Reaction prediction & retrosynthesis planning

Deciding how to make a molecule is a slow, expert-driven search through millions of known reactions. Computer-aided synthesis planning (CASP) tools search backwards from the target to purchasable building blocks, rank routes and predict whether each step will work. Chemists use them to shorten route design, find cheaper or greener alternatives and feed automated labs. Output is always a hypothesis to be tested at the bench.

Problem & value

• Today: medicinal and process chemists design routes from experience and literature searches (SciFinder, Reaxys). Route design for complex targets can take days to weeks; failed steps cost lab time and material.
• AI value: broader search than one person can do, fast ranking by steps, cost and risk, discovery of non-obvious disconnections, and reuse of company-internal reaction data.
• Benefit evidence: published work on expert-coded systems (Synthia, formerly Chematica) and template-free and template-based neural planners shows routes that experts judge viable for many drug-like targets, but benefit is measured in route-design time and fewer failed experiments, and no universal accuracy figure exists; benchmarks are mostly retrospective on patent data [1][2][3].

Variants

• Medicinal chemistry: fast routes to analogues from catalogue building blocks, tied to generative design (SCI-002).
• Process chemistry: cost, safety, solvent and impurity-driven route selection; scale-up.
• Materials / specialty chemicals: monomers, dyes, catalysts, agrochemicals (SCI-003, SCI-014).
• Forward prediction: product, selectivity and yield prediction for a proposed step, plus condition recommendation and Bayesian reaction optimisation.
• Academic and teaching use: open tools for synthesis-planning courses and research groups.

How it works

1. target molecule
2. single-step retrosynthesis model proposes precursors (template-based, template-free or LLM)
3. multi-step search (Monte Carlo tree search / A* / retro*) toward purchasable building blocks
4. scoring: step count, cost, availability, reaction feasibility, filters (safety, regulated reagents)
5. forward-prediction check, condition and yield models
6. chemist review and edit of the route tree
7. ELN record / execution on lab automation (SCI-007)

• Single-step models: reaction templates extracted from databases (e.g. AiZynthFinder), or sequence/graph neural nets (Transformer SMILES-to-SMILES, graph edit models) that need no templates.
• Search: tree search guided by learned policy and value networks; building-block stock defined by a catalogue or in-house inventory.
• LLM role: reasoning models read literature, explain a route, suggest conditions or call CASP tools as agents; they are not reliable standalone predictors of reaction outcomes.
• Human-in-the-loop: a chemist reviews every route for chemistry validity, safety, selectivity, protecting groups and scalability before anything is run.

Data preparation

Inputs are target molecules, reaction records (own ELN entries, literature, patents) and the building-block inventory. They live in ELN/LIMS, chemistry databases, inventory or ERP systems and PDF reports.

Typical gaps: reaction entries without yields, conditions or reagents in structured form, failed reactions not recorded, building-block stock lists in Excel with free-text names and no structures, and duplicated or inconsistently drawn molecules.

Preparation steps: 1. Inventory reaction and stock data sources and owners (PREP-001, PREP-004). 2. Export ELN and ERP data into tabular form (PREP-005). 3. Convert building-block names into standardised structures and keep stock lists current (PREP-011, PREP-012). 4. Extract reactions from PDFs where needed (DOC-014). 5. Harvest outcome labels such as yield and success from existing records (PREP-020). 6. Keep a test set of known routes for evaluation (PREP-018).

Effort share: medium. Licensed tools bring their own reaction data, so preparation mainly concerns the in-house building-block inventory and, if used, own reaction data for fine-tuning. See GUIDE-003.

Implementation paths

Which path for whom

• Enterprise: Path A - mature licensed retrosynthesis tools (e.g. Synthia, SciFinder, Reaxys-based planners) fit a commodity planning task and need little integration; choose B/C only for proprietary reaction data.
• Developer: Path C - AiZynthFinder and other open tools run on modest hardware (HW2) and can be tailored to own building-block stocks.
• Public: Path C - universities can run open tools on-premises and avoid licence costs and sending unpublished structures to third parties.

