Reaction Mode for Synthesis Planning: Tools That Think in Mechanisms, Not Just Structures

The instinct when a synthesis chemist hears “reaction mode chemistry software” is to picture a retrosynthesis engine — SYNTHIA, AiZynthFinder, ASKCOS — chewing through a target structure and spitting out a tree of disconnections. That is one kind of synthesis-planning software, but it is not what most bench chemists need most days. The everyday job is closer to: draw a reaction, capture stoichiometry, predict the product or sanity-check the one you proposed, and write it down somewhere you can find it again.

This post separates the three jobs that “synthesis-planning software” actually covers, points to the tools that fit each, and gives a concrete decision rule for picking between them. The shortest version: if you are planning a route from a target you have never made, you want a retrosynthesis engine. If you are setting up tomorrow morning’s reaction, you want a structure editor with a reaction mode and a reagent table. They are not interchangeable.

The three jobs synthesis-planning software does

Practitioners use the phrase “synthesis planning” to mean three different things, and the right tool depends on which one you are doing.

1. Retrosynthetic route search — given a target structure, propose disconnections back to commercially available starting materials. This is what SYNTHIA, AiZynthFinder, ASKCOS, Reaxys, and SynRoute do. The output is a tree of candidate routes ranked by some scoring function (commercial availability, step count, predicted yield).
2. Forward reaction prediction — given reactants and conditions, predict the product. This is what most AI-augmented structure editors do (including ChemStitch’s reaction-prediction tool) and what literature-search engines like Reaxys do when they retrieve published precedents. It answers “will this work?” not “how do I get to X?”
3. Reaction capture and stoichiometry — draw the reaction you are about to run, attach reagents and amounts, get a clean record for the lab notebook. This is what ChemDraw’s reaction mode and ChemStitch’s reaction mode do. It is not predictive at all; it is the digital equivalent of writing the scheme on a whiteboard and adding the reagent table.

Conflating these is the most common reason practitioners feel like “the software does not do what I need.” A retrosynthesis engine will not help you set up Friday’s amide coupling. A reaction-mode editor will not invent a six-step route to your target. Pick the tool that matches the job.

Job 1: retrosynthetic route search

Retrosynthesis engines walk a target backwards through a library of disconnection rules until they reach a set of commercially available precursors. The differences between them come down to (a) how the rules are sourced (expert-curated vs. literature-mined vs. machine-learned), (b) how routes are ranked, and (c) whether the catalog of starting materials is realistic for your buying patterns.

• SYNTHIA (Merck/MilliporeSigma) — expert-coded rule library developed over roughly two decades; catalog of around 12 million commercially available starting materials. Strong on classical disconnections, weaker on novel chemistry the rules don’t cover. Commercial license. SYNTHIA
• ASKCOS (originally MIT) — open-source suite combining template-based and template-free models for retrosynthesis, plus condition recommendation and forward prediction. Self-hosted; the tradeoff is that you are running and updating the models. ASKCOS
• AiZynthFinder (AstraZeneca, open-source) — Monte Carlo tree search over a learned policy network. Lightweight to run, easy to retrain on your own reaction data if you have it. AiZynthFinder on GitHub
• Reaxys Predictive Retrosynthesis (Elsevier) — AI ranking on top of the Reaxys reaction database. Unique strength is grounding in published precedents; the cost is the Reaxys subscription. Reaxys
• SynRoute — uses a small library of general templates (around 263 in the published version) plus a literature-grounded reaction database to surface short, practical routes. Useful when you want a parsimonious answer rather than the largest possible tree.

Job 2: forward reaction prediction

Forward prediction asks “if I run these reactants under these conditions, what comes out?” This is the question you have when you have an idea for a reaction and want a sanity check before you spend a day at the bench. Forward prediction comes in three flavors:

• Literature lookup (Reaxys, SciFinder) — find the closest published precedent. Strong when the substrate or transformation is well-precedented; weak when you are doing something genuinely new.
• Template-based prediction — encoded mechanism templates applied to your substrate. Predictable, explainable, but limited to chemistry the templates cover.
• AI-assisted prediction — LLM- or graph-neural-network-based prediction of products from drawn reactants. Useful for triage, but the failure mode is plausible-looking-but-wrong predictions. Treat outputs as suggestions, never as ground truth.

