Run modes, fragment constraints, SMILES formats, and quality filter configuration for GenMol.

You provide a generation mode and, for most modes, one or more seed SMILES — a reference molecule, scaffold, or fragments. The mode you choose is what most shapes the result: it determines whether generation runs unconstrained (de novo, PMO) or is anchored to the structures you supply (lead optimization, scaffold decoration, linker design, motif extension, fragment-constrained).

Webapp input methods

The .bloom* webapp provides several input methods depending on the selected generation mode.

Input type selection

Input Type	Applicable Modes	Description
No input	de novo, PMO	Fully unconstrained generation
Single SMILES	lead optimization, motif extension	Seed molecule or motif as SMILES string
Scaffold SMILES	scaffold decoration	Core scaffold with SAFE attachment points
Two fragment SMILES	linker design	Two fragments to be connected by a generated linker
SAFE constraints	fragment constrained	Fixed SAFE tokens with mask tokens for generation
From previous job	all modes	Import SMILES from fragment library, PubChem, or other tools

Generation modes

De novo generation

Input: None required
Behavior: Generates molecules from scratch by sampling the full learned distribution
Best for: Chemical space exploration, library enumeration, diversity-focused campaigns

Lead optimization

Input: Reference SMILES string
Additional parameter: similarity_threshold (0.0-1.0)
Behavior: Generates analogs constrained to be within the specified Tanimoto similarity to the seed molecule
Best for: SAR exploration around a known hit or lead compound

Scaffold decoration

Input: Scaffold SMILES with SAFE attachment points
Behavior: Keeps the core scaffold fixed and generates R-group substituents at marked positions
Best for: Exploring R-group variations on a validated scaffold

Linker design

Input: Two fragment SMILES
Behavior: Generates molecular linkers connecting the two fragments into a single molecule
Best for: Fragment-based drug design where fragment hits need to be merged

Motif extension

Input: Motif SMILES (pharmacophore or substructure)
Behavior: Grows the motif into a full drug-like molecule while preserving the core substructure
Best for: Growing fragment hits or pharmacophore motifs into full molecules

PMO (Practical Molecular Optimization)

Input: Oracle name (e.g., qed, gsk3b, jnk3, drd2)
Behavior: Goal-directed generation targeting high scores on the specified oracle
Best for: Benchmark optimization, targeting specific computational endpoints

Fragment-constrained generation

Input: SAFE-encoded molecule with fixed and masked tokens
Behavior: Keeps specified SAFE fragments fixed while generating the remaining portions
Best for: Fine-grained control over which parts of the molecule are generated vs. preserved

SMILES input requirements

Valid SMILES

Standard notation: Use canonical or isomeric SMILES (e.g., CC(=O)Oc1ccccc1C(=O)O)
Validation: Real-time checking via RDKit; invalid SMILES are rejected before submission
Stereochemistry: Supported but not required (e.g., C[C@@H](O)CC for specified chirality)
Case sensitivity: Element symbols are case-sensitive (e.g., c for aromatic carbon, C for aliphatic)

SAFE notation

Fragment representation: Molecules are decomposed into fragments with numbered attachment points (e.g., [1]c1ccccc1.[1]N)
Attachment points: Numbered brackets indicate where fragments connect (e.g., [1], [2])
Dot separation: Fragments are separated by dots in SAFE strings
Automatic conversion: .bloom* can convert SMILES to SAFE notation automatically

Size constraints

Maximum SMILES length: ~200 characters for seed molecules
Minimum fragment size: At least 3 heavy atoms per fragment
Maximum fragments: Up to 10 fixed fragments per constrained generation job

Generation parameters

Core parameters

Parameter	Type	Default	Description
run_type	string	`denovo`	Generation mode (denovo, fragment, linker_design, linker_design_onestep, motif_extension, scaffold_decoration, superstructure_generation, pmo, lead_optimization)
num_molecules	integer	100	Number of molecules to generate (1-10,000)
fragments	string[]	-	Seed SMILES fragments (scaffold_decoration, linker_design, motif_extension, fragment modes)
starting_molecule	string	-	Seed SMILES for lead_optimization mode
num_diffusion_steps	integer	50	Diffusion denoising steps per molecule (10-200)
temperature	float	1.0	Sampling temperature (0.1-2.0)

