Skip to content

RetroCast: A Unified Framework for Multistep Retrosynthesis

isChemist Protocol v1.0.0 Ruff ty coverage arXiv

RetroCast is a comprehensive toolkit for standardizing, scoring, and analyzing multistep retrosynthesis models. It decouples prediction from evaluation, enabling rigorous, apples-to-apples comparison of disparate algorithms on a unified playing field.

The Problem

The field of retrosynthesis evaluation is fragmented:

  1. Incompatible Outputs: AiZynthFinder outputs bipartite graphs; Retro* outputs precursor maps; DirectMultiStep outputs recursive dictionaries. Comparing them requires bespoke parsers for every paper.
  2. Ad-Hoc Metrics: "Solvability" is calculated differently across publications, with varying stock definitions (e.g., made-to-order libraries vs. actual off-the-shelf compounds).
  3. Flawed Benchmarks: Standard datasets are heavily skewed (74% of PaRoutes routes are length 3-4), masking performance failures on complex targets.

RetroCast solves this. It provides a canonical schema, adapters for 10+ models, and a rigorous statistical pipeline to turn retrosynthesis from a qualitative art into a quantitative science.

Key Features

  • Universal Adapters: Translation layers for AiZynthFinder, *Retro*, DirectMultiStep, SynPlanner, Syntheseus, ASKCOS, RetroChimera, DreamRetro, MultiStepTTL, SynLlama, and PaRoutes*.
  • Curated Benchmarks: Reference Series (algorithm comparison) and Market Series (practical utility), stratified by route length and topology.
  • Rigorous Statistics: Built-in bootstrapping (95% CI), pairwise tournaments, and probabilistic ranking.
  • Reproducibility: Every artifact tracked via cryptographic manifests (SHA256).

Installation

We recommend using uv for fast, reliable dependency management:

# Install as a standalone tool
uv tool install retrocast

# Or add to your project
uv add retrocast

If you don't have uv, install it in one minute:

curl -LsSf https://astral.sh/uv/install.sh | sh

pip install retrocast

Getting Started

New to RetroCast? Start here:

Benchmarks

RetroCast introduces two benchmark series derived from PaRoutes:

Reference Series (ref-)

For algorithm developers

Benchmark Targets Description
ref-lin-600 600 Linear routes stratified by length (100 each for lengths 2–7)
ref-cnv-400 400 Convergent routes stratified by length (100 each for lengths 2–5)
ref-lng-84 84 All available routes of extreme length (8–10 steps)

Market Series (mkt-)

For practical utility

Benchmark Targets Description
mkt-lin-500 500 Linear routes solvable with commercial buyables (stratified)
mkt-cnv-160 160 Convergent routes solvable with commercial buyables (stratified)

See Benchmarks Guide for details.

Get Data

For the most up-to-date benchmarks and stocks, use the RetroCast CLI:

# Download a specific benchmark (includes definition + required stock)
retrocast get-data mkt-cnv-160

The complete data/ folder as used in the preprint is available at files.ischemist.com/retrocast/publication-data:

# Show available folders and their sizes
curl -fsSL https://files.ischemist.com/retrocast/get-pub-data.sh | bash -s

# Download all benchmark definitions
curl -fsSL https://files.ischemist.com/retrocast/get-pub-data.sh | bash -s -- definitions

Verify integrity against manifests:

retrocast verify --all

Visualization: SynthArena

RetroCast powers SynthArena, an open-source platform for visualizing and comparing retrosynthetic routes.

  • Compare predictions from any two models side-by-side
  • Visualize ground truth vs. predicted routes with diff overlays
  • Inspect stratified performance metrics interactively

Citation

If you use RetroCast in your research, please cite: arXiv:2512.07079

@misc{retrocast,
title         = {Procrustean Bed for AI-Driven Retrosynthesis: A Unified Framework for Reproducible Evaluation},
author        = {Anton Morgunov and Victor S. Batista},
year          = {2025},
eprint        = {2512.07079},
archiveprefix = {arXiv},
primaryclass  = {cs.LG},
url           = {https://arxiv.org/abs/2512.07079}
}

Vision: Structural AI for Chemistry

We distinguish between two fundamental classes of problems in scientific machine learning: quantitative (predicting scalar targets like toxicity or binding affinity) and structural (generating complex objects governed by an underlying grammar). Quantitative problems, analogous to early NLP challenges like sentiment analysis, are often constrained by data scarcity. In contrast, the most transformative AI breakthroughs—from large language models to AlphaFold—have occurred in structural domains.

Mastery of structure is a prerequisite for solving downstream quantitative tasks. Foundation models trained on the structure of language, for instance, now excel at sentiment analysis with little to no task-specific fine-tuning. In organic chemistry, the paramount structural challenge is retrosynthesis: designing a valid synthetic pathway to a molecule of interest. This capability is the key to unlocking critical quantitative problems like predicting synthetic accessibility, a significant bottleneck in drug discovery. Current accessibility heuristics, however, bypass the core structural challenge, relying on learned patterns that correlate with accessibility without ever generating the pathway itself.

A model cannot judge the difficulty of a journey it cannot first articulate.

Achieving structural mastery in retrosynthesis is a long journey—one that requires moving beyond fragmented data formats, inconsistent evaluation methods, and unreliable metrics. Progress demands unified, rigorous infrastructure to standardize outputs, track provenance, and measure improvements with statistical rigor.

RetroCast is that infrastructure.

License

MIT License. See LICENSE for details.

!!! note "From isChemist: Structure precedes quantity."

Essays and software that make better scientific questions possible.

Subscribe at [ischemist.com/newsletter](https://ischemist.com/newsletter), or check [service status](https://status.ischemist.com) if RetroCast docs, data downloads, or SynthArena look off.