Harnessing AI-ML for Chemical Innovation, Process Optimization, and Sustainable Impact

Biodegradation is an important depletion pathway for mitigating environmental and human health risks associated with chemical exposure, bioaccumulation and toxicity. Compounds that are persistent in the environment can accumulate in ecosystems, human membranes and tissues, which can limit efficacy and increase the risk of adverse effects. Due to the large number of chemicals produced worldwide, experimental biodegradation studies are impractical. Traditional in silico models often underperform due to the lack of mechanistic insight, as many rely on structural (vs. chemical) information derived from 2-D (vs. 3-D) data. To address these challenges, we developed a series of decision-tree models for binary classification of readily biodegradable (RB) versus non-readily biodegradable (non-RB) compounds, with molecular weights up to 1,000 g/mol. These models encompass the vast majority of industrial chemicals as well as small-molecule Active Pharmaceutical Ingredients (APIs). Using a curated and structurally diverse chemical library, we first developed a model that uses pharmacophore-based matching to predict biodegradation. This approach was adapted from drug discovery, where compounds likely to bind enzymatic targets share similar distributions of features and shape. To account for uptake into living systems and oxidation (and reduction) reactions post binding, we developed models based on physicochemical properties related to bioavailability and QM-based descriptors that gauge the molecular response to an electron-density flux, respectively. Combined, these models offer a comprehensive and mechanistically defensible in silico approach to evaluate biodegradability of small but chemically complex molecules. Our protocol allows for fast screening of large libraries, identifying potentially persistent chemicals faster, and reducing the demand for experimental synthesis and tests. The underlying structure-property relationships (SPRs) lend themselves uniquely to new and safer chemical design, where biodegradation can be tuned and optimized along other vectors related to chemical hazard and performance.

Conventional synthesis of the key Edoxaban intermediate, tert-butyl (1R,2S,5S)-2-amino-5-(dimethylcarbamoyl)cyclohexylcarbamate, relies on hazardous azide chemistry and multi-step protecting-group manipulations, posing safety concerns and environmental drawbacks. To enable a greener and more sustainable manufacturing route, we attempted a biocatalytic alternative using transaminases. The significant steric hindrance of the prochiral ketone substrate, however, presented a major challenge. To address this, we developed an AI-assisted, structure-guided workflow to mine public databases, leading to the discovery of transaminase PaTA as a viable starting point. Systematic engineering of its active site produced a triple-mutant variant with a remarkable >400-fold increase in activity. This engineered biocatalyst achieves >90% conversion with excellent selectivity within 12 hours under mild, aqueous conditions and low catalyst loading. Our work establishes a safe, efficient, and intrinsically green biocatalytic process that eliminates the use of explosive reagents, reduces synthetic steps, and offers a sustainable alternative to traditional chemical synthesis.

Deep eutectic solvents (DES) are emerging as green alternatives to the traditional organic solvents used in chemical processing including lignocellulosic biorefining. Despite their potential, it is highly challenging to rapidly design and discover novel effective DES due to the requirement of complex experimentation to generate comprehensive solvation properties. We have overcome this challenge by simultaneous prediction of three solvation parameters of DESs using machine learning models to analyze experimental data and quantum chemistry calculations of 522 compounds. The model has achieved high accuracy and guided the discovery of new DES (TBACl-LA) for highly efficient delignification and tailored lignin structural design for downstream usage. The efficient lignin removal and the avoidance of inhibitors have empowered superior hydrolysate processability comparable to pure glucose. Furthermore, the lignin has been fractionated to a hydrophobic macromolecule for quality biomaterials, where we demonstrated the polymer-network for slow-release fertilizer with significant improvement in plant growth. Techno-economic analysis (TEA) indicates that the novel DES-based process enables economically viable biorefinery. Together, artificial intelligence has guided the design of new DES that improves delignification, substantially increases hydrolysate processability and biodiesel production, and empowers quality lignin biomaterials, all of which together could revolutionize the next generation biorefinery with better economics and environmental impact.

