REINVENT 4 — Tool Review — Pub-AI: AI in Science

What it does

REINVENT 4 is a molecular design platform that generates new drug-like molecules optimized toward a set of properties you specify. You define scoring functions — desirable ranges for potency, selectivity, solubility, synthetic accessibility, or any computed property — and REINVENT uses reinforcement learning to iteratively generate, score, and refine molecules toward those targets.

Unlike virtual screening (which evaluates existing molecules from a library) or retrosynthesis tools (which plan routes to a specified target), REINVENT generates new molecules from scratch. It’s the core computational tool used in AstraZeneca’s small-molecule drug discovery pipeline and has been open-sourced with full documentation and training.

The current version (REINVENT 4, published in Journal of Cheminformatics 2024) consolidates several earlier REINVENT variants into a unified framework supporting:

Reinforcement learning (RL) for multi-parameter optimization
Transfer learning to bias generation toward a target chemical space
Curriculum learning to progressively tighten optimization objectives
Scaffold decoration and linker design for lead optimization from a known starting scaffold

Best for

Drug discovery chemists and computational chemists who need to generate novel compounds optimized for multiple properties simultaneously. Particularly well suited for lead optimization (you have a hit scaffold and need analogues with improved properties) and multi-parameter optimization problems where simple enumeration of analogues is insufficient.

Pricing

Free and open-source (Apache 2.0 license). Available on GitHub. Requires Python, GPU recommended for training and sampling.

Strengths

Multi-parameter optimization: specify multiple scoring functions and weight their trade-offs — REINVENT balances them during generation rather than optimizing for one property at a time
Transfer learning support: pre-train on a focused chemical library (natural products, known actives) to bias generation toward a relevant chemical space before fine-tuning with RL
Modular scoring: integrate any external scoring function — docking scores, ADMET predictors, Chemprop models, or your own experimental assay readout
Backed by AstraZeneca’s production use: the codebase reflects years of practical refinement in an industrial drug discovery context
Active community and documentation, with tutorials and example configurations in the GitHub repository

Limitations

Requires significant setup: defining effective scoring functions, tuning RL hyperparameters, and interpreting output requires both computational and medicinal chemistry expertise
Generated molecules still need synthetic feasibility evaluation — REINVENT can include a synthetic accessibility score but this is a heuristic, not a guarantee
RL-generated molecules may exploit scoring function weaknesses (“gaming” the score) — careful scoring function design and medicinal chemistry review of outputs is essential
GPU hardware needed for reasonable throughput; CPU-only training is slow

How it compares

vs.	Key difference
IBM RXN	IBM RXN plans synthetic routes to a specified target molecule; REINVENT generates new target molecules — different stages of the discovery pipeline, often used together
Chemprop	Chemprop predicts properties of existing molecules; REINVENT generates new molecules optimized toward predicted properties — complementary tools
BayBE	BayBE is a Bayesian optimization framework for experimental design across any domain; REINVENT uses RL specifically for molecular generation
Schrödinger’s Maestro	Commercial platform with broader integrated capabilities; REINVENT is open-source and ML-focused, without the physics-based simulation features

Field Guide: Chemistry
Tool: IBM RXN — for planning synthetic routes to REINVENT-generated targets
Tool: Chemprop — for property prediction scoring functions
Glossary: De Novo Design
Glossary: Latent Space