Quality filter parameters

These filters change what you get back

enforce_validity and enforce_uniqueness are on by default, so a run returns only parseable molecules and no repeated SMILES — the validity and uniqueness rates still describe everything that was generated, so they can sit below 100% while every molecule you receive passes. The thresholds below default to unset and bind only when you set them, except in lead_optimization, which applies its own QED floor and SA ceiling (see mode-specific parameters). A molecule whose property could not be computed is dropped by a filter you asked for rather than passed through. See interpreting outputs.

Parameter	Type	Default	Description
enforce_validity	boolean	true	Drop molecules RDKit cannot parse
enforce_uniqueness	boolean	true	Drop repeats of a SMILES already returned
apply_druglikeness_filter	boolean	false	Drop molecules breaking any of Lipinski's four bounds (MW ≤ 500, logP ≤ 5, HBD ≤ 5, HBA ≤ 10)
min_qed	float	unset	Drop molecules whose RDKit QED is below this value (0-1)
max_sa_score	float	unset	Drop molecules whose synthetic accessibility is above this value (1 easy to 10 hard)

The result carries a quality_filters block with the counts in, the counts out and how many molecules each filter removed, plus a plain-language caveat for any filter that could not be evaluated.

Mode-specific parameters

Parameter	Mode	Type	Default	Description
similarity_threshold	lead optimization	float	0.4	Minimum Tanimoto similarity to seed
oracle_name	PMO	string	required	TDC scoring oracle for goal-directed generation (e.g. `jnk3`, `gsk3b`, `drd2`)
optimization_budget	PMO	integer	10,000	Maximum oracle evaluations (100-100,000)
fragments	linker design	string[]	required	Both fragment SMILES to connect, as a two-element list
min_qed_lead	lead optimization	float	0.5	QED floor applied to a lead run's output. Set `min_qed` to override it
max_sa_lead	lead optimization	float	5.0	Synthetic-accessibility ceiling applied to a lead run's output. Set `max_sa_score` to override it

The two lead bounds are GenMol's own softer band rather than thresholds invented here; the paper's stricter 0.6 / 4 pair excludes approved drugs such as aspirin (QED 0.550), which is why it is not the default. A lead run that returns nothing says which bound removed everything and names the parameter that widens it.

Previous job integration

Compatible tools

Fragment library: SAFE fragments from previous GenMol, REINVENT4, or fragment extraction jobs
PubChem: Compound SMILES from structure lookups
Fragment search/extraction: Fragment hits from screening campaigns
Other GenMol runs: Top-scoring molecules for iterative optimization

Data handling

Multi-molecule files: Select individual molecules from multi-compound outputs
Format conversion: Automatic handling of SDF, MOL, and other chemical file formats
Provenance tracking: Maintains connection to original data source for reproducibility

Batch processing strategies

Single-mode batches

Recommended: Submit one job per generation mode for clear result organization
Molecule count: 100-1,000 per job for good balance of speed and diversity
Error isolation: Issues affect only individual jobs

Iterative optimization

Round 1: De novo generation with 1,000 molecules and strict quality filters
Round 2: Lead optimization using top hits from Round 1 as seeds
Round 3: Scaffold decoration using the best scaffold from Round 2
Track lineage: .bloom* maintains provenance across iterative runs

Pre-submission validation

SMILES quality checks

Validity: Ensure SMILES parses correctly (use the real-time validator in .bloom*)
Sanitization: Check for valence errors, impossible bond orders, or invalid atom types
Stereochemistry: Verify chirality specifications are intentional
Canonicalization: Canonical SMILES avoids duplicate representations of the same molecule

Parameter optimization

Start conservative: Use defaults for first runs, then tune based on results
Temperature exploration: Try 0.7 (conservative), 1.0 (balanced), and 1.3 (diverse) to find the sweet spot
Step count: More steps generally improve quality but increase runtime

GenMol inputs and parameters