Path A — Buy (recommended for enterprises)

• Synthia (Merck KGaA / MilliporeSigma, EU vendor; expert-encoded rules plus ML; on-prem or cloud).
• Reaxys Predictive Retrosynthesis (Elsevier, built on Iktos) and CAS SciFinder retrosynthesis (CAS), integrated with large curated reaction collections.
• Spaya (Iktos, France; EU vendor) for API-driven retrosynthesis and synthetic accessibility scoring.
• IBM RXN for Chemistry (cloud, free tier and API), ChemAxon and Schrödinger tooling, and Postera Manifold for catalogue-driven route and building-block sourcing.
• Choose A when you need validated, supported tools with curated data and ELN integration. Check licence terms on use of the vendor's reaction data and on where queried structures are stored, since target structures are confidential IP.

Path B — Build on cloud APIs

• Call vendor APIs (IBM RXN, Spaya, Postera) from internal tools; combine with frontier reasoning LLMs (Claude, GPT, Gemini, Mistral on EU regions) for literature-grounded explanation and agentic orchestration, with RDKit as a deterministic validator of every proposed structure.
• Choose B for custom workflows, e.g. batch route scoring for virtual libraries.

Path C — Self-host (open source)

• AiZynthFinder (AstraZeneca, MIT-style open source; template-based MCTS), Syntheseus (Microsoft, benchmarking and wrapping of single-step models), ASKCOS (MIT, open-source suite incl. forward prediction and condition recommendation), Molecular Transformer, LocalRetro, RetroBioCat (biocatalysis), RDKit and Open Reaction Database data.
• Training data: USPTO patent reactions (open) or licensed Pistachio / Reaxys extracts; building blocks from vendor catalogues (e.g. eMolecules, Enamine, Sigma-Aldrich).
• Hardware: inference and search run on CPU or HW1; HW2 (one GPU) is enough for training and serving Transformer models. Ops effort is moderate; most of the work is data curation.

Path D — Custom training

Worthwhile for large pharma and chemical companies with decades of internal ELN reaction data: fine-tune single-step and yield models on proprietary (including failed) reactions, which public data lack. Needs thousands to hundreds of thousands of curated, atom-mapped reactions with outcomes. Standardise reaction records first (Open Reaction Database schema, SURF format).

State of the art (2026)

• Mature: multi-step search over template-based and neural single-step models solves many literature-like targets; Synthia and open tools have shown wet-lab-validated routes [1][2][3].
• Trend: LLM-based reasoning and tool-using chemistry agents, reaction foundation models, and closed loops linking planning to robotic synthesis and Bayesian optimisation (SCI-007).
• Data bias: public data are dominated by successful, patent-style reactions; negative results, yields and conditions are scarce, so yield and selectivity predictions are weak outside the training distribution.
• Limits and failure modes: models propose chemically implausible or non-selective steps, miss protecting-group and stereochemistry issues, overestimate feasibility for novel scaffolds, and evaluate poorly against single "ground-truth" routes (many valid routes exist). Treat scores as ranking aids, not probabilities of success.

Shared resources & combinations

• Combines with: SCI-002 (generate molecules, then check synthesisability), SCI-003 and SCI-014 (materials and formulation), SCI-007 (route to robotic execution), SCI-018 (analytics to confirm products), SCI-015 (toxicology screening of intermediates and reagents), KNOW-008 (literature).
• Shared building blocks: scientific foundation models and graph ML, RDKit cheminformatics stack, reaction database and knowledge graph, ELN integration; one GPU server (HW2) can host these models together with SCI-002 and SCI-003 workloads.

Effort, cost & KPIs

• Path A: weeks to onboard; commercial licences per seat or site, typically a significant annual software cost; reaction-database access is a major cost driver.
• Path B: 1-3 months to integrate APIs and ELN.
• Path C: 2-4 months with a cheminformatics engineer; mostly data and building-block stock set-up.
• Path D: 6-12+ months; requires data engineering of the internal reaction archive.
• KPIs: route-design time, share of proposed routes judged viable by chemists, experimental success rate of first-pass steps, number of steps and cost per target, share of routes from in-stock building blocks.

