---
title: REINVENT Quickstart
description: 'Pick a REINVENT run type — sampling, scoring, transfer or staged learning — set molecule count and scoring, and submit; sampling needs no scaffold.'
---

You want to generate optimized molecules toward a target property profile: pick a [REINVENT](/docs/reinvent) run type, set your generation and scoring parameters, and submit — the fastest path is a Sampling run, which generates diverse compounds straight from the prior model with no scaffold required.

> **Soft-deprecated**

> REINVENT is soft-deprecated in favor of [GenMol](/docs/genmol/quickstart), the platform's default generator. Use this quickstart when you specifically need REINVENT's RL-based scoring, diversity filters, or transfer learning.

## Webapp input methods

The .bloom* webapp provides a streamlined interface for REINVENT de novo molecular design.

## Run type selection

| Run Type | Description | Use Case |
|----------|-------------|----------|
| **Sampling** | Generate molecules from prior model | Initial exploration, diverse compounds |
| **Scoring** | Score existing molecules | Evaluate compound libraries |
| **Sampling + Scoring** | Generate and immediately score | Full optimization workflow |
| **Transfer Learning** | Fine-tune model on new data | Customize for specific chemical space |
| **Staged Learning** | Progressive curriculum training | Complex multi-objective optimization |

## Scaffold source

| Source | Description |
|--------|-------------|
| **Manual Entry** | Enter scaffold SMILES directly |
| **From Previous Job** | Import from ZINC, PubChem, or other chemistry tools |

### From previous job

- **Job Output Selector**: Browse completed jobs with compound outputs
- **SMILES extraction**: Automatic extraction from job outputs
- **Compatible tools**: ZINC search, PubChem lookup, REINVENT outputs

## Webapp parameters

### Scaffold configuration (when applicable)

- **Scaffold SMILES**: Comma-separated list of scaffold structures
- **Required for**: Scoring runs; optional for sampling runs
- **Format**: Standard SMILES notation

### Generation parameters

| Parameter | Description | Applies To |
|-----------|-------------|------------|
| **Number of SMILES to Generate** | Target molecule count | All run types |
| **Population Size** | Population for evolutionary optimization | Transfer/Staged Learning, Sampling |
| **Max Generations** | Maximum optimization iterations | Transfer/Staged Learning, Sampling |

### Scoring configuration

| Option | Description |
|--------|-------------|
| **Standard** | Use default scoring function configuration |
| **Custom JSON** | Provide custom diversity filter and scoring function |

### Custom scoring (when selected)

- **Diversity Filter (JSON)**: Custom diversity filter configuration
- **Scoring Function (JSON)**: Custom scoring function with components and weights

### Job options

- **Job Name (Optional)**: Custom identifier (default: `Denovo REINVENT (timestamp)`)
- **Use Personalized Settings**: Toggle for project-specific configuration from Supabase

## Workflow integration

### Sampling + Scoring confirmation

When selecting "Sampling + Scoring" run type, a confirmation dialog appears:
- Shows number of molecules to generate
- Confirms the combined workflow
- Allows cancellation before proceeding

### Continue to scoring

After a sampling run completes:
- **"Continue to Scoring with Sampling Results"** button appears
- Automatically uses sampling output URI
- Seamlessly chains sampling → scoring workflow

## Generative molecule design with REINVENT in the platform webapp

**REINVENT 4** is AstraZeneca's modern open-source generative AI framework for molecular design, published in the Journal of Cheminformatics (2024). It uses recurrent neural networks and transformer architectures with **reinforcement learning (RL)** to generate optimized molecules compliant with user-defined property profiles.

## Citations

> Loeffler HH, He J, Tibo A, Janet JP, Voronov A, Mervin LH, Engkvist O. Reinvent 4: Modern AI–driven generative molecule design. *Journal of Cheminformatics*. 2024;16:20. doi: [10.1186/s13321-024-00812-5](https://doi.org/10.1186/s13321-024-00812-5)

> Blaschke T, Arús-Pous J, Chen H, Margreitter C, Tyrchan C, Engkvist O, Papadopoulos K, Patronov A. REINVENT 2.0: An AI Tool for De Novo Drug Design. *Journal of Chemical Information and Modeling*. 2020 Dec 28;60(12):5918-5922. doi: [10.1021/acs.jcim.0c00915](https://doi.org/10.1021/acs.jcim.0c00915)

## Key capabilities

- **De novo design**: Create molecules from scratch
- **Scaffold hopping**: Generate novel scaffolds maintaining activity
- **R-group replacement**: Optimize substituents systematically
- **Linker design**: Connect molecular fragments
- **Molecule optimization**: Multi-objective property optimization
- **Library design**: Generate focused compound libraries

Available at https://github.com/MolecularAI/REINVENT4 under Apache 2.0 license.

