RESEARCH

Automated discovery of toxic chemicals in complicated mixtures with machine learning and complementary analytical techniques

Anneli Kruve, Jonathan W. Martin, Matthew MacLeod, Frederic Been, Isak Samsten

FORMAS 2026 – 2030

Advanced treatment is soon to be required in the EU to mitigate the impact of micropollutants in wastewater. Identification of the chemicals contributing to the toxicity of sludge, waste- and raw drinking water is essential for intelligent design of water treatment solutions. Unfortunately, only a small fraction of chemical mixture toxicity of wastewater can be explained by known micropollutants monitored today. We have shown that nontarget screening with liquid chromatography high resolution mass spectrometry (LC/HRMS) in combination with machine learning allows rapid ecotoxicity and exposure evaluation of detected chemicals. Importantly, our model MS2Tox enables to overcome the need for unequivocal structural annotation that was hitherto impeding the toxicity assessment. Nevertheless, the toxicity assessment is limited by (1) the detection of chemicals causing the toxicity and (2) the prediction accuracy of the machine learning models. Here we will close the gap in accounting for the effect of polar and (semi-)volatile chemicals by expanding MS2Tox for chemicals detected with hydrophilic interaction chromatography and gas chromatography coupled to HRMS. Simultaneously, interpretable machine learning expedites understanding of the knowledge learned by MS2Tox and its gaps. Furthermore, we are going to leverage the expanding toolbox of HRMS and MS2Tox to assess the toxicity of the whole range of chemicals present in waste- and raw drinking water as well as sludge in Sweden.

Lignin2MS: Unravelling lignin structure with mass spectrometry and generative
machine learning

Eliise Tammekivi, Mika Sipponen, Anneli Kruve

Wenner-Gren 2024-2026

Lignin is one of the main components of lignocellulosic biomass (e.g., wood, energy crops, stalks, leaves), the most abundant natural material on earth. Cellulose and hemicellulose are isolated from the lignocellulosic biomass in paper and ethanol industries, and the residual lignin is burned for heat and power production. However, lignin holds potential for sustainable production of chemicals, biofuels, or functional biomaterials, as it is the most abundant natural resource for aromatic compounds.1,2 Owing to its heterogeneous and complex structure the valorisation processes remain inefficient today. Establishing structure-function-performance correlations is severely hampered due to the missing structural information of lignin. These obstacles hamper effective valorisation of lignin that potentially contains biologically active (antimicrobial, antioxidant, UV-protective, etc.) molecules that could be discovered and more efficiently extracted from complex mixtures if their structures can be predicted.

LearningStructurE: Machine Learning and Mass Spectrometry for Structural Elucidation of Novel Toxic Chemicals

Anneli Kruve

ERC CoG 2024 – 2028

Nearly half a million known chemicals have been deemed relevant for exposure studies and an even larger number of their transformation products are likely to co-occur in the environment. This mind-blowing number of possible chemical structures makes it impossible to in-silico generate all these structures, let alone synthesise and analytically confirm them, thereby limiting the discovery of novel chemicals. Today, the structural elucidation of chemicals detected with high resolution mass spectrometry relies on databases and machine learning models trained on the known chemical space. Both are fundamentally ill-suited for discovering novel chemical structures. As a result, only a few percent of the toxic activity of the environmental samples is explained by the currently known and monitored chemicals. It is crucial to access the novel chemical space to improve our understanding of the origin, fate, and impact of these chemicals.
The aim of LearningStructurE is to turn the discovery of novel chemical structures from serendipity to routine. As a steppingstone in this pursuit, I will combine the fundamental understanding of chromatography and high resolution mass spectrometry with machine learning to pinpoint novel toxic chemical structures based on their empirical analytical information. To significantly advance the predictive power of machine learning models for empirical analytical information, I will take advantage of the candidate structures as a sample specific training set for machine learning models. The improved predictive power will feed into in-silico structure generation, allowing to elucidate the structure directly from the empirical analytical information.
LearningStructurE will pave the way for exploration of the unknown chemical space detected from environmental samples, and thereby improve our understanding of the emissions, chemical processes transforming the emitted chemicals, and close the gap in measured and explained toxicity.

Mobility grant: Development of rapid chemical screening techniques in recycled textiles for a non-toxic future

Stockholm University & University of Copenhagen for Drew Szabo

VINNOVA 2023

Upcycling and recycling textile into new materials will be a key strategy for the sustainable development of textile production in the future. However, recycling process may introduce and concentrate potentially harmful chemicals from the feed material and the risk of potentially harmful chemicals in recycled textile needs to be assessed. The aim of this project is to evaluate desorption electrospray ionization (DESI) technique for its ability to rapidly identify and quantify potentially hazardous substances from recycled textiles.

