Skip to main content
NCBI home page
ACS AuthorChoice logoLink to ACS AuthorChoice
. 2026 Sep 28;98(40):29273–29284. doi: 10.1021/acs.analchem.6c03711

CCSBase2: A Unified Benchmark Dataset and Machine Learning Framework for Robust CCS Prediction

Amogh Bantwal 1, Reuben Santoso 1, Ryan Nguyen 1, Griffin Rangel 1, Libin Xu 1,*
PMCID: PMC13670893  PMID: 42916131

Abstract

The identification of unknown analytes remains a persistent challenge in untargeted mass spectrometry-based analyses. Ion mobility–mass spectrometry (IM-MS) has expanded identification capabilities by enabling the use of collision cross section (CCS) as an orthogonal molecular identifier, but experimentally measured CCS values remain sparse relative to the breadth of known chemical space. Machine learning (ML) approaches have emerged as a complementary strategy for CCS prediction. Many existing models, however, suffer from poor generalizability and accuracy across structurally diverse chemical classes. Here, we present CCSBase2, a large, unified CCS database comprising 66,153 measurements assembled by integrating multiple public datasets, along with a classical ML framework for CCS prediction using Morgan count fingerprints. CCSBase2 achieved a mean relative error of 1.73%, median relative error of 1.23%, and root mean squared error of 4.89 Å2 on a held-out test set. CCSBase2 was trained on a consumer-grade Apple Silicon M4 Pro CPU, demonstrating that accurate CCS prediction can be achieved with minimal computational resources while maintaining fair performance when tested against out-of-distribution compounds. The final model is available at https://ccsbase.net/ccsbase2, and the codebase and datasets are freely available on GitHub (https://github.com/libinxulab/ccsbase2)­.

graphic file with name ac6c03711_0006.webp

graphic file with name ac6c03711_0005.webp

Introduction

Ion mobility–mass spectrometry (IM-MS) has emerged as a powerful analytical technique for characterizing and identifying unknown analytes. − While mass spectrometry provides mass-to-charge (m/z) information, many analytes, particularly structural isomers, cannot be uniquely identified based on mass alone. , Collision cross section (CCS) measurements from IM-MS provide complementary structural information by reflecting the gas-phase size and shape of an ion. , When combined with m/z and MS/MS fragmentation data, CCS can improve confidence in compound identification and enable more reliable discrimination between isomeric species. As a result, CCS has become an increasingly valuable parameter in metabolomics, lipidomics, proteomics, and environmental analysis. −

Three widely used ion mobility platforms for CCS measurement include drift tube ion mobility spectrometry (DTIMS), trapped ion mobility spectrometry (TIMS), and traveling wave ion mobility spectrometry (TWIMS). , Although all three techniques determine CCS by measuring ion mobility through a buffer gas, they differ in their operational principles and experimental conditions. DTIMS directly determines CCS by measuring the time required for ions to traverse a uniform electric field using the stepped-field method. ,, In contrast, TIMS and TWIMS measure CCS through ion trapping or propagation within dynamic electric fields in combination with CCS calibrants. Despite these methodological differences, reported inter-platform CCS deviations are typically on the order of 1–2%. ,

A key limitation of relying solely on experimentally measured CCS values for analyte identification is the limited coverage of public CCS reference databases. To address these constraints, machine learning (ML) approaches have increasingly been explored as an alternative strategy to predict CCS values for unknown analytes. − Prior to 2023, the largest publicly available CCS resource was CCSBase, comprising 16,840 unique compound-adduct pairs. More recently, METLIN-CCS has expanded this landscape, now reporting 53,893 unique compound-adduct pairs, representing the largest CCS database available to date. However, this database is limited to ions with <650 m/z values. Despite these recent advances, the large but heterogeneous CCS datasets also pose challenges for benchmarking predictive models. CCS prediction methods are often evaluated using disparate training and test splits, making direct comparison between models challenging. The lack of a standardized, community-wide evaluation test dataset limits reproducibility and objective assessment of model performance across studies.

