We plan to conduct secondary analysis on GREI datasets with the primary aim of elucidating molecular conformations and stereochemistry. We will utilize the Geometric Ensemble of Molecules (GEOM) dataset, containing over 37 million annotated molecular conformations. The analysis will be supplemented by two additional publicly available datasets: (a) the Natural Products Magnetic Resonance Database, containing annotated 1D ¹H and ¹³C NMR spectra for natural products; and (b) the GlycoNMR dataset, with 1D NMR spectra for carbohydrates. We will develop machine learning models capable of (a) predicting 1D/2D NMR shifts for molecules; and (b) elucidating molecular structures, particularly stereochemistry and conformational features, from experimental NMR data. This project will enhance our scientific understanding of molecular structures, improve the interpretation of complex compounds, and accelerate drug discovery, advancing disease research and precision in therapeutic interventions.
Nuclear Magnetic Resonance (NMR) spectroscopy is the most widely used and arguably the most powerful tool for elucidating the structure and stereochemistry of molecules, especially natural products and complex carbohydrates (glycans). NMR spectra encode the local environment information of the atoms that make up a molecule, providing molecular 'fingerprints' that can be used to interpreted conformational properties and relative stereochemistry. However, interpreting NMR spectra and extracting critical stereochemical and conformational details from NMR spectra are time consuming and require substantial expertise. Advanced machine learning techniques offer significant potential to improve this process.
We propose to conduct a secondary data analysis by integrating three datasets: the GEOM dataset from the Generalist Repository Ecosystem Initiative (GREI), the Natural Products Magnetic Resonance Database (NP-MRD: : https://np-mrd.org/), and the GlycoNMR dataset (https://github.com/Cyrus9721/GlycoNMR). The GEOM dataset, available through Dataverse, contains over 37 million molecular conformations annotated by energy and statistical weight for more than 450,000 molecules. This dataset will complement the NP-MRD dataset, which contains ~18,000 annotated experimental 1D NMR spectra and ~780 annotated experimental HSQC (one type of 2D NMR) spectra of natural products, and the GlycoNMR dataset, which contains ~300 annotated 1D NMR spectra of carbohydrates.
Our prior works have demonstrated that deep learning techniques can be effectively used to predict 1D NMR shifts (https://data.mlr.press/assets/pdf/v01-11.pdf) and 2D NMR cross-peaks (https://arxiv.org/abs/2403.11353v2). In this proposal, by combining the GEOM dataset with the two publicly available NMR datasets, we aim to develop interpretable models that not only predict both 1D and 2D NMR shifts with high accuracy but also provide insight into how specific spectral features correspond to structural characteristics, particularly stereochemistry. The GEOM dataset will be pivotal in linking molecular conformations with NMR spectral features. We will train our model to incorporate all conformations of each molecule into predicting its NMR shifts, allowing our model to more effectively capture the impact of different molecular conformations on NMR spectra. Model performance will be evaluated using mean squared error (MSE) for NMR shift prediction and structural correctness for molecular elucidation.
Timeline:
We anticipate several key outcomes from this project. First, we aim to develop a robust machine learning model capable of accurately predicting 1D and 2D NMR chemical shifts for both natural products and glycans. Additionally, we will create a tool for molecular structure elucidation, providing key stereochemical and conformational information directly from NMR spectra. These findings will provide insights into the relationship between NMR spectral features and molecular structures, which could be generalized to other classes of molecules beyond glycans. Another significant outcome will be the development of a web service that allows users to upload NMR spectra for automated analysis, providing them with predicted molecular candidates and NMR shift assignments.
We will publish our results in peer-reviewed journals and present at relevant conferences. Additionally, we will make the code, models, and data available through open-access platforms such as GitHub. Data management plan will adhere to FAIR (Findable, Accessible, Interoperable, Reusable) Guiding Principles, which emphasize rich metadata and machine-actionability. Specifically, all datasets and models will be assigned unique, persistent identifiers in accessible repositories, with detailed metadata ensuring they are easily searchable. The data and models will be freely available and easily retrievable, with unrestricted access. We will use standardized and widely accepted formats and vocabularies to ensure that the data is interoperable and usable by a broad range of researchers. The proposed web service will be released under an open license, accompanied by comprehensive documentation to ensure clarity and ease of use.
To ensure the replicability and reproducibility, we will provide thorough documentation of all preprocessing steps, model development, and validation procedures, enabling other researchers to replicate our work and apply the methods to their own datasets.
We anticipate that the development of proposed method and tools will significantly enhance the ability to interpret NMR spectra for complex molecules, reducing reliance on traditional time-consuming experimental methods. Natural products are essential to the normal growth, development and reproduction of their hosts and provide physiological benefits. They are of considerable interest to a variety of industries (such as, pharmaceuticals, agriculture, food and beverages, environmental and biotechnology, nutraceuticals and dietary supplements, cosmetics and personal care, and so on). Natural products are structurally complex, where accurate stereochemical analysis is key to understanding their therapeutic potential. Small changes in stereochemistry can drastically affect biological activity, making precise analysis essential. Similarly, glycans play vital roles in processes like immune response and pathogen recognition. Their conformational flexibility and branching structures make them challenging to interpret. A key focus on glycans is particularly impactful for fields such as glycomics, where understanding carbohydrate structures is essential for advancing knowledge of biological function and disease mechanisms. Improved interpretation of glycan structures will deepen the understanding on molecular basis of many diseases, including those linked to immune response and pathogen interactions.
By incorporating cutting-edge machine learning and advanced data analysis techniques, this project aims to push the boundaries of NMR-based structural elucidation, particularly for complex molecules such as natural products and glycans. The proposed work will have wide-ranging implications for both fundamental research and practical applications in fields such as medicinal chemistry and structural biology.
Our team consists of three researchers with expertise in NMR spectroscopy, machine learning, and spectra analysis. We came together through a shared interest in applying advanced data analysis and machine learning techniques to molecular structure elucidation. As a collaborative team, we meet regularly to review progress, exchange ideas, and troubleshoot challenges.
Pengyu specializes in NMR spectroscopy and molecular analysis, bringing a deep understanding of the intricacies of NMR data and how it relates to molecular structures.
Yunrui has extensive experience in developing and applying machine learning models, particularly in graph neural networks (GNNs) and message passing neural networks (MPNNs). She will focus on designing the predictive models that will process the structural data and NMR spectra.
Hongrui will bring expertise in statistical analysis and data preprocessing. He will be responsible for transforming the raw data from multiple sources into machine-learning-compatible formats and applying advanced statistical methods to ensure data consistency and integrity. He will also lead the development of the project’s web service, which will allow users to upload 1D and 2D NMR spectra for automated analysis, providing predicted molecular candidates and atom-to-NMR shift assignments.
The success of the proposed research project will be ensured by several critical factors. First, the use of three high-quality, well-annotated datasets will provide a robust foundation for developing predictive models. Both datasets are comprehensive and compliant with FAIR principles. We will also incorporate several smaller NMR datasets available on Dataverse that are not included in the NP-MRD and GlycoNMR collections, further enriching our dataset diversity. Secondly, the expertise of the research team, which spans the domains of machine learning, NMR spectroscopy, and structural biology, will be essential in the interpretation and modeling of complex NMR spectra. Cutting-edge machine learning techniques will be employed to deliver state-of-the-art predictive capabilities. By emphasizing model interpretability and reproducibility, we will enhance the reliability and applicability of the results. Together, these considerations will ensure the success of the proposed project.