MS2Tox: Deep Learning for Automated Prediction of the Endocrine Disruptive Potency of Chemicals in Complex Mixtures

Anneli Kruve, Jonathan Martin, Stefano Papazian, Ulf Norinder (SU), and Swapnil Chavan, Ian Cotgreave (RISE)

VR 3R grant 2023 – 2025

The safety of water, food, and new materials used in daily life is essential for human life and the ecosystem. Nontarget liquid chromatography high resolution mass spectrometry (LC/HRMS) is increasingly used to detect chemicals in such samples. Here we will develop machine learning methods for, evaluating the hazard possessed by the chemicals in these complex mixtures, especially endocrine disruptive potency. Up to now the main limitation in applying machine learning has been that for predicting toxic endpoints of complex mixtures the single chemical constituents in the mixture need to be first unequivocally identified. Here we will predicting the hazard of chemicals directly from the empirical spectral information acquired rapidly in nontarget LC/HRMS chemical analysis.

Self-learning artificial intelligence for detection of toxic chemicals in recycled materials

Anneli Kruve (SU)

Carl Tryggers Foundation 2023 – 2024

Machine learning could contribute enormously is identifying small potentially toxic chemicals from big data acquired from the analysis of recycled materials with LC/ESI/HRMS. Today, the unawareness of possible chemical risk possessed by the chemicals from the highly varying source material limits the safe usage of recycled materials. To overcome this risk and enable efficient and safe recycling, we need to be able to guarantee that no toxic chemicals are present in the final product. Analytical methods to detect such chemicals exist. What is lacking is a fast way to pinpoint if and which potentially toxic chemicals are present. Here we are developing new workflows for identifying the chemicals detected with LC/ESI/HRMS.

Toxicity guided inverse design of materials for environmental remediation

Anneli Kruve, Aji Mathew, Isak Samsten

Wallenberg Initiative Materials Science for Sustainability 2023 – 2026

Environmental remediation of water and soil as well as clean-up of waste streams relies on separation of pollutants followed by their transformation to non-hazardous products. The vast majority of contaminated sites and waste streams are polluted with complex mixtures of chemicals, where the chemical complexity poses a challenge to developing highly efficient (nano)materials for remediation.
For efficient remediation we need to focus on all hazardous chemicals present in the contaminated environment. Recent developments in LC/HRMS have opened up possibilities for the detection of a broad spectrum of chemicals from contaminated sites and my group has recently expanded this methodology to evaluate for toxicity and concentration of detected chemicals with machine learning. These discoveries will allow us to combine the screening for chemical contaminants and material design that have remained independent processes until now. The goal of this project is to reach highly efficient materials for environmental remediation that perform under real world conditions by using the removal efficiency of toxic chemicals for the machine learning based inverse design.

Quantification of emerging contaminants with liquid chromatography high resolution mass spectrometry and graph-based machine learning

Anneli Kruve & Meelis Kull

VR 2022 – 2025

Understanding the toxicity of real-world complex mixtures, and identification of the most toxic chemicals is an essential starting point for the intelligent design of water treatment solutions and chemical regulation. To evaluate the risk possessed by water contaminants to humans and wildlife we need to know the structure, concentration (exposure), and toxic endpoint of the contaminants. Liquid chromatography-electrospray mass spectrometry (LC/ESI/HRMS) in combination with machine learning has offered significant improvements in detecting, identifying, and estimating the toxic endpoint for hundreds and thousands of water contaminants simultaneously; however, the quantification is lagging. The main reason is that analytical standards are required for the quantification of the detected contaminants.

In this project, we will develop machine learning methods to quantify the contaminants detected with LC/ESI/HRMS even if analytical standards are not available. Specifically, we will develop machine learning models to predict the response of the detected compounds in LC/ESI/HRMS from the structure of the compounds and use this for the quantification of these compounds. This will allow us to evaluate the risk possessed by the detected compounds, without time-consuming synthesis of the analytical standards. Moreover, it will become possible to pinpoint the compound(s) with the greatest contribution to the mixture toxicity.