In this work, we introduce CCSBase2, an ML framework and unified evaluation dataset for CCS prediction, achieving strong accuracy given limited computational power. We demonstrate that an XGBoost ML approach can be effectively applied to CCS prediction while maintaining generalization. Additionally, our unified dataset can serve as a final test dataset for future research. CCSBase2 outperforms most state-of-the-art deep learning architectures for a fraction of the cost and compute time. To enable fair and reproducible model comparisons, all models CCSBase2 was benchmarked against were retrained on the same training data and assessed on the same held-out test data.

Experimental Section

Database Aggregation

A multi-stage database aggregation workflow was developed, consisting of SMILES computation (Figure A), deduplication (Figure B), and chemical subclass classification (Figure C).

1.

1

Open in a new tab

Simplified overview of the CCSBase2 aggregation workflow.

External CCS datasets from METLIN-CCS, AllCCS, , PNNL, and Song et al. were combined with CCSBase and subjected to a two-step data cleaning workflow. Conditional filtering steps adapted from prior work were implemented to: (i) exclude dimer species, (ii) standardize adduct notation, (iii) reject entries lacking valid SMILES representations, and (iv) exclude entries flagged as low-reliability by the source dataset's own CCS quality metric. Dimers were identified using annotations supplied by the source datasets. For datasets containing adduct annotations, entries with a multiplicative coefficient greater than one on the neutral molecule were classified as dimeric. For METLIN-CCS, the dataset’s explicit monomer/dimer classification was used directly. Dimers were excluded because the SMILES representation used for molecular featurization does not explicitly encode noncovalent intermolecular interactions. Because dimer formation is common in electrospray ionization, reliable application of CCSBase2 to experimental data requires that putative dimeric features be identified prior to CCS prediction, as the current model was not trained to predict CCS values for these species. All adducts were reformatted using a common notation that includes square brackets and an explicit charge designation (e.g., [M + 2Na–3H]−). Dataset-specific CCS reliability metrics were also leveraged to further improve data quality. METLIN-CCS reports CCS precision using percent coefficient of variation (%CV) across three replicate measurements, while AllCCS assigns confidence levels ranging from 1 to 4. Only CCS values with %CV less than 1% (METLIN-CCS) or confidence level 1 (AllCCS), indicating high reproducibility across independent measurements from more than one researcher, were retained.

SMILES Computation

PubChem is a widely used chemical information resource, and we operated under the assumption that external CCS datasets relied on the PubChem API to generate SMILES representations. However, the PubChem API will, at times, return multiple candidate SMILES that require manual disambiguation, or produce salted SMILES whose molecular masses are inflated relative to experimentally measured values. For instance, querying thioridazine hydrocholoride returns the SMILES string CN1CCCCC1CCN2C3=CC=CC=C3SC4=C2C=C­(C=C4)­SC·Cl, corresponding to a molecular weight of 407.0 Da. The experimentally recorded mass of this analyte reported by Hines et al. is 371.161 Da. Following desalting, the corresponding neutral SMILES yields a molecular mass of 370.1537, with the remaining 1.0073 Da difference attributable to the proton mass of the [M + H]+ adduct that is not explicitly encoded in the SMILES representation. In other cases, PubChem returns multiple structural isomers for a given name. Querying 2-(phenylmethylene)­octanal, for example, yields three possible isomeric SMILES: CCCCCC/C­(=C\\C1=CC=CC=C1)/C=O,CCCCCCC­(=CC1=CC=CC=C1)­C=O, and CCCCCC/C­(=C/C1=CC=CC=C1)/C=O, all sharing the same molecular weight but differing in double-bond geometry.

