IBM RXN review and guide

IBM RXN review and guide

In 2018 IBM released a web based AI system for chemical reaction predictions, free to use for all users. The system uses simplified molecular-input line-entry system or ‘SMILES’ to make predictions of synthetic routes compiles from a database of over 2 million publications. By treating the SMILES system in a similar way to translation AI systems translate between two spoken languages, it is IBM’s attempt to help chemists across all sectors to better understand the ‘language’ they try to understand with the aid of AI.
Here we will review the usability and crucially the usefulness of the system, including the newly updated retrosynthetic feature.
Source: ibm.com

Once signing up and creating a free account at https://rxn.res.ibm.com you’ll be directed to your profile page where you’ll find all of you previous projects and from here you can also start new projects. 

Frequent users of services such as reaxys and alike will be familiar with the drawing input page which is where the conversion to the SMILES input language will input your reactants into the AI program.

Reaction Predictions

For demonstration we’ve picked a simple Friedel Crafts Alkylation reaction shown in the work space, with reagents and catalysts drawn collectively in the same space

Pressing the “New Reaction” button in the top right hand corner this will begin the run of the prediction and you’ll be prompted to select the AI model version, which the most recent is preselected and assumed to be the most relevant and accurate. 

The string generated for this reaction is:


This is as aforementioned the language the imputed reaction is translated to.  

As you can see the model was able to successfully predict the outcome and gives you a confidence rating (1.00 for this prediction).  The grid shown is a depiction of the attention applied to each individual atom and allows an insight into the assumptions it has made. If the confidence is low or the prediction isn’t as expected there is an option for other outcomes to be listed, shown in the top right hand corner.

Retrosynthetic Predictions

This feature was added later and allows a target molecule to have a retrosynthetic prediction generate. The molecule is drawn in the same workspace as the reaction predictions. 

Let’s try a well know retrosynthetic example for aspirin: 

This time the “New Retrosynthesis” button is selected from the top right and the prediction will be run.

The string: C(OC(=O)C)1=C(C(C)=O)C=CC=C1 is generated for the aspirin molecule

Here you are presented with more options;

Automatic mode where you can select or remove structures or substructures to fine tune the results available to you, imputed and translated into the SMILES language from the ketcher drawing system. 

Interactive mode which has a slider where you can customise based on price of starting materials (USD per g/mL)

Automatic Mode
Interactive Mode

For this example interactive mode was selected with no stipulation on molecules cost and the retrosynthetic prediction was generated. 

After expanding the results you will see a single step retro-synthetic step analysis with multiple alternate reactions each with respective confidence levels.

If at this point there is a component of the proposed synthesis which isn’t commercially available there is an option to add an additional step for the molecule and apply a separate retro-synthetic step to you reaction.  

This reaction is now saved to your workspace for you to view later.


The application of AI technology within the laboratory environment is often difficult and creating a bespoke program for your particular uses is costly. What IBM have created is a very useful tool to make more informed synthetic choices and a great step in the wide spread implementation of these technologies to a wide audience. The software is easy to use from undergraduate student level reactions to more complex research level predictions. Some predictions however, should be taken with a pinch of salt as IBM themselves have stated the system has a bias to some inefficient reactions where better alternatives may be available. The AI model is constantly being updated however and more widespread adoption will only increase the accuracy of predictions.

A feature you may notice while using the program is the RoboRXN. This is through a collaboration with Chemspeed technologies allowing users to make predictions and subsequently running the outputted reactions on a compatible robotic device. We will cover this in a separate story with a more hands on review in the future.   

Leave a Reply

Your email address will not be published. Required fields are marked *