Fluorescent Brightening Agents (FBAs) are synthetic organic compounds designed to enhance the perceived whiteness and brightness of materials. They act by absorbing light in the near ultraviolet region and re-emitting visible blue-violet fluorescence, counteracting the yellowish cast present in many materials. FBAs are used extensively in textiles, paper, plastics, and detergents. Despite decades of use, toxicological and environmental fate data remain limited for many commercial variants, particularly across the structural classes that dominate production. This study focuses on the FBA classes that represent most commercial production: stilbene derivatives, coumarins, benzoxazoles, pyrazolines, and naphthalimides. Using the GreenScreen for Safer Chemicals and the EPA Design for the Environment (DfE) alternatives assessment framework, each class is evaluated against established hazard benchmarks, including persistence, bioaccumulation, aquatic toxicity and mammalian toxicity endpoints. Where experimental data are lacking, predictive modeling (VEGA, TEST, ECOSAR) is used to fill gaps. By combining experimental and computational data, this study informs FBAs that may warrant further investigation or substitution. Structural features, such as aromatic rings, water-soluble groups and stable linkers, which can be associated with elevated hazard or persistence, are identified. Based on the developed structure activity relationships (SARs), the study provides a framework to guide redesign of whitening agents with improved safety profiles.

As the backbone of innovation, chemistry determines not only the functionality of products but also their usability, market viability, and quality.
But with millions of chemical compounds supplied by thousands of manufacturers, understanding the toxicity effects of these chemicals and their combined effect is very difficult.

If scientists can get trusted information about the toxicity of raw material and can select the least harmful alternatives that still deliver the performance needed, they can create safer products while minimizing hazardous waste. And companies can bring high-quality products to market much faster while reducing the R&D cost by more than 25%. ViridisChem is dedicated to delivering comprehensive toxicity insights on chemicals, reactions, complex mixtures and formulations—empowering researchers to design better, greener solutions.

This presentation will describe how ViridisChem has utilized the latest AI technologies to build a unique chemically intelligent platform containing meticulously curated toxicity databases and advanced analytical tools; and how its Gen-AI seamlessly integrates the power of these tools to provide highly collaborative and chemically intelligent solution for advanced insights. We will also discuss our experience in how best we could take advantage of the AI, and the pitfalls we avoided to minimize the hallucinations issues, while giving the ultimate final decision-making control to the scientists.

The selection of more sustainable, greener, and “better” chemicals for green chemistry solutions has been the focus of sustainable and green chemistry practitioners for many years. For industries like Pharmaceuticals and Agrochemicals, where complex synthetic routes are routinely developed for a target molecule, identification of intermediates, reagents, solvents, reactants, etc. with known environmental, safety and health hazards has been key in evaluating competing routes to a commercial manufacturing route. This talk will discuss ViridisChem’s Reaction Analyzer, an AI-enabled tool that allows chemists to rapidly assess the EHS hazards of multiple routes.

In material recovery and recycling systems, conventional computer vision datasets and RGB-based approaches often fail to capture chemical heterogeneity and surface contamination that directly affect downstream process performance. This work presents an integrated framework that combines computer vision with hyperspectral imaging (HSI) to enable non-destructive, real-time characterization of complex material streams. Hyperspectral data provide chemically relevant spectral signatures linked to material composition, coatings, and contamination, complementing spatial features extracted from RGB imagery. In parallel, a domain-specific dataset was developed to reflect realistic operational conditions, including mixed materials, variable degradation states, and conveyor-based environments.
Models trained on this dataset were systematically evaluated against commonly used benchmark datasets, demonstrating that chemically informed sensing and dataset design significantly improve classification robustness and process-relevant decision accuracy. By explicitly linking spectral chemistry, data quality, and engineering performance, this work highlights how advanced imaging and dataset-centered design can reduce misclassification-driven waste, improve material circularity, and support scalable green chemistry–enabled infrastructure.

The choice of molecular representation (e.g., SMILES, graphs, fingerprints) affects both the scope and accuracy of AI models. For safer chemical design, what molecular representation is selected will determine the specific functions and hazards that may be controlled. Diverging from the use of structure- or physicochemical-based information (e.g., functional groups, partitioning properties), we present a representation based on nuclear magnetic resonance (NMR) spectra and show how it may be used to predict and control molecular reactivity. By carefully selecting information from NMR spectra, we show that it is possible to preserve physical meaning relating to the electronic environments within a molecule and achieve meaningful accuracy when used with AI models. Using model interpretability methods, we illustrate rational and intuitive linkages between NMR spectra and chemical reactivity.