Since CCSBase2 relies on SMILES-derived molecular features for model training, all structural representations were independently retrieved and standardized. Structures were initially retrieved from PubChem as isomeric SMILES to preserve available stereochemical information. Following salt removal using RDKit’s SaltRemover, each structure was regenerated using the MolToSmiles function, yielding a canonicalized SMILES representation while retaining stereochemistry. This procedure was selected because stereoisomers with identical canonical SMILES can exhibit different CCS values; for instance, fumaric acid (118.1 Å2) and maleic acid (113.1 Å2).

Validation workflow scripts first retrieved SMILES using either a PubChem compound identifier (CID) or a chemical name. Entries originating from METLIN-CCS, PNNL, and Song et al. include PubChem CIDs, whereas AllCCS and CCSBase provide only chemical names. In cases where PubChem failed to return a SMILES, often due to the use of lipid shorthand notation, a fallback strategy was implemented using the Lipid Maps API. For example, the lipid PE­(18:0/18:0) is recognized by Lipid Maps but not PubChem. All returned SMILES were subsequently desalted, and their neutral molecular masses were validated against experimentally reported masses. Exact molecular masses were computed using RDKit’s CalcExactMolWt function. An appropriate adduct mass offset was then applied to calculate a theoretical ion mass, which was compared against the experimental mass. Compounds were not assigned a SMILES if their experimental mass deviated by more than 1% from the theoretical value. When multiple candidate SMILES were returned from the PubChem API, each was independently validated using the same desalting and mass-matching criteria; however, no SMILES string was assigned if more than one candidate satisfied the mass validation thresholds.

Outlier Filtering and Deduplication

Outlier filtering was designed to address discrepancies among repeated SMILES-adduct observations. Remaining entries were grouped by SMILES–adduct pair, and the median CCS was calculated for each group. Measurements deviating by more than 1% from the group median were removed (Figure S1). The choice of a 1% threshold is based on our objective of establishing a dataset that can be used for high-accuracy CCS prediction in this and future studies. Given the large and diverse nature of our dataset, removing a small number of entries with relatively larger uncertainty to achieve higher-quality data is justifiable for training a more accurate predictive model. All remaining independent measurements were retained individually, thereby preserving the experimentally observed CCS variability available to the model. Separately, exact duplicate rows reflecting the same underlying measurement reported more than once across source datasets were dropped, retaining only a single copy of each. To characterize the measurement variability present in the aggregated data prior to outlier filtering, we examined the distribution of CCS deviations across all repeated compound-adduct measurements and evaluated whether high-variance entries were specific to structural classes (Figure S2).

Chemical Classification

A key distinction between CCSBase2 and the original CCSBase preprocessing workflow is the use of refined classification labels, rather than broad (e.g., lipid, peptide, carbohydrate, small molecule) categorical groupings. Chemical classification was performed using the ClassyFire API. For compounds with validated SMILES, the corresponding InChIKey was retrieved via the PubChem API and supplied as input to ClassyFire. This process returned a hierarchical classification consisting of superclass, class, and subclass, which were recorded in the CCSBase2 database. While ClassyFire provided reliable superclass annotations, class and subclass-level assignments were often lacking. In these cases, ClassyFire returned a “Not Found” exception, indicating that no classification could be assigned at these hierarchical levels. For example, 1-hydroxypyrene was classified within the benzenoid superclass and pyrene class, but no subclass annotation. In contrast, decabromodiphenyl ether returned no superclass, class, or subclass assignment from ClassyFire.

Compounds lacking subclass annotations represent a substantial source of untapped data that could otherwise contribute to downstream model training and analysis. To address this gap, we leveraged compounds with known subclass assignments to train a supervised XGBoost classifier (XGBClassifier). A training set was created by grouping unique SMILES by subclass and excluding groups with fewer than 30 compounds to ensure there was sufficient representation for each type of subclass. Excluded subclasses were withheld entirely from model training and reserved as a proxy out-of-distribution (OOD) evaluation dataset.

Hyperparameters for the XGBoost subclass classifier were optimized using scikit-learn’s GridSearchCV with 5-fold cross-validation. The search space consisted of n_estimators = [500, 600, 700], max_depth = [8, 9, 10], learning_rate = [0.05, 0.07, 0.08], subsample = [0.9, 1], colsample_bytree = [0.9, 1], reg_lambda = [5, 10], and gamma = [0, 1]. Average accuracy, precision, recall, and F1 score were computed across folds, with mean log loss used for model selection. The optimal configuration consisted of n_estimators = 600, max_depth = 8, learning_rate = 0.08, subsample = 1, colsample_bytree = 1, reg_lambda = 5, and gamma = 0.

Molecular structures were encoded using un-hashed Morgan count fingerprints (radius 2, chirality enabled). To enable an aligned feature matrix for model training, a common fingerprint vocabulary was constructed from the unique substructure identifiers observed across the dataset. Each identifier was assigned a fixed feature column, and each molecule was vectorized according to the number of occurrences of each vocabulary substructure. To ensure accurate predictions at inference time, we implemented a gate to prevent the classifier from making subclass predictions on chemicals that are chemically different from the training set. For each compound requiring prediction, we computed its Morgan count fingerprint and measured its Tanimoto similarity against every compound in the training set, retaining the maximum similarity value as a nearest-neighbor similarity score. Compounds whose nearest-neighbor Tanimoto similarity fell below an empirically chosen threshold were considered structurally dissimilar from any compound the classifier had been trained on and were not assigned a label. A Tanimoto similarity threshold was determined by sampling compounds from each subclass within the training set and computing pairwise Tanimoto nearest-neighbor similarities among them, establishing a baseline for the median similarity expected among structurally related, in-distribution compounds.

To evaluate the classifier’s ability to correctly abstain from making uncertain predictions, we constructed a proxy out-of-distribution (OOD) evaluation set composed of subclasses excluded from classifier training. Because these subclass labels were never represented during training, the classifier could not correctly assign them to their true subclass. We therefore evaluated whether the Tanimoto similarity gate prevented predictions for compounds that were structurally dissimilar to the training set. OOD rejection behavior and subclass prediction coverage are summarized in Table S1. In-distribution classifier performance is reported in Table S2.

CCSBase2 Model Development

CCS prediction was performed using an XGBoost regressor (XGBRegressor) implemented in Python 3.12. Hyperparameters were optimized using GridSearchCV with 5-fold cross-validation and a random seed of 26. The search space consisted of n_estimators = [6000, 7000, 8000, 10000, 12000, 14000, 20000], max_depth = [8, 10, 12, 13, 15], learning_rate = [0.01, 0.02, 0.03], subsample = [0.9], colsample_bytree = [0.5, 0.9], reg_lambda = [30], min_child_weight = [1], and gamma = [1]. Mean absolute error (MAE), median absolute error (MDAE), root mean squared error (RMSE), mean relative error (MRE), median relative error (MDRE), and R 2 were calculated across folds, with RMSE used for model selection. The optimal configuration consisted of n_estimators = 2000, max_depth = 12, learning_rate = 0.02, subsample = 0.9, colsample_bytree = 0.9, reg_lambda = 30, min_child_weight = 1, and gamma = 1. CCSBase2 was trained using the same Morgan count fingerprints vocabulary as our subclass classifier, in addition to one-hot encoded (OHE) adduct identity, signed adduct mass, and molecular mass. Our OHE representation had a dimension size of 14, representing the number of adducts in our database with more than 50 examples (n = 13) plus an additional dimension for all adducts with fewer than 50 examples. Molecular mass and Morgan count fingerprints were computed using RDKit (CalcExactMolWt and GetMorganGenerator) methods, and signed adduct mass was derived by subtracting the exact molecular mass from the experimental ion mass.