For the practitioner who draws a structure, sketches a reaction, and wants the AI to comment, ChemStitch and a handful of other AI-augmented structure editors fill this niche. The discipline that distinguishes a useful AI-prediction tool from a misleading one is confidence labeling: results need to be marked as predicted, not as measured. Wrong predictions stated with the same visual treatment as RDKit-computed properties teach the user to trust the wrong thing.

Job 3: reaction capture and stoichiometry

This is the job most chemists actually do every week, and it gets the least attention from synthesis-planning software vendors. The workflow is:

1. Draw the reaction scheme (reactants, products, optionally a reagent or condition above the arrow).
2. Build the reagent table: name, MW, density (for liquids), equivalents, mmol, mass or volume.
3. Use the table at the bench to weigh out reagents.
4. Record actual yield and update the table with isolated yield.
5. Export the scheme + table to the lab notebook or report.

A “reaction mode” in a structure editor is the feature set that makes step 1 cleanly distinct from drawing single molecules — a persistent indicator that you are in reaction-drawing context, an arrow tool, and an export format that round-trips reactions (RXN, MOL V3000 for reactions). ChemDraw has had a reaction mode for years; ChemStitch’s reaction mode is the same idea: a persistent toolbar toggle, a context badge that updates with reactant and product counts, and RXN export. Neither predicts products; both make the reaction the unit of editing.

The reagent table (step 2) is where most editors fall short. ChemDraw has a reagent-table panel as a paid add-on; most free editors have no equivalent. Spreadsheets fill the gap, with the usual costs of formula errors and version drift. Once the reagent table exists, downstream calculations are straightforward — if you have a target equivalents value and a scale, the stoichiometry calculator handles the equiv-to-mmol-to-mass conversion. Connecting it back to the drawn reaction so MW auto-populates is what makes the workflow feel native rather than bolted-on.

How to choose

The decision rule that holds up in practice:

• You have a target you have never made and need a route → retrosynthesis engine (SYNTHIA, AiZynthFinder, ASKCOS, Reaxys). Expect to evaluate several candidate routes and pick on practical grounds the engine cannot score.
• You have a reaction in mind and want a sanity check → literature lookup first (Reaxys, SciFinder), then AI forward prediction as triage. Verify against precedent before committing material.
• You have a route and need to set up tomorrow’s reaction → structure editor with reaction mode and a reagent table (ChemDraw, ChemStitch). Predictive features are nice-to-have; reaction-aware drawing and stoichiometry are required.
• You are scaling a known reaction → reagent table tool with cumulative-yield tracking. See our companion post on cumulative yield in multi-step synthesis for how step-yield assumptions stack up across a route.

Practitioners often layer these. A typical week might look like: SYNTHIA proposes a four-step route on Monday, Reaxys verifies the proposed amide coupling has precedent on Tuesday, the bench reaction goes into the structure editor with a reagent table on Wednesday, and the isolated yield gets entered Friday. No single tool covers the whole loop yet; the working pattern is to pick the right tool for each step and accept the handoffs.

Where the gaps are

The interesting gap, and the reason newer entrants like ChemStitch position around it, is the connection between the structural data the editor already has and the quantitative workflows that follow. MW from a drawn structure should populate the reagent table automatically. The reagent ratios in a published procedure should populate as equivalents in the canonical form chemists actually use, not as raw moles. Yield calculations should chain from one step to the next without the chemist re-keying data.

None of the established tools do this end-to-end. ChemDraw has the drawing and reaction mode but the reagent-table experience is dated and the AI is absent. SYNTHIA and Reaxys are upstream of the bench. Open-source retrosynthesis tools are powerful but have no UI for the bench-side work. The opening for a new tool is to make the reaction the unit of editing — structure, reagent table, conditions, yield — and let AI handle the parts that are genuinely predictive without dressing up uncertain output as fact.