ReactionPredictor Help



There are two modes of operation: Single-step or pathway prediction.

Single-step prediction

Upon entering a set of reactants and a description of reaction conditions, the system returns coarsely filtered reactive sites and a list of highly ranked mechanistic reactions.

Pathway prediction

Upon entering a set of reactants, reaction conditions, and a target molecule, the system runs a tree-search using the single-step prediction outlined above at each step for a multi-step reaction prediction.


Reaction prediction is the problem of predicting 'productive' mechanistic reactions given a set of reactants and conditions. This webserver is an implementation of a novel machine learning approach to reaction prediction first prototyped in 2011 and expanded to radical, pericyclic chemistry in 2012. A manuscript detailing the prototype was published in the Journal of Chemical Information and Modeling. (See the References).

See the prototype for details, but briefly, the approach is based on a few key ideas:


A reaction is modeled as an interaction between electron filled (source) and electron unfilled (sink) idealized molecular orbitals. Using this representation, all polar reactions can be enumerated by simply pairing all sources and sinks over a set of reactants.


Using the existing rules-based Reaction Explorer system and manually curating reactions from graduate level textbooks, we have developed a database of several thousand known productive reactions.

Reactive Site Filtering

In order to reduce the space of possible reactions, we first coarsely filter source and sink reactive sites using machine learning techniques. These filters are trained using the curated data.

Reaction Ranking

Once the space of reactions has been filtered, we then rank the remaining reactions by productivity using machine learning ranking techniques. The ranking models are trained using the curated data.

Input description


Enter the smiles for up to two reactants. Reactant size is currently limited to a maximum 75 character smiles string. See the Chemistry covered section for details on what types of chemistry are covered.

Target (Multi-step only)

Enter the smiles for one reactant. Reactant size is currently limited to a maximum 75 character smiles string. See the Chemistry covered section for details on what types of chemistry are covered.

Reaction Conditions

Select the reaction reaction conditions you would like to use. See the Kayala 2011 paper in References for more details.

Number of reactions (Single-step only)

Enter in how many of the top ranked reactions you would like to see returned. Set to 0 to see all filtered reactions.

Chemistry type

Choose one or more chemistry types to use in the predictor. For single-step, choosing Predict will make polar, radical, and pericyclic predictions and then use a trained neural network to decide which one of the three types to use. For the multi-step reaction predictor, if more than one chemistry type is selected, then the trained neural chemistry type predictor will be used at each step to decide which path to explore.

Output description

Reactive Sites

Predicted source and sink reactive sites are represented with atoms mapped to the value 1 in two different depictions. If no image is shown, then no atoms passed the particular filter.

Filtered Reactions

The top ranked productive reactions are shown in ranked order with the most productive at the top. The table of these reactions includes an internal score (output by the machine learning ranking model), a depiction of the arrow-pushing mechanism, and a smiles and electron flow string representation of the reaction. See the Chen & Baldi paper in References for a detailed explanation of the electron flow string specification.

Pathway Reactions

In the multi-step prediction, if a pathway is discovered, the ordered steps will be shown.

Note on the images

The images for the mechanisms are generated automatically with ChemAxon Marvin. Because of the automatic image generation software, sometimes the arrows in the images are improperly aligned.

Chemistry covered

The original prototype was limited to polar mechanisms over C, H, N, O, F, Cl, Br, I, and Li. However, the latest version has no atom type restrictions and can predict polar, radical, and pericyclic mechanisms. Stereo-specific mechanisms are not proposed. Furthermore, reactant stereo-information is not used in the ranking prediction.



Email Amin with questions, comments, or problems.