Use Reinvent Finetune Online
Commercially Available Reinvent Finetune No-Code Web Server
REINVENT 4: Generative AI for De Novo Small-Molecule Design and Fine-Tuning
REINVENT 4 is an open-source generative AI framework for small-molecule design that combines recurrent neural networks and transformers with transfer learning, reinforcement learning and curriculum learning. It was described by Hannes H. Loeffler, Jiazhen He, Alessandro Tibo, Jon Paul Janet, Alexey Voronov, Lewis H. Mervin and colleagues, with Ola Engkvist among the authors, in the Journal of Cheminformatics (2024). The article states that the software is released under the Apache 2.0 license.
Designing molecules that satisfy many objectives at once is a core challenge in drug discovery. The authors present REINVENT 4 as a reference implementation of common AI-based molecule generation algorithms, intended for education and further innovation, and note that it is also used in production for in-house drug discovery.
The Reinvent Finetune tool on Tamarind Bio focuses on adapting a pretrained generator to your own chemistry, which in the paper is done through transfer learning on a small, task-focused set of molecules.
Attribute | Detail |
|---|---|
Molecule representation | SMILES strings, tokenized from a fixed vocabulary |
Model families | Recurrent neural networks and transformers (autoregressive) |
Generators | Reinvent, Libinvent, Linkinvent and Mol2Mol |
Run modes | Scoring, sampling, transfer learning and staged learning |
Configuration | Command-line tool driven by TOML or JSON files |
License | Apache 2.0 |
Generation Modes and Fine-Tuning Strategies in REINVENT 4
Four generators for different design tasks
Reinvent: de novo design of new molecules.
Libinvent: scaffold decoration with up to four attachment points.
Linkinvent: linker design between two warheads.
Mol2Mol: similarity-constrained optimization of an input molecule; the only generator that supports stereochemistry.
How a prior is adapted
Prior models: foundation models are trained with teacher forcing on large public molecule sets, and randomized SMILES are supported.
Transfer learning: a prior is retrained on a small, task-focused set of molecules so that it generates chemistry closer to your project.
Reinforcement learning: an augmented likelihood balances a score against the prior, with the Difference between Augmented and Posterior (DAP) loss recommended by the authors.
Staged learning: multi-stage reinforcement learning implements curriculum learning, with checkpoints between stages.
Optional components include a scaffold-based diversity filter and an inception memory that replays top-scoring molecules.
What the REINVENT 4 Paper Reports
Item | Reported value |
|---|---|
Mol2Mol training pairs | Over 200 billion PubChem pairs |
Pair similarity filter | ECFP4 count Tanimoto of at least 0.50 |
Sampling from a prior trained on 1 million molecules | Hundreds of millions of unique, valid molecules |
Molecular-weight experiment | Agent steered to 1500 Da targets, then recovered molecules of 500 Da or less by epoch 550 after the score was reversed |
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with Reinvent Finetune on Tamarind Bio
Project-specific generators: fine-tune a prior on a small set of known actives so new molecules resemble your chemical series.
Scaffold hopping and R-group replacement: explore alternatives around a core, as the paper lists among supported tasks.
Library and linker design: decorate scaffolds or connect fragments.
Molecule optimization: improve an input molecule while staying within a similarity constraint.
How to Use Reinvent Finetune on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and select Reinvent Finetune.
Choose the generator: select the model family, such as Reinvent or Mol2Mol, if exposed.
Provide training molecules: upload the small, task-focused set of SMILES used for transfer learning.
Set parameters: adjust epochs, batch size and sampling temperature if exposed; the paper's default temperature is 1.
Run the job: launch it and track progress from your dashboard.
Download outputs: retrieve the fine-tuned model and any sampled molecules, if exposed.
Check the output: canonicalize and review the generated SMILES, since the authors note users handle post-processing of raw logs.
Use results downstream: score candidates with docking or property predictors on Tamarind Bio.
Things to Keep in Mind
Reinforcement learning performance depends strongly on sigma, the balance between reward and prior regularization.
Multinomial sampling may suffer from mode collapse, and beam search is more computationally expensive.
Input SMILES must match the model's fixed vocabulary.
Only Mol2Mol supports stereochemistry.
Users are responsible for post-processing outputs and for train and validation splits.
Source: Hannes H. Loeffler, Jiazhen He, Alessandro Tibo, Jon Paul Janet, Alexey Voronov, Lewis H. Mervin et al., Reinvent 4: Modern AI-driven generative molecule design, Journal of Cheminformatics 16, article 20 (2024). DOI 10.1186/s13321-024-00812-5. Paper: https://link.springer.com/article/10.1186/s13321-024-00812-5. Code: https://github.com/MolecularAI/REINVENT4