILES toggle: Switch between SAFE fragment view and SMILES display

Generation mode indicators

Mode	Input	Output
denovo	None (random generation)	Novel molecules
scaffold_decoration	Fragment SMILES	Decorated scaffolds
linker_design	Two fragments	Linked molecules
motif_extension	Core motif	Extended molecules
pmo	Oracle name	Optimized molecules
lead_optimization	Starting molecule + similarity	Improved analogs

Interpreting quality scores

QED (Quantitative Estimate of Drug-likeness)

QED Range	Interpretation
0.7-1.0	High drug-likeness; suitable for lead optimization
0.5-0.7	Moderate drug-likeness; may need property optimization
0.3-0.5	Low drug-likeness; review specific property violations
<0.3	Poor drug-likeness; likely needs significant modification

Synthetic accessibility

Each molecule carries sa_score from RDKit's SA_Score (Ertl and Schuffenhauer), 1 easy to 10 hard, and the run reports sa_score_mean and sa_score_std over the returned set. Set max_sa_score to drop molecules above a ceiling; a molecule with no score is dropped by that filter rather than passed through, and if the scorer could not run at all the filter is reported as not applied instead of silently letting everything through.

⚠︎ It is a heuristic over fragment frequencies and structural complexity, not a route. It ranks molecules within a chemical series far better than across series, so treat makeability as unsettled until you have a retrosynthesis plan from AiZynthFinder and a chemist's read. The published literature on this question is in how GenMol results compare to the bench.

Diversity index

>0.8: Highly diverse set covering broad chemical space
0.5-0.8: Moderate diversity; some structural clustering
<0.5: Low diversity; molecules are structurally similar (expected in lead optimization mode)

Interpreting GenMol outputs