Exploring MIT’s ASKCOS Tool
A user-friendly program that uses advanced machine learning to assist in organic synthesis planning
Chemical compounds lie at the heart of all chemical engineering jobs and projects. Specifically, organic compounds are crucial to multi-billion dollar companies in the pharmaceutical, speciality chemicals, and food processing industries. For every compound that is to be manufactured, a detailed synthesis plan has to be made involving reagents, catalysts, and reaction conditions. In this post I explore ASKCOS, a new online tool from MIT that utilizes neural networks to automatically plan organic syntheses.
ASKCOS, which stands for Automated System for Knowledge-based Continuous Organic Synthesis, is a research project hosted by the Jenson Lab at the Massachusetts Institute of Technology. When you visit the main website, you first have to create an account and get it approved via email. Then, after logging in, there are a set of tools that are available to tinker with, which are called ‘modules.’ I experimented with the One Step Retrosynthesis, Context Recommendation, and Reaction Evaluation modules.
One-step Retrosynthesis
The first module that I opened was One-step Retrosynthesis, which inputs a target compound and outputs precursor compounds that react to produce the target in a single step reaction. The main page for this module looks like the following:

Essentially the only user parameter needed is the identity of the compound that you wish to synthesize. For popular chemicals such as aspirin (which was my choice), atropine, etc. simply typing the common name will work, but otherwise you can draw the compound using the tool provided. The other parameters are for advanced tweaking, and I left them at their defaults.
A neural network and a custom heuristic combine to form the heart of the prediction algorithm. The neural net is trained on thousands of predefined reaction boilerplates (‘templates’) and is used to identify which templates are most relevant for the target compound. Then, a heuristic is used to pick precursors according to the following two conditions:
- “Buyability” — whether or not the chemical is available to purchase, and at what price
- Simplicity — the more stable the compound, the better
After entering the target compound, all that’s needed is to hit the search button, and the program outputs a result with multiple solutions, similar to the picture below.

The results are ranked according to a score generated from the neural net and heuristic. As you can see, the program decided that Acetic anhydride and Salicylic acid were the two reactants best suited to form aspirin. The prices of the chemicals is also provided, or else ‘not buyable’ is specified.
Context Recommendations
After completing the One-step Retrosynthesis of your desired compound, it is possible to get more details of the reaction through the Context Recommendations module. To go to the module, I clicked the ->? icon at the end of the first reaction and was redirected to a page that included the following information:

ASKCOS uses a (different) neural network to predict which conditions would be best for carrying forward the desired reaction. In my case, it suggested that mixing the reactants at 108ºC would be enough to produce aspirin. The fact that multiple configurations are presented is nice, as it provides options depending on what may be available in a particular scenario.
Synthesis Predictor
The final module that I used in my aspirin synthesis experiment was the Synthesis Predictor. As the name suggests, the program predicts the products of a reaction given the reactants, reagents, solvents, etc. As ASKCOS had informed me that I needed only Acetic anhydride and Salicylic acid, I entered them into the reactants box and pressed ‘Go’. Sure enough, aspirin was the first result with a probability of 80%.

Even my small experimentation with ASKCOS was enough to show how powerful it is in assisting with charting organic synthesis pathways. With better development of multi-step syntheses (the “Tree Builder” module), it will become an indispensable tool to chemists and chemical engineers who want to save time planning reactions.
I think the possibilities of the project really expand when it is combined with other computational tools. For example, there is a lot of research being done into using deep learning to generate new drugs and pharmaceutical molecules. Combining that tool with this one could mean not only the discovery of potentially life-saving medicine, but also an insight of how to synthesize this medicine from pre-existing compounds under ideal conditions.
It’s really exciting to see labs at schools such as MIT actively harnessing machine learning to improve the chemical industry. I hope that I get the chance to experiment with more amazing programs such as ASKCOS in the very near future.
