Graph-based machine learning interprets and predicts diagnostic isomer-selective ion-molecule reactions in tandem mass spectrometry

Recent Developments in Machine Learning for Mass Spectrometry

Statistical analysis and modeling of mass spectrometry (MS) data have a long and rich history with several modern MS-based applications using statistical and chemometric methods. Recently, machine learning (ML) has experienced a renaissance due to advents in computational hardware and the development of new algorithms for artificial neural networks (ANN) and deep learning architectures. Moreover, recent successes of new ANN and deep learning architectures in several areas of science, engineering, and society have further strengthened the ML field. Importantly, modern ML methods and architectures have enabled new approaches for tasks related to MS that are now widely adopted in several popular MS-based subdisciplines, such as mass spectrometry imaging and proteomics. Herein, we aim to provide an introductory summary of the practical aspects of ML methodology relevant to MS. Additionally, we seek to provide an up-to-date review of the most recent developments in ML integration with MS-based techniques while also providing critical insights into the future direction of the field.

Graph to Activation Energy Models Easily Reach Irreducible Errors but Show Limited Transferability

Activation energy characterization of competing reactions is a costly but crucial step for understanding the kinetic relevance of distinct reaction pathways, product yields, and myriad other properties of reacting systems. The standard methodology for activation energy characterization has historically been a transition state search using the highest level of theory that can be afforded. However, recently, several groups have popularized the idea of predicting activation energies directly based on nothing more than the reactant and product graphs, a sufficiently complex neural network, and a broad enough data set. Here, we have revisited this task using the recently developed Reaction Graph Depth 1 (RGD1) transition state data set and several newly developed graph attention architectures. All of these new architectures achieve similar state-of-the-art results of ∼4 kcal/mol mean absolute error on withheld testing sets of reactions but poor performance on external testing sets composed of reactions with differing mechanisms, reaction molecularity, or reactant size distribution. Limited transferability is also shown to be shared by other contemporary graph to activation energy architectures through a series of case studies. We conclude that an array of standard graph architectures can already achieve results comparable to the irreducible error of available reaction data sets but that out-of-distribution performance remains poor.

Structure Based Machine Learning Prediction of Retention Times for LC Method Development of Pharmaceuticals

PURPOSE

Significant resources are spent on developing robust liquid chromatography (LC) methods with optimum conditions for all project in the pipeline. Although, data-driven computer assisted modelling has been implemented to shorten the method development timelines, these modelling approaches require project-specific screening data to model retention time (RT) as function of method parameters. Sometimes method re-development is required, leading to additional investments and redundant laboratory work. Cheminformatics techniques have been successfully used to predict the RT of metabolites & other component mixtures for similar use cases. Here we will show that these techniques can be used to model structurally diverse molecules and predictions of these models trained on multiple LC conditions can be used for downstream data-driven modelling.

METHODS

The Molecular Operating Environment (MOE) was used to calculate over 800 descriptors using the strucutres of the analytes. These descriptors were used to model the RT of the analytes under four chromatographic conditions. These models were then used to create data-driven models using LC-SIM.

RESULTS

A structural-based Random Forest (RF) model outperformed other techniques in cross-validation studies and predicted the RTs of a randomized test set with a median percentage error less than 4% for all LC conditions. RTs predicted by this structure-based model were used to fit a data-driven model that identifies optimum LC conditions without any additional experimental work.

CONCLUSIONS

These results show that small training sets yield pharmaceutically relevant models when used in a combination of structure-based and data-driven model.

Learning Relationships Between Chemical and Physical Stability for Peptide Drug Development

PURPOSE OR OBJECTIVE

Chemical and physical stabilities are two key features considered in pharmaceutical development. Chemical stability is typically reported as a combination of potency and degradation product. Moreover, fluorescent reporter Thioflavin-T is commonly used to measure physical stability. Executing stability studies is a lengthy process and requires extensive resources. To reduce the resources and shorten the process for stability studies during the development of a drug product, we introduce a machine learning-based model for predicting the chemical stability over time using both formulation conditions as well as aggregation curves.

METHODS

In this work, we develop the relationships between the formulation, stability timepoint, and the chemical stability measurements and evaluated the performance on a random test set. We have developed a multilayer perceptron (MLP) for total degradation prediction and a random forest (RF) model for potency.