## Step-by-step molecule generation workflow

### 1. Access REINVENT design tools

Ask Ivy in chat to run a REINVENT job, or submit it directly with `POST /api/v1/jobs` (see the [API reference](/docs/api)). REINVENT provides advanced molecule generation capabilities using reinforcement learning to optimize compounds for specific properties and targets.

### 2. Define your design objectives

**Property optimization:**
- **Single property**: Target specific physicochemical properties (logP, MW, etc.)
- **Multi-objective**: Balance multiple properties simultaneously
- **Activity optimization**: Optimize for predicted biological activity
- **ADMET optimization**: Design for favorable pharmacokinetic profiles

**Starting points:**
- **De novo generation**: Create molecules from scratch
- **Template-based**: Use scaffold hopping from known compounds
- **Fragment growing**: Extend fragment hits into lead-like molecules
- **Analog design**: Generate analogs of existing compounds

### 3. Configure generation parameters

**Model selection:**
- **Pre-trained models**: Choose from available generative models
- **Fine-tuned models**: Use models trained on specific chemical spaces
- **Custom models**: Upload or reference trained REINVENT models
- **Model versions**: Select appropriate model architectures

**Generation settings:**
- **Number of molecules**: Set target compound count (10-10,000+)
- **Diversity controls**: Balance novelty vs. similarity to training data
- **Convergence criteria**: Define when to stop optimization
- **Sampling strategies**: Control exploration vs. exploitation

### 4. Set up scoring functions

**Built-in scorers:**
- **Property predictors**: LogP, TPSA, molecular weight, etc.
- **ADMET models**: solubility, permeability, toxicity predictions
- **Target activity**: a QSAR or docking model you supply. Its output is a prediction, and reinforcement learning will optimize against *its* blind spots — published work found RL reaching scores of 0.7-1.0 on a scoring model that rated the compounds a real project actually advanced only 0.1-0.8
- **Synthetic accessibility**: heuristic feasibility scores. These rank within a chemical series and do not transfer across series, so treat them as a triage step and settle makeability with [AiZynthFinder](/docs/aizynthfinder)

**Custom scoring:**
- **Upload models**: Import trained predictive models
- **Script integration**: Connect to custom scoring functions
- **Multi-objective**: Combine multiple scoring criteria
- **Constraint satisfaction**: Hard limits on unacceptable properties

### 5. Execute molecule generation

**Training progress:**
- **Epoch monitoring**: Track learning progress over iterations
- **Score improvement**: Monitor optimization of target properties
- **Diversity metrics**: Assess chemical space coverage
- **Convergence tracking**: Evaluate algorithm convergence

**Real-time results:**
- **Generated molecules**: View newly created compounds
- **Property distributions**: Analyze generated compound properties
- **Score progression**: Track improvement over training epochs
- **Diversity analysis**: Assess molecular diversity and novelty

### 6. Analyze and refine results

**Compound evaluation:**
- **Structure visualization**: 2D depictions of generated molecules
- **Property analysis**: Detailed property profiles for each compound
- **Scoring breakdown**: Understand how each compound was scored
- **Similarity analysis**: Compare to training data and known compounds

**Refinement options:**
- **Iterative optimization**: Continue training with updated objectives
- **Parameter adjustment**: Modify generation parameters based on results
- **Scoring refinement**: Adjust scoring functions for better results
- **Constraint modification**: Update design constraints

**Export capabilities:**
- **SMILES libraries**: Generated compound collections
- **Property data**: Comprehensive scoring and property information
- **Training logs**: Complete optimization history
- **Model checkpoints**: Save trained models for future use

## Common generative design scenarios

### Property optimization campaigns
1. Define target property profile (logP, MW, TPSA ranges)
2. Select appropriate generative model and scoring functions
3. Run optimization for 50-200 epochs
4. Analyze property distributions and select top candidates
5. Export optimized compound library for further evaluation

### Scaffold hopping studies
1. Provide starting scaffold or known active compound
2. Set similarity constraints for scaffold hopping
3. Include activity and property scoring functions
4. Generate diverse analogs maintaining key pharmacophores
5. Evaluate synthetic feasibility and drug-likeness

### Fragment-to-lead optimization
1. Start with fragment hits from screening
2. Use fragment growing objectives in scoring
3. Include binding affinity predictions for target
4. Generate lead-like molecules from fragments
5. Assess ADMET properties of grown compounds

### Multi-target optimization
1. Define multiple targets of interest
2. Create composite scoring function for selectivity
3. Include off-target activity penalties
4. Generate compounds optimized for target profile
5. Evaluate polypharmacology potential

