Forward Synthesis

Forward Synthesis tools answer the opposite question to retrosynthesis: given a set of reactants, what happens when you run the reaction? These are supplementary analysis tools alongside the retrosynthetic tools (Interactive Planner, Tree Builder), useful for checking a route you've already planned or exploring a reaction's real-world behavior.

Access

Chemistry Tools requires an active subscription to any subject — see Chemistry Tools Overview for the full access model.

Opening Forward Synthesis

Click Chemistry Tools in the sidebar — it opens in a new browser tab — then open Forward Synthesis from the tools menu.

Available analyses

Forward Synthesis bundles five specialized prediction tools:

Tool	What it predicts
Product Prediction	The major product formed from a given set of reactants.
Condition Recommendation	Likely reagents/conditions (solvent, catalyst, temperature) for a reactants → target transformation.
Impurity Prediction	Likely by-products/impurities, including over-reaction scenarios.
Regioselectivity Prediction	Which position on a molecule a reaction is likely to favor.
Site Selectivity Prediction	Selectivity specifically for aromatic C–H functionalization reactions.

Product Prediction is the most reliable of the five — it's the most well-studied task, with models reaching over 80% top-1 accuracy on standard benchmarks. The other four are more specialized, trained on smaller or noisier datasets, so treat their output as a useful hint rather than a certain answer — always sanity-check against your own chemical reasoning.

Product Prediction walkthrough

Enter your reactants as a SMILES string (or draw them with Ketcher). For a mixture of reactants, separate them with a . in SMILES — for example, acetic acid and propan-1-ol as a Fischer esterification pair.
Check the model settings — you can select:
- The Condition Recommendation model, if you want suggested conditions alongside the product.
- The forward prediction model variant and its training dataset source. Defaults are reasonable for a first attempt.
Run the prediction. Results come back as a ranked list of possible products, each with a confidence/probability score.

Worked example: acetic acid + propan-1-ol predicts propyl acetate (ester) as the top-1 product, with high confidence (around 0.94 in typical runs) — a textbook Fischer esterification.

Using the other four tools

Condition Recommendation, Impurity Prediction, Regioselectivity Prediction, and Site Selectivity Prediction follow the same basic pattern as Product Prediction — enter your reactants/target, run the tool, read the ranked output — but keep their more limited training data in mind:

Condition Recommendation is most useful as a starting point for literature comparison, not a definitive answer.
Impurity Prediction is best used to sanity-check a route for known problem patterns (e.g. over-oxidation, over-alkylation) rather than an exhaustive impurity list.
Regioselectivity and Site Selectivity predictions are most reliable for reaction types well-represented in the training data (e.g. common aromatic functionalizations for Site Selectivity) — treat unusual substrates' predictions cautiously.

Tips

Use Product Prediction to double-check a step you found via Interactive Planner or Tree Builder — running the forward direction on a proposed disconnection is a quick sanity check that the reaction actually goes the way the retrosynthesis suggested.
A low top-1 confidence score doesn't necessarily mean the reaction fails — it can mean the model is genuinely uncertain between several plausible products. Check the next few ranked results, not just the top one.

Troubleshooting

Symptom	Likely cause
No product predicted	Check the reactants' SMILES is valid and complete (all reactants included, separated by `.`).
Low-confidence or implausible top result	Try the next-ranked results — the correct product may be ranked 2nd or 3rd, especially for less common reaction types.
A prediction seems to hang	Wait briefly and retry — if it persists, reload the Chemistry Tools tab.

Still stuck? See Chemistry Tools Overview for general troubleshooting, or email support@examcatalyst.ai.

Advanced: a closer look at each tool

Each of the five tools has its own settings and best-use pattern worth knowing once you've run the basic Product Prediction walkthrough above.

Condition Recommendation, in detail

Unlike Product Prediction, Condition Recommendation needs both reactants and the expected/desired product as input (e.g. CC(=O)O.CCCO reacting to give CCCOC(C)=O) — you're asking "what conditions would make this specific transformation happen," not "what happens with these reactants."

Two model choices are available under Settings:

Neural Network — predicts solvent, reagent, catalyst, and temperature, plus a solvent "greenness" score.
Neural Network v2 — predicts reactants/reagents together with their stoichiometries and temperature; you additionally choose between a graph-based or fingerprint-based variant.

Results are ranked recommendations, each shown with a solvent score on a 1–6 scale, where 1 is the greenest/safest solvent choice — useful if you're comparing options with sustainability in mind, not just feasibility. You can adjust how many recommendations come back (10 by default). Each recommendation links directly into Product Prediction so you can immediately verify that the suggested conditions actually give your intended product.

Product Prediction — model choices

Three underlying model variants are available for Product Prediction, each a different machine-learning approach to the same question:

wldn5 — a template-free Weisfeiler-Lehman Network model, reasoning directly about bond changes.
graph2smiles — a template-free, end-to-end graph-to-SMILES model.
augmented_transformer — a template-free transformer-based model.

Results show rank, structure, a probability score, a "Max Score" (which accounts for cases where multiple predicted structures collapse to the same canonical SMILES), and molecular weight — plus direct links onward into Impurity Prediction and Regioselectivity Prediction for the same reaction.

Impurity Prediction, in detail

Impurity Prediction checks for five different impurity-formation modes, not just one:

Minor product — lower-ranked candidates from Product Prediction's own top-k output.
Over-reaction — where the intended product itself reacts further, becoming a reactant in a follow-on reaction.
Dimerization — two molecules of the same reactant/product combining.
Solvent adduct formation — the solvent itself reacting into the product.
Subset reactant reactions — a subset of the full reactant list reacting on its own.

Configurable settings include the top-k number of forward predictions considered per mode (default 3), an inspection threshold that filters out low-scoring candidate impurities (default 0.1), a choice of inspector model used to verify each candidate, and an atom-mapping toggle to help confirm the proposed mechanism visually.

Don't compare impurity scores across formation modes

Different impurities come from different reactant/reagent combinations and different formation modes — their probability scores are not directly comparable to rank "most likely impurity" against each other. Use the scores to flag candidates worth a closer look, not as a precise likelihood ranking.

Regioselectivity Prediction, in detail

Built on a QM-GNN model — combining graph-based reasoning (WLN) with quantum-chemical descriptors in one multitask network. You supply reactants and the product skeleton; you can optionally exclude reagents from the atom-mapping step. Results are ranked regiochemical outcomes with probability scores, and can be exported to CSV for further analysis outside the tool.

Site Selectivity (Aromatic C–H Functionalization), in detail

Also a multitask WLN-based model, specifically for predicting which site on an aromatic ring is favored in a C–H functionalization reaction:

Enter your substrate — aromatic carbon atoms eligible for functionalization are highlighted in blue on the drawn structure.
Click a highlighted atom to see the most probable functionalizing reactants for that specific site.
Use the Filter Reactants box to score site-selectivity for a specific reactant you already have in mind, rather than browsing all possibilities.

Suggested workflow

Run Product Prediction first to confirm the expected major product.
If you need real-world conditions, follow up with Condition Recommendation using your reactants and that predicted product.
Use Impurity Prediction on a route you're about to commit to, to catch likely by-products before they surprise you in practice.
Reach for Regioselectivity / Site Selectivity specifically when your reaction has more than one plausible position for reaction to occur — most reactions won't need them.