Version Compatibility

Reference examples tested with: REINVENT 4.0+, RDKit 2024.09+, PyTorch 2.1+, MolMIM (NVIDIA BioNeMo), chemprop 2.0+.

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Generative Molecular Design

Hard rules

1. No fabricated citations. Every cited work must resolve to a verifiable
2. No claim without provenance. Every quantitative or factual claim
3. No silent failure. Every script invocation, API call, or tool use must declare its exit status and what to do on non-zero. A skill that silently swallows errors is a violation.

When to use

Load this skill when the user asks a question that matches its declared trigger conditions (see the frontmatter description for the most common ones). Do not load it for questions outside its scope — defer to the appropriate ORS skill instead.

When NOT to use

Do not load this skill if the question is in a sibling skill's domain (see ## Cross-references), if the user explicitly asks for a different tool, or if the task is outside the skill's declared category.

Generator Mode Taxonomy

Learning Algorithm Taxonomy

Decision Tree by Scenario

REINVENT 4 Setup

REINVENT 4 uses a TOML configuration file specifying generator, algorithm, prior model, and scoring functions.

[parameters]"
prior_file = "priors/reinvent.prior"
agent_file = "priors/reinvent.prior"
batch_size = 64
unique_sequences = true

[[stage]]
max_steps = 1000
chkpt_file = "checkpoints/agent.chkpt"

[[stage.scoring.component]]
name = "QED"

[[stage.scoring.component]]
name = "custom_activity"
weight = 1.0

Multi-Parameter Optimization (MPO)

The art of generative design lies in the scoring function. Common components:

Critical pitfall: Reward hacking. If activity model is biased, generator produces structures that exploit the bias. Mitigations:

• Use ensemble of models (5+)
• Constrain to chemically reasonable substructures
• Validate top-100 by orthogonal in silico methods (docking, FEP)

MolMIM (NVIDIA BioNeMo)

MolMIM is a property-guided latent-variable model:

from molmint import MolMIM
from rdkit import Chem

model = MolMIM()
smiles = 'CCO'
optimized = model.optimize(smiles, target_logp=2.5, target_sas=2.0)
mol = Chem.MolFromSmiles(optimized)

Strength: Continuous property optimization in latent space. Weakness: Latent space not always semantically meaningful.

Diffusion-Based Generators (DiffSMol, DiGress)

Diffusion models generate molecules by iteratively denoising:

# DiffSMol / DiGress pseudo-API; verify against current release.
# from diffsmol import generate
# mols = generate(n_samples=1000, scaffold='aryl_sulfonamide')

Strength: State-of-the-art sample quality on MOSES benchmark. Weakness: Slower than RL; harder to condition on multiple objectives.

JT-VAE (Latent-Space Optimization)

Junction Tree VAE optimizes in latent space then decodes:

# Pseudo-API
# from jtvae import optimize
# best_mol = optimize(smiles='CCO', target='activity', iterations=100)

Strength: Smooth latent space for optimization. Weakness: Outdated vs transformers; reconstruction quality lower.

Per-Tool Failure Modes

REINVENT -- reward hacking

Trigger: Generated molecules score high but don't bind target.

Mechanism: Generator exploits QSAR model weaknesses (e.g., simple features that correlate spuriously).

Symptom: Top-100 in silico but 0% hit rate in vitro.

Fix: Ensemble of 5+ scoring models; structural diversity constraint; orthogonal validation (docking).

Scaffold decoration -- diversity loss

Trigger: After 1000 REINVENT steps, all generated molecules are near-same scaffold.

Mechanism: Generator converges to local optimum.

Symptom: Generated SMILES have Tanimoto > 0.9 to scaffold.

Fix: Add Tanimoto distance to known actives as bonus; restart with new seed; increase stochasticity.

Generative vs docking mismatch

Trigger: Generated molecules have high predicted QED but Vina score = 0.

Mechanism: Generator not aware of binding pocket geometry.

Symptom: Synthesizable but non-binders.

Fix: Add docking score (Vina or GNINA) to scoring function; dock top candidates as filter.

MolMIM -- discontinuous objective

Trigger: Optimizing a property with sharp boundaries (e.g., exactly 1 sulfonamide).

Mechanism: Latent-space optimization uses smooth gradient; sharp objectives don't have it.

Symptom: Generator oscillates around target.

Fix: Use reward-style scoring instead of property-distance; post-filter for hard constraints.

Common Errors

References

• Loeffler et al., J. Cheminform. -- REINVENT 4.
• Sanchez-Lengeling et al., ACS Cent. Sci. -- generative chemistry review.
• Jin et al. -- JT-VAE.

Related Skills

• chemoinformatics/qsar-modeling - Scoring models
• chemoinformatics/retrosynthesis - Synthetic feasibility
• chemoinformatics/reaction-enumeration - Library enumeration alternative
• chemoinformatics/protac-degraders - Linker design for PROTACs

Cross-references

Other skills in this category:

• admet-prediction
• conformer-generation
• covalent-design
• docking-rescoring
• free-energy-calculations
• molecular-descriptors
• molecular-io
• molecular-standardization
• pharmacophore-modeling
• pose-validation
• protac-degraders
• qsar-modeling
• retrosynthesis
• scaffold-analysis
• shape-similarity
• similarity-searching
• substructure-search
• virtual-screening

Changelog

• 1.1.0 (migration) — Bulk-migrated to v0.4.0 schema: canonical metadata block, base Hard rules, Cross-references. Body content unchanged; author should review and fill in any domain-specific extensions to the Hard rules.
• 1.0.0 — Initial release.