Putting Chemical Knowledge to Work in Machine Learning for Reactivity - Abstract

Aug 3, 2026

Upload by Logic

Machine learning has been used for decades to study chemical reactivity in areas such as physical organic chemistry, chemometrics, and cheminformatics. Deep neural networks learn directly from molecular structures, but many chemical datasets are small, requiring chemistry-informed modeling. Performance can improve by adding molecular descriptors from computed quantum-chemical properties and by using differentiable programming to merge neural networks with chemistry and physics models. The resulting approaches are more data-efficient and generalize better to new chemical spaces, accelerating drug design, materials design, catalysis, and reactivity research.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

AccessFree to download

Abstract

Background: AI and ML in Chemistry

Deep learning shift and successes

Generalization, Chemical Space and Applicability Domain

doi:10 .2533/chimia.2023.22 Chimia 77 (2023) 22–30 © K. Jorner Putting Chemical Knowledge to Work in Machine Learning for Reactivity Kjell Jorner* Abstract: Machine learning has been used to study chemical reactivity for a long time in fields such as physical organic chemistry, chemometrics and cheminformatics. Recent advances in computer science have resulted in deep neural networks that can learn directly from the molecular structure. Neural networks are a good choice when large amounts of data are available. However, many datasets in chemistry are small, and models utilizing chemical knowledge are required for good performance. Adding chemical knowledge can be achieved either by adding more information about the molecules or by adjusting the model architecture itself. The current method of choice for adding more information is descriptors based on computed quantum-chemical properties. Exciting new research directions show that it is possible to augment deep learning with such descriptors for better performance in the low-data regime. To modify the models, differentiable programming enables seamless merging of neural networks with mathematical models from chemistry and physics. The resulting methods are also more data-efficient and make better predictions for molecules that are different from the initial dataset on which they were trained. Application of these chemistry-informed machine learning methods promise to accelerate research in fields such as drug design, materials design, catalysis and reactivity. Keywords: Digital chemistry · Machine learning · Reactivity Kjell Jorner received his PhD (2018) in computational physical organic chemistry from Uppsala University, Sweden, under the supervision of Henrik Ottosson. After postdoctoral work at AstraZeneca UK (2018– 2020) on reaction prediction models in process chemistry, he received an International Postdoc grant from the Swedish Research Council to work withAlánAspuru-Guzik at the University ofToronto (2021–2022) . His research there focused on computer-assisted design of functional molecules and catalysts using machine learning and artificial intelligence. Since January 2023, Kjell is Assistant Professor of Digital Chemistry at ETH Zurich, where he focuses his research on chemistry-informed machine learning for catalysis. Artificial intelligence (AI) is a broad research field that aims to use computers to accomplish tasks that were previously only achievable with active input from intelligent humans. Some of the most famous tools in the AI toolbox is machine learning (ML), which refers to algorithms that learn from data, and deep learning (DL), which refers to such algorithms based on neural networks (NNs) with many layers. During the last 5–10 years, interest in applying ML in chemistry has exploded, with countless research articles, reviews,[1,2] perspectives[3,4] and books. [5,6] The large interest in these methods stems from their potential to accelerate discovery and development of chemical solutions to important societal challenges and to bring these to the market faster. Sustainable energy production,[7] chemical production,[8,9] drug design,[10] and the computer-aided synthesis of (drug-like) molecules[11] are just some of the challenges where the application of AI in chemistry can make a difference. The new wave of AI methods in chemistry follows significant advances in computer science. [12] One application where DL ex- cels is image recognition (Fig. 1a), with models trained on the large datasets of the internet era.[13] These powerful models represent a shift from the previous hand-crafted, rule-based expert AI systems, to NNs that learn the rules implicitly from the data. The rationale behind this shift is that ever more flexible NNs can come up with more and sometimes better rules than experts, provided that they are given sufficient data to learn from. One example of this transition is when Google Translate went from an expert system con

FAQs

Why do machine-learning models for chemical reactivity need chemistry-informed methods?

Many chemistry datasets are small, so flexible deep learning models need additional chemical information to achieve good performance.

How is chemical knowledge incorporated into deep learning for reactivity prediction?

One common strategy adds descriptors based on computed quantum-chemical properties, and newer methods augment deep learning with these descriptors in low-data regimes.

What role does differentiable programming play in chemistry-informed machine learning?

Differentiable programming enables seamless integration of neural networks with mathematical models from chemistry and physics, yielding more data-efficient and more accurate predictions.

Aug 3, 2026

Upload by Logic