People & change

Involve early: management and the chemistry department lead decide scope and licences. Data owners (ELN, stockroom/purchasing) deliver inventory data. IT handles accounts and integration; legal/compliance checks licence terms, data sharing with the vendor and dual-use or controlled-substance aspects. Synthetic chemists as end users should test routes in a pilot.

Works council: depends. The tool does not evaluate staff, but licensed cloud tools often log user queries; if such logs allow monitoring, §87(1) no. 6 BetrVG applies, as the Federal Labour Court treats objective suitability as sufficient. Inform the works council in good time (§90(1) no. 3 BetrVG); public bodies involve the Personalrat. See GUIDE-004.

Data protection: no personal data in the molecules; check accounts and logs with the DPO. Check that vendor terms do not claim rights to your structures (GUIDE-010).

AI literacy: support chemists with training on route plausibility checks and typical failure modes such as invented reactions (GUIDE-006).

Communication and pilot: test the tool on three to five known syntheses with a few volunteer chemists, compare with the published route and report results (GUIDE-009, GUIDE-008).

Typical concerns: loss of craft and doubts about route quality. Chemists stay in control: the tool proposes, the chemist judges feasibility and safety.

Risks, regulation & compliance

• EU AI Act: minimal risk; no Annex III category. AI literacy duty (Art. 4) applies. Use inside a regulated product (e.g. GMP process development) does not make the tool itself high risk.
• Dual use (important): synthesis-planning tools can in principle assist production of toxic or controlled substances. Operators should screen targets and routes against controlled-substance and chemical-weapons lists (Chemical Weapons Convention schedules, EU precursor and drug-precursor rules, Australia Group lists), apply access control and logging, and restrict public-facing exposure. Published studies on LLM chemistry agents recommend such safeguards. This file deliberately omits operational detail.
• Chemicals law: REACH and CLP duties apply to the substances made, not the planning tool; coordinate with SCI-015 and DOC-023.
• GDPR: generally no personal data (molecular data); user logs and ELN entries may contain employee data.
• IP and confidentiality: target structures are trade secrets; prefer on-prem, EU-hosted or contractually isolated tools; check vendor terms on training with queries.
• Safety: predicted routes may involve hazardous reagents or exotherms; independent hazard review (process safety) is required before any scale-up.

Getting started

1. Pick 20-30 past in-house targets with known routes and compare tool output with them.
2. Define a building-block stock (in-house inventory plus 1-2 supplier catalogues) and exclusion rules.
3. Run AiZynthFinder or ASKCOS locally as a baseline, then trial one commercial tool (e.g. Synthia, Spaya).
4. Set up review workflow and access/audit controls, including target screening for controlled substances.
5. Capture executed routes and failures in the ELN in a structured format to improve models later.

Sources

1. Klucznik et al., "Efficient syntheses of diverse, medicinally relevant targets planned by computer and executed in the laboratory", Chem 2018 — https://doi.org/10.1016/j.chempr.2018.02.002
2. Genheden et al., "AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning", J. Cheminform. 2020 — https://github.com/MolecularAI/aizynthfinder
3. Maziarz et al., "Re-evaluating retrosynthesis algorithms with Syntheseus", Faraday Discussions 2025 — https://github.com/microsoft/syntheseus
4. Tu et al. / Coley group — ASKCOS open-source synthesis planning suite — https://askcos-docs.mit.edu/
5. Schwaller et al., "Molecular Transformer", ACS Cent. Sci. 2019 — https://doi.org/10.1021/acscentsci.9b00576
6. Merck KGaA — Synthia retrosynthesis software — https://www.merckgroup.com/en/research/ai/synthia.html
7. Iktos — Spaya retrosynthesis — https://iktos.ai/
8. Open Reaction Database — https://open-reaction-database.org/
9. Bran et al., "Augmenting large language models with chemistry tools" (ChemCrow), Nature Machine Intelligence 2024 — https://doi.org/10.1038/s42256-024-00832-8
10. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj