The IBM RoboRXN in Zurich runs on artificial intelligence. The researchers hope to find drugs to treat diseases much faster with the project.
Image: IBM

The synthesis of organic molecules is integral to scientific and commercial product development ranging from aspirin to nylon. Up until now, this process has been highly manual and labor intensive, and it has consisted of multiple steps. The process slows down development and can be frustrating—especially when trying to develop therapeutic treatments and vaccines for illnesses like COVID-19.

This is why the debut of IBM’s RoboRXN for Chemistry last week was so exciting.

SEE: TechRepublic Premium editorial calendar: IT policies, checklists, toolkits, and research for download (TechRepublic Premium)

Teodoro Laino, a chemist and researcher for IBM Zurich, works on the RoboRXN project.
Image: IBM

In a cloud-based demo conducted from Zurich, Switzerland, an IBM science and research team demonstrated the end-to-end development of a molecule through a successful integration of collaborative research, artificial intelligence (AI) and machine learning (ML), and robotic automation in a lab that performed the actual physical blending and testing of molecular components.

Chris Sciacca, communications manager for IBM Research Europe, likened the creation of a new molecule to the baking of an apple pie. “You have ingredients like apples, sugar, flour, a binding agent, and so on,” Sciacca said. “Then you follow instructions on how to blend these ingredients to make the final pie.”

SEE: Drug development in a year? IBM bets on AI, robotics to ramp up molecule research (ZDNet)

A recipe of essential ingredients for the construction of a molecule is similar. You blend the required ingredients according to the exact increments specified. Then there is a series of actions as to how to handle and blend the ingredients, and a robot that automatically follows the specified instructions.

While this process is specifically targeted for individuals in the scientific research field, there are also general lessons for IT, AI, ML, and robotics integrators.

1. How to deal with noisy data

The preliminary step of the development of the molecule in the demo was ingesting the vast amounts of worldwide research that already had been developed by researchers and scientists. Reams of raw data were entered into the system initially, and there had to be an automated way of separating the data that was relevant to the experiment from data that was not so that data “noise” could be eliminated upfront.

SEE: Natural language processing: A cheat sheet (TechRepublic)

“We had already trained a system AI and machine learning model to examine the data,” Sciacca said. “As the raw data streamed in, the system AI and ML model would ask itself, ‘Is this data that I never learned, or that was rarely learned, or data that has been learned many times?’ If the data was rarely or never learned, it was classified as irrelevant noise and removed. If it was data that had come up frequently in the model, it was retained. This was how we performed data quality and cleaning.”

2. How to use AI in real time for high-confidence production analytics

Before the experiment, team researchers already had a profile of the molecule they wanted to construct. They used the data that was cleaned in preliminary steps through the AI model and then integrated that data with a coded “recipe” of ingredients and actions to be taken on those ingredients for molecular formation.

SEE: How to help data scientists adapt to business culture (TechRepublic)

In turn, this recipe with instructions ran the robotics in the lab. The robotics performed injections into test tubes and mixed the ingredients. During this process, real-time analytics were run.

The end goal was to identify “fingerprints” of the molecule that researchers hoped to develop. Molecule fingerprint analytics and matching in real time was what the AI performed during process execution. As this analytics step was performed, the AI assigned a confidence level on each molecular reaction that was based on the learning models it was using.

Speeding up research and development

For life science and researchers, AI, ML and robotic automation has the potential to reduce research cycles from an average of ten years and $10 million to one year and $1 million. However, even if you’re not in research or the life sciences, the potential of the data techniques and analytics IBM is using can benefit your applications.

The data noise elimination techniques and real-time analytics are likely to be integral to many automated processes in a wide array of industries.

IBM demonstrated how easy this could be in a practical context that can help AI professionals as they navigate through the many “boots on the ground” integration issues that AI, ML, and machine automation present.