Skip to content

Open-Source Retrosynthesis Tools: A Feature Comparison

This page compares RENKIN against three other open-source computer-aided synthesis planning (CASP) projects on verifiable, structural differences — language, dependencies, deployment model, and feature set. It does not rank success rate or speed: these tools use different stock databases, template sets, search budgets, and evaluation methodologies, so a head-to-head performance number would not be a fair comparison without a matched-condition experiment, which none of these projects (including RENKIN) has published. See RENKIN's own Benchmark page for what limited, heavily caveated performance data does exist for RENKIN specifically.

At a Glance

RENKIN AiZynthFinder ASKCOS v2 Syntheseus
Maintainer Independent AstraZeneca (MolecularAI) MIT (mlpds_mit consortium) Microsoft Research
Core language Rust Python Python (multi-service) Python
License (code) MIT MIT MIT (current v2; the archived v1 GitHub repo was MPL 2.0, with data/models under CC BY-NC-SA) MIT
Install pip / cargo / npm, single binary or wheel pip install aizynthfinder[all] Docker Compose or Kubernetes; no simple pip install pip install "syntheseus[all]"
Runs in a browser (WASM) Yes — a full client-side build No No No
Chemistry backend chematic (pure Rust, no C/C++) RDKit RDKit + trained models Depends on the wrapped model(s)
Ships its own retrosynthesis engine Yes — rules + search built in Yes — trained expansion policy + MCTS Yes — multiple built-in template-based and ML models No — orchestrates/benchmarks external models (LocalRetro, MEGAN, Chemformer, RootAligned, ...)
Search algorithm A* / beam search Monte Carlo Tree Search over a trained policy Multiple (template-based + template-free/Transformer) Pluggable — implements common search algorithms over whichever model you plug in
Custom reaction templates Yes (--templates, up to 50k SMIRKS) Yes (trainable expansion policy + custom stock) Yes, via self-hosted configuration Depends on the wrapped model
Curated per-template evidence (DOI/patent/yield/conditions) Yes — native --template-metadata sidecar; user-supplied, no bundled evidence corpus No native template-ID evidence sidecar documented in the current official docs No native template-ID evidence sidecar documented in the current official docs No native template-ID evidence sidecar documented in the current official docs
Minimum local footprint Single binary, no network calls Local Python process Per current ASKCOS v2 deployment docs: 4+ CPU cores, 32 GB+ RAM, x86-only (no Apple Silicon) for self-hosting Local Python process (plus whatever the wrapped model needs, e.g. PyTorch/GPU)

Facts above were checked directly against each project's own repository and docs (linked in the table) as of this page's last update; ASKCOS in particular has migrated from the archived ASKCOS/ASKCOS GitHub repo (v1, no longer updated per its own README) to askcos2_core on GitLab (v2, the row above) — if you land on the old repo, follow its own link to the current one. If anything else has changed since, each project's own README/docs are the source of truth.

What Each Tool Is Actually For

These aren't four competing implementations of the same idea — they solve different problems:

  • RENKIN is a small, embeddable engine: a single Rust crate compiled to a CLI binary, a Python wheel, a Rust library, or a WASM module that runs entirely in a browser tab with no server. If you want retrosynthesis search inside another tool, a browser demo, an MCP-connected AI agent, or a resource-constrained environment, this is the shape that fits.
  • AiZynthFinder is a planning tool built around trained neural expansion policies and Monte Carlo Tree Search, developed at AstraZeneca (MolecularAI) and published as open source. It's the closest architectural peer to RENKIN in scope (a standalone library you run locally), but it's Python/RDKit-based and its search relies on trained models rather than (only) hand-curated or extracted rules.
  • ASKCOS is a full synthesis-planning platform, not a library — current ASKCOS v2 is designed to be self-hosted as a multi-service application (Docker Compose or Kubernetes) with template-based and template-free forward/retro models, a reaction-condition recommender, and more. If you want an organization-wide deployed service rather than an embeddable engine, this is that shape — at the cost of a much heavier install (4+ cores, 32 GB+ RAM, x86-only, per its own deployment docs).
  • Syntheseus isn't a standalone retrosynthesis engine at all — it's a benchmarking/orchestration framework from Microsoft Research that wraps other published models (LocalRetro, MEGAN, Chemformer, RootAligned, and others) behind one common interface, so you can swap models and search algorithms and compare them fairly. If you're doing retrosynthesis-model research rather than looking for a planning engine to embed, this is the tool built for that.

Where RENKIN Is a Better Fit

  • You need the engine to run without a Python/Docker runtime — in a browser, a Rust service, or a CLI tool distributed as a single binary.
  • You want to attach curated, citable evidence you supply (a DOI, a reported yield, a known side reaction) to specific templates, not just a bare disconnection — see Reaction Evidence Metadata. RENKIN doesn't ship a bundled evidence corpus either; the sidecar is a mechanism for evidence you curate yourself.
  • You want zero C/C++ dependencies and a cargo build/pip install install story with no GPU, no Docker, and no multi-service deployment.

Where Another Tool Is a Better Fit

  • You need trained-model-driven route ranking (expansion policy + MCTS) and are comfortable with a Python/RDKit dependency stack → AiZynthFinder.
  • You're deploying a shared, organization-wide synthesis-planning service with forward prediction, condition recommendation, and more, and can dedicate the infrastructure to it → ASKCOS.
  • You're doing retrosynthesis-model research and need to benchmark multiple published models under one harness → Syntheseus.

Next Steps