Google DeepMind Releases Chemistry Research Tools for Wrangling Molecular Data
The company released Cheminformatics Skills, a toolkit that helps researchers find molecular data and perform molecular calculations. The AI coordinates the order of operations while chemistry-specific tools handle the molecular calculations.
Google DeepMind introduced Cheminformatics Skills, a toolkit for chemistry research, on November 3 local time. Designed for use in Google Antigravity, it helps with the computational handling of chemical information. Google DeepMind says the goal is to cut down the time researchers spend wrangling data so they can focus on forming hypotheses.
Why is organizing molecular data so hard?
In chemistry research, the same molecule is written differently depending on the source. In papers it appears as a systematic compound name. In patents it shows up as a scheme that bundles many structures together. In vendor catalogs it arrives as a structure data file. Researchers have to connect these scattered representations to one another and filter out errors, because a difference of a single atom or a single bond makes it an entirely different compound.
Image: Google DeepMind
The new toolkit brings AI and chemistry-specific tools into the research workflows where this kind of data is handled. Google DeepMind is targeting both computationally oriented chemists and chemists who run experiments at the bench. The idea is to create a collaborative research environment with access to the tools already on a researcher’s computer. The focus is on binding the data a researcher handles and the computational tools they use into a single workflow.
The AI and the computational tools divide the labor. The language model coordinates the order of operations, searches databases, and supports scientific judgment. The molecular calculations are handled by chemistry tools that run on fixed rules. The skills are built on RDKit, an open-source chemistry toolkit, and also connect to ChEMBL and PubChem, which compile compound information. Google DeepMind says it has baked recommended practices from the field of chemistry into this process.
What do you gain by restricting a molecule’s motion?
As a use case, Google DeepMind pointed to a drug design technique that limits a molecule’s motion. A drug molecule can rotate around some of its bonds and adopt many three-dimensional shapes. To bind to its target, it has to hold one particular shape. The energy cost of giving up that freely moving state can weaken binding. That is why motion is taken into account when designing molecular structures.
Image: Google DeepMind
The approach here is to add new bonds that narrow the range of motion available to the molecule. Creating a bond in the right place can help the molecule hold the shape best suited to binding its target. Google DeepMind cited lorlatinib, which is structurally similar to crizotinib, a drug used to treat some non-small cell lung cancers. Crizotinib has 5 major rotatable bonds. Lorlatinib, the company explained, has a ring structure that reduces the effective number of rotatable bonds to 0, which translated into benefits such as improved potency.
The toolkit aims to ease the burden on researchers of organizing molecular data and setting up calculations. Google DeepMind has put forward an approach that combines the AI’s ability to coordinate tasks with the calculations done by chemistry-specific tools. The questions now are how much it shortens data processing time in real research, and how the accuracy of converting between molecular representations will be verified.