Relative to the original CCSBase model, three key modifications were introduced to CCSBase2: (i) adoption of a new ML algorithm: XGBoostRegressor, (ii) implementation of a revised train-test splitting strategy, and (iii) inclusion of an additional adduct-specific descriptor. CCSBase’s train-test split method uses scikit-learn’s StratifiedShuffleSplit method to partition a training (80%) and test (20%) set by binning six discrete intervals defined by the dataset's quartile distribution (Q1, Q2, Q3) and intermediate midpoints (m 1,2,m 2,3), ensuring training and test sets maintained identical, representative distributions across the entire CCS range. However, a limitation of this splitting method is that it can cause the training and test sets to have different chemical and adduct distributions, which may prevent the model from learning, for certain compounds, how CCS varies across different adducts. Hence, CCSBase2’s approach is to first group by subclass, and within each subclass, all CCS measurements associated with a given SMILES string were treated as a single indivisible unit. These compound-level groups, rather than individual measurements, were then strategically assigned to a test set while the remaining were assigned to a training set.

Test-set assignment was performed using a greedy selection procedure. Candidate groups were iteratively added to the test set, at each step selecting the group that minimized the deviation between the test set's cumulative adduct distribution and the adduct distribution of the subclass overall, until approximately 20% of the subclass's measurements were allocated to the test set. This approach was designed to preserve, within the training set, multiple adduct-specific measurements for the same compound, as this comparison is expected to provide the model with information relevant to learning adduct-dependent effects on CCS.

However, because subclass-level grouping requires a subclass assignment, compounds lacking such an assignment were split using a hierarchical fallback. Class assignment was used where available, followed by superclass assignment. In the absence of a superclass, class, and subclass, compounds that fell in this bucket were committed exclusively to the training set. The training set was evaluated using 5-fold cross-validation, and performance was assessed using average mean absolute error, median absolute error, root mean squared error, mean relative error, and median relative error, consistent with prior CCS modeling studies. Final model performance was then assessed on the held-out global test set, representing in-distribution data not encountered during training.

Results and Discussion

CCSBase2 Composition and Distribution

Following validation and cleaning, including removal of dimers, entries lacking validated SMILES representations, discordant CCS measurements, and duplicate records, 6,757 entries were removed from the initial aggregated dataset (72,910 entries), yielding a final CCSBase2 dataset of 66,153 measurements (Table and Figure ) spanning CCS values from 108.8 Å2 to 756.79 Å2. METLIN-CCS contributed the largest proportion of measurements (72.4%), followed by CCSBase (24.9%), Song et al. (1.3%), PNNL (0.7%), and AllCCS , (0.7%).

1. Attrition Counts after Each Dataset Cleaning Stage during Aggregation of External CCS Datasets into CCSBase2 .

cleaning stage datapoints remaining (#)
all data combined 72,910
dimer removal 69,868 (−3042)
SMILES validation 69,556 (−312)
CCS outlier filtering 66,653 (−2903)
deduplication 66,153 (−500)
Open in a new tab
a

Values in parentheses indicate the number of datapoints removed at each stage relative to the previous stage.

Notably, CCSBase broadened the CCS range represented in the aggregated dataset, exhibiting the greatest CCS dispersion and extending the experimental range to 756.79 Å2 (Table and Figure A).

2. Summary Statistics of CCS Values for Each Dataset Contributing to CCSBase2, Reported Prior to Cleaning and Aggregation.

dataset CSS mean (Å2) CCS median (Å2) CCS deviation (Å2)
CCSBase 197.12 180.60 55.47
PNNL 182.52 173.21 47.74
METLIN 193.11 193.50 16.35
AllCCS 172.51 158.30 40.90
Song et al 182.90 179.91 39.46
Open in a new tab

2.