## Quality control and validation

### Generation quality assessment

**Molecular validity:**
- [ ] All generated SMILES are chemically valid
- [ ] No unusual or reactive functional groups
- [ ] Appropriate molecular weight and complexity
- [ ] Drug-like properties within acceptable ranges

**Scoring sanity:**
- [ ] Property predictions align with expectations
- [ ] Your scoring model rates your *known actives* highly — if it does not, RL will optimize away from them
- [ ] Diversity metrics meet project requirements
- [ ] You are treating the output as design ideas for a chemist, not as candidates shown to be makeable or active ([translatability](/docs/reinvent/translatability))

### Optimization validation

**Training progress:**
- [ ] Score improvement over training epochs
- [ ] Convergence to desired property ranges
- [ ] Diversity maintained throughout optimization
- [ ] No mode collapse or repetitive structures

**Result diversity:**
- [ ] Chemical space coverage appropriate for project
- [ ] Scaffold diversity vs. focused optimization balance
- [ ] Property distribution matches target profile
- [ ] Novelty assessment vs. training data

## Integration with platform workflows

### Upstream design inputs

**Biological data:**
- **Target structures**: Use DiffDock results for binding optimization
- **Activity data**: Incorporate screening results into scoring
- **Property requirements**: Define from ADMET profiling studies
- **Synthetic constraints**: Include feasibility from AiZynthFinder

**Chemical data:**
- **Known ligands**: Use scaffold hopping from active compounds
- **Property filters**: Apply from DeepSol solubility predictions
- **SAR data**: Incorporate structure-activity relationships

### Downstream evaluation

**Property assessment:**
- **ADMET profiling**: evaluate with ADMET-AI predictions
- **Solubility**: use DeepSol for predicted aqueous solubility
- **Synthetic planning**: check accessibility with [AiZynthFinder](/docs/aizynthfinder)
- **Target binding**: propose a pose with [DiffDock](/docs/diffdock). Docking does not establish that a compound binds

**Experimental validation:**
- **Compound synthesis**: Plan routes with retrosynthesis tools
- **Biological testing**: Design assays based on predicted activity
- **Analytical characterization**: Predict properties for QC planning
- **Formulation studies**: Assess developability early

## Performance optimization

### Model selection and configuration

**Model considerations:**
- Choose appropriate model size for computational resources
- Select pre-trained models for faster convergence
- Consider fine-tuned models for specific chemical spaces
- Balance model complexity with generation speed

**Training parameters:**
- Set appropriate batch sizes for available memory
- Configure learning rates for stable optimization
- Adjust exploration parameters for diversity vs. convergence
- Monitor training stability and adjust as needed

### Computational efficiency

**Resource optimization:**
- Utilize GPU acceleration when available
- Configure parallel scoring for faster evaluation
- Use appropriate batch sizes for memory efficiency
- Monitor computational costs vs. result quality

**Scalability strategies:**
- Start with smaller runs to validate approaches
- Scale up generation size based on initial results
- Use iterative refinement rather than single large runs
- Implement result filtering to manage output volume

## Best practices for generative design

### Project planning

**Objective definition:**
- Clearly define target property profiles and constraints
- Establish success criteria before starting generation
- Consider synthetic feasibility alongside optimization goals
- Plan for experimental validation of generated compounds

**Data preparation:**
- Curate high-quality training data when fine-tuning models
- Validate scoring functions against known active compounds
- Ensure property predictions are calibrated for your targets
- Consider chemical space coverage and diversity requirements

### Generation strategy

**Iterative optimization:**
- Start with broad objectives and refine based on results
- Use intermediate evaluation to guide parameter adjustment
- Balance exploration of chemical space with focused optimization
- Consider multiple rounds of generation and selection

**Quality assurance:**
- Implement post-generation filtering and validation
- Assess synthetic accessibility of promising compounds
- Evaluate drug-likeness and developability early
- Cross-validate predictions with orthogonal methods

### Result evaluation

**Compound assessment:**
- Evaluate generated compounds against project criteria
- Consider synthetic accessibility and cost of goods
- Assess intellectual property and patentability
- Plan experimental testing priorities

**Success metrics:**
- Measure achievement of target property profiles
- Assess diversity and novelty of generated compounds
- Evaluate hit rates in downstream screening
- Track progression toward lead optimization goals

### Research reproducibility

**Documentation:**
- Record all generation parameters and scoring functions
- Preserve training data and model versions
- Document evaluation criteria and selection processes
- Generate comprehensive reports for project archives

**Collaboration:**
- Share generation protocols with team members
- Generate standardized compound evaluation reports
- Document methods for peer review and validation
- Maintain searchable compound libraries and results

## Read next