RESULTS

The coefficient of determination (R²) of 0.945 and a mean absolute error (MAE) of 0.421 were achieved on the test set when using MLP for total degradation. Similarly, we achieved a R² of 0.908 and MAE of 1.435 when predicting potency using the RF model. When physical stability measurements are included into the MLP model, the MAE of predicting TD decreases to 0.148. Using a similar strategy for potency prediction, the MAE decreases to 0.705 for the RF model.

CONCLUSIONS

We conclude two important points: first, chemical stability can be modeled using machine learning techniques and second there is a relationship between the physical stability of a peptide and its chemical stability.

Modification of a Quadrupole/Orbitrap/Linear Quadrupole Ion Trap Tribrid Mass Spectrometer for Diagnostic Gas-Phase Ion-Molecule Reactions

Tandem mass spectrometry based on diagnostic gas-phase ion-molecule reactions represents a robust method for functional group identification in unknown compounds. To date, most of these reactions have been studied using unit-resolution instruments, such as linear quadrupole ion traps and triple quadrupoles, which cannot be used to obtain elemental composition information for the species of interest. In this study, a high-resolution mass spectrometer, a quadrupole/orbitrap/linear quadrupole ion trap tribrid, was modified by installing a portable reagent inlet system to obtain high-resolution data for ion-molecule reactions. Examination of a previously published test system, the reaction between protonated 1,1'-sulfonyldiimizadole with 2-methoxypropene, demonstrated the ability to perform ion-molecule reactions on the modified tribrid mass spectrometer. High-resolution data were obtained for ion-molecule reactions of three isobaric ions (protonated glycylalanine, protonated glutamine, and protonated lysine) with diethylmethoxyborane. On the basis of these data, the isobaric ions can be differentiated based on both their measured accurate mass as well as the different product ions they generated upon the ion-molecule reactions. In a different experiment, analyte ions were subjected to collision-induced dissociation (CID), and the structures of the resulting fragment ions were examined via diagnostic ion-molecule reactions. This experiment allows for the functional group interrogation of fragment ions and can be used to improve the understanding of the structures of fragment ions generated in the gas phase.

Recent Advances in Mass Spectrometry-Based Structural Elucidation Techniques

Mass spectrometry (MS) has become the central technique that is extensively used for the analysis of molecular structures of unknown compounds in the gas phase. It manipulates the molecules by converting them into ions using various ionization sources. With high-resolution MS, accurate molecular weights (MW) of the intact molecular ions can be measured so that they can be assigned a molecular formula with high confidence. Furthermore, the application of tandem MS has enabled detailed structural characterization by breaking the intact molecular ions and protonated or deprotonated molecules into key fragment ions. This approach is not only used for the structural elucidation of small molecules (MW < 2000 Da), but also crucial biopolymers such as proteins and polypeptides; therefore, MS has been extensively used in multiomics studies for revealing the structures and functions of important biomolecules and their interactions with each other. The high sensitivity of MS has enabled the analysis of low-level analytes in complex matrices. It is also a versatile technique that can be coupled with separation techniques, including chromatography and ion mobility, and many other analytical instruments such as NMR. In this review, we aim to focus on the technical advances of MS-based structural elucidation methods over the past five years, and provide an overview of their applications in complex mixture analysis. We hope this review can be of interest for a wide range of audiences who may not have extensive experience in MS-based techniques.

Sodium adduct formation with graph-based machine learning can aid structural elucidation in non-targeted LC/ESI/HRMS

Non-targeted screening with LC/ESI/HRMS aims to identify the structure of the detected compounds using their retention time, exact mass, and fragmentation pattern. Challenges remain in differentiating between isomeric compounds. One untapped possibility to facilitate identification of isomers relies on different ionic species formed in electrospray. In positive ESI mode, both protonated molecules and adducts can be formed; however, not all isomeric structures form the same ionic species. The complicated mechanism of adduct formation has hindered the use of this molecular characteristic in the structural elucidation in non-targeted screening. Here, we have studied the adduct formation for 94 small molecules with ion mobility spectra and compared collision cross-sections of the respective ions. Based on the results we developed a fast support vector machine classifier with polynomial kernels for accurately predicting the sodium adduct formation in ESI/HRMS. The model is trained on five independent data sets from different laboratories and uses the graph-based connectivity of functional groups and PubChem fingerprints to predict the sodium adduct formation in ESI/HRMS. The validation of the model showed an accuracy of 74.7% (balanced accuracy 70.0%) on a dataset from an independent laboratory, which was not used in the training of the model. Lastly, we applied the classification algorithm to the SusDat database by NORMAN network to evaluate the proportion of isomeric compounds that could be distinguished based on predicted sodium adduct formation. It was observed that sodium adduct formation probability can provide additional selectivity for about one quarter of the exact masses and, therefore, shows practical utility for structural assignment in non-targeted screening.