2

Open in a new tab

Overview of CCSBase2. (A) Boxplots showing the distribution of collision cross section (CCS) values across source datasets contributing to CCSBase2. Boxes indicate interquartile ranges (IQR), center lines denote medians, and whiskers extend to 1.5× the IQR. (B) Histogram of CCS values for all entries in CCSBase2. (C) Histogram of mass-to-charge (m/z) values for all entries in CCSBase2. (D) Kernel density estimates of CCS values for positive and negative ions. (E) Violin plots depicting CCS distributions of common adduct types in CCSBase2. (F) Scatter plot of CCS versus m/z (i.e., IM-MS conformational space plot) colored by adduct type. Note that only datapoints for [M + H]+, [M – H]−, [M + Na]+, [M + NH4]+, [M + H – H2O]+, and [M + K]+ adducts are displayed for clarity.

The overall CCS distribution of CCSBase2 is approximately normal with a slight right skew, centered at a mean of 193.8 Å2 and a median of 191.9 Å2 (Figure B).

The relatively low global dispersion suggests that most of the training data occupy a dense, well-populated region of chemical space (∼150–250 Å2), which supports high predictive accuracy for commonly observed analytes while still allowing the model to learn from other high-CCS species. In contrast to the approximately normal CCS distribution, the mass-to-charge ratio (m/z) distribution is distinctly multimodal, mostly reflecting the distribution of positively and negatively charged ions (Figure C). The m/z values span a wide range, from ions such as pyrazole (m/z 68) reported by PNNL to polypeptides such as melittin (m/z 2,848) from CCSBase. This breadth highlights the inclusion of both small molecules and high-mass biomolecular ions in CCSBase2. Kernel density estimated (KDEs) for positive (n = 47,540) and negative (n = 18,613) polarity ions show substantial overlap (Figure D). Despite differences in ionization mode, both polarities exhibit a dominant density peak centered near 192 Å2.

A major challenge in CCS prediction lies in accounting for the conformational changes induced by different adducts. Our dataset contains 48 unique adduct types, with the most prevalent being [M + H]+, [M – H]−, and [M + Na]+, each represented by more than 15,000 observations. Less common adducts collectively account for approximately 7.4% of the dataset (n = 4,888). Importantly, distinct CCS distributional patterns across adduct types were observed, which our model had to adequately capture (Figure E). For example, [M + H]+ ions span the widest CCS range, whereas other adduct types, such as [M + NH4]+ ions, are more tightly clustered in m/z-CCS space. (Figure F). Together, these trends emphasize the need for the model to learn adduct-specific structural responses rather than relying on a single, unified representation.

Development of a Subclass Prediction Model

A major challenge encountered during dataset aggregation was the incompleteness of chemical taxonomy labels. Although ClassyFire successfully assigned superclass labels to 56,261 compounds, only 50,069 compounds were annotated at the more granular subclass level, yielding a total of 435 distinct subclasses. Because subclass assignments were required for subsequent train-test splitting, we trained an XGBClassifier model to infer missing subclass labels and increase subclass coverage. The resulting training set consisted of 21,447 unique SMILES spanning 113 distinct subclasses, while 2,218 datapoints, spanning 322 subclasses, were held out for the proxy OOD evaluation set. As shown in Table S1, our Tanimoto threshold of 0.7 blocked 81.1% of OOD compounds from being inferred by the classifier. This configuration allowed the model to confidently predict 6,360 out of 16,084 (∼40%) unlabeled samples.

The classifier exhibited strong performance on the in-distribution validation set. Using 5-fold cross-validation, the model achieved an average accuracy of 88.48% and an F1 score of 84.28%, as summarized in Table S2. Incorporating predicted subclass labels increased the subclass-labeled data to 56,429 out of 66,153, comprising 26,607 unique SMILES, with an average of two adducts per compound.

CCSBase2 Training Set vs Test Set Analysis

For CCSBase2 model development, the dataset was partitioned using an 80/20 train-test split, stratified by both subclass and adduct while ensuring the same SMILES did not show up across training and test. In total, 53,351 datapoints were assigned for training, while 12,802