Tandem development of waste textile recycling process and chemical screening for a non-toxic Re:Start

Aji Pallikunnel Mathew & Anneli Kruve

FORMAS dual project 2021 – 2023

Re:start aims to establish a tandem method combining sustainable processes and online chemical analysis to enhance the sustainability and safety related to recycling of waste textiles. The focus will be on reclaiming PET, acrylics, nanocellulose and its hybrids from textile waste for subsequent conversion to monomers, chemicals or new materials for recirculation. This will yield positive benefits as i) reduced need for virgin resources in industrial process ii) decreased materials that will be phased out and end up in incineration or landfill and iii) safe and toxic free products. However, the circulation should not come at the cost of quality, human health, or ecosystem. The project will therefore develop fast and non-invasive analysis methods to detect possibly harmful chemicals before and during recycling processes. Furthermore, our approach to partially fractionate the textiles to suitable hybrids without separating them completely into the component chemical fractions is expected to be more toxic free and resource efficient than chemical routes used today. New thermomechanical recycling processes will contribute significantly to green recycling and also put Sweden as forerunners of sustainable textile recycling and recovery of valuable resources.The developed tandem method is expectd to enable the emergence of a resource-efficient industries, chemical-smart society and new start ups focusing on toxic free and sustainable products from recycling.

Design of chemical reagents for fast and sensitive detection of pesticides in water and food

Anneli Kruve & Nicklas Selander, supported by Berit Olofsson, Meelis Kull, Riin Rebane

SUCCeSS 2022 – 2024

The sustainability goals put forward by the United Nations stress the need for clean and accessible drinking water and keeping the seas clean while still enabling industrial growth. The focus of this project is on the sustainable development of chemical reagents for the detection of glyphosate, the most widely used pesticide in the world, which affects the abundance of aquatic life as well as the quality of drinking water for human consumption. We will focus on developing high sensitivity, synthetically easily accessible, nonhazardous reagents for the determination of glyphosate. To achieve this,we will employ machine learning to generate and evaluate chemical reagents in-silico allow the algorithms to iteratively learn which structures provide the highest sensitivity while having high sustainability index. The machine learning algorithms will allow the evaluation of thousands of chemical reagents without producing chemical waste and the most promising reagents will be synthesized as well as their suitability in the practical analysis of glyphosate will be evaluated.

RapMixTox: Rapid and automated prediction of complex mixture toxicity by nontarget instrumental analysis

Anneli Kruve, Magnus Breitholtz, Jonathan Martin, Matthew MacLeod

FORMAS 2021 – 2023

The Swedish Government recently launched a comprehensive report on the future of chemical risk management accounting for combination effects and assessing chemicals in groups. The major conclusions from the report were that there is a need to improve the assessment and management of mixture risks and increase and develop group-wise assessment and management of chemicals. In more detail, modelling both co-exposures and combined toxicity was identified as key for regulatory progress.

Unfortunately, assessing the effect of the mixtures traditionally requires expensive and slow testing in vivo, and the constituents of the mixture need to be known. Real-world environmental water samples may contain hundreds if not thousands of chemical compounds, which vary strongly depending on the location and proximity to industrial and agricultural activities. Water samples are routinely analysed using liquid chromatography mass spectrometry, which allows separating the components and measuring several characteristics of these compounds. Still, identification of all of the constituents of an environmental water sample is a very challenging task, and all constituents of a sample do not contribute equally to the mixture effect.

In the current project, we will develop a method to predict the mixture toxicity of water samples without the need for identification of each constituent, and without biological testing. Specifically, we will use the information collected while analysing surface water sample with liquid chromatography mass spectrometry directly for predicting the toxic effect of the compounds. This will allow us to evaluate the mixture toxicity rapidly, without identifying the structure of all chemical compounds in the sample. Moreover, it will become possible to pinpoint the compound(s) with the greatest contribution to the mixture effect prior to the identification of the compounds and without the need for time-consuming and expensive animal testing.

Interlaboratory comparison of (semi-)quantitative LC/HRMS non-targeted screening

Anneli Kruve & Nikolaos S. Thomaidis

NORMAN 2020 – 2022

The identification and quantification of emerging contaminants from complicated organic mixtures have traditionally been based on the “calibration” standard substances. Non-targeted analysis has redefined the (tentative) identification of the compounds. Now several new methods for (semi-)quantification of the tentatively identified compounds are emerging both inside and outside NORMAN non-targeted community. However, the validity of the methods remains ambiguous as (1) they have been tested on limited type of samples and (2) commonly accepted datasets suitable for benchmarking of the
(semi-)quantification approaches are unavailable. A possibility to overcome and speed up the development of (semi-)quantification approaches is to provide a common ground for testing the developed approaches via a collaborative trial.