Explainable graph neural networks for organic cages

The development of accurate and explicable machine learning models to predict the properties of topologically complex systems is a challenge in materials science. Porous organic cages, a class of polycyclic molecular materials, have potential application in molecular separations, catalysis and encapsulation. For most applications of porous organic cages, having a permanent internal cavity in the absence of solvent, a property termed “shape persistence” is critical. Here, we report the development of Graph Neural Networks (GNNs) to predict the shape persistence of organic cages. Graph neural networks are a class of neural networks where the data, in our case that of organic cages, are represented by graphs. The performance of the GNN models was measured against a previously reported computational database of organic cages formed through a range of [4 + 6] reactions with a variety of reaction chemistries. The reported GNNs have an improved prediction accuracy and transferability compared to random forest predictions. Apart from the improvement in predictive power, we explored the explicability of the GNNs by computing the integrated gradient of the GNN input. The contribution of monomers and molecular fragments to the shape persistence of the organic cages could be quantitatively evaluated with integrated gradients. With the added explicability of the GNNs, it was possible not only to accurately predict the property of organic materials, but also to interpret the predictions of the deep learning models and provide structural insights for the discovery of future materials.

We report the development of explainable Graph Neural Networks to predict shape persistence of organic cages. Integrated gradient analysis identifies collapse-inducing molecular fragments and helps chemists design more shape persistent structures.

Determination of the compound class and functional groups in protonated analytes via diagnostic gas-phase ion-molecule reactions

Diagnostic gas-phase ion-molecule reactions serve as a powerful alternative to collision-activated dissociation for the structural elucidation of analytes when using tandem mass spectrometry. The use of such ion-molecule reactions has been demonstrated to provide a robust tool for the identification of specific functional groups in unknown ionized analytes, differentiation of isomeric ions, and classification of unknown ions into different compound classes. During the past several years, considerable efforts have been dedicated to exploring various reagents and reagent inlet systems for functional-group selective ion-molecule reactions with protonated analytes. This review provides a comprehensive coverage of literature since 2006 on general and predictable functional-group selective ion-molecule reactions of protonated analytes, including simple monofunctional and complex polyfunctional analytes, whose mechanisms have been explored computationally. Detection limits for experiments involving high-performance liquid chromatography coupled with tandem mass spectrometry based on ion-molecule reactions and the application of machine learning to predict diagnostic ion-molecule reactions are also discussed.

Characterization of Protonated Substituted Ureas by Using Diagnostic Gas-Phase Ion-Molecule Reactions Followed by Collision-Activated Dissociation in Tandem Mass Spectrometry Experiments

Substituted ureas correspond to a class of organic compounds commonly used in agricultural and chemical fields. However, distinguishing between different ureas and differentiating substituted ureas from other compounds with similar structures, such as amides, N-oxides, and carbamates, are challenging. In this paper, a four-stage tandem mass spectrometry method (MS⁴) is introduced for this purpose. This method is based on gas-phase ion-molecule reactions of isolated, protonated analytes ([M + H]⁺) with tris(dimethylamino)borane (TDMAB) (MS²) followed by subjecting a diagnostic product ion to two steps of collision-activated dissociation (CAD) (MS³ and MS⁴). All the analyte ions reacted with TDMAB to form a product ion [M + H + TDMAB - HN(CH₃)₂]⁺. The product ion formed for substituted ureas and amides eliminated another HN(CH₃)₂ molecule upon CAD to generate a fragment ion [M + H + TDMAB - 2HN(CH₃)₂]⁺, which was not observed for many other analytes, such as N-oxides, sulfoxides, and pyridines (studied previously). When the [M + H + TDMAB - 2HN(CH₃)₂]⁺ fragment ion was subjected to CAD, different fragment ions were generated for ureas, amides, and carbamates. Fragment ions diagnostic for the ureas were formed via elimination of R-N═C═O (R = hydrogen atom or a substituent), which enabled the differentiation of ureas from amides and carbamates. Furthermore, these fragment ions can be utilized to classify differently substituted ureas. Quantum chemical calculations were employed to explore the mechanisms of the reactions. The limit of detection for the diagnostic ion-molecule reaction product ion in HPLC/MS² experiments was found to range from 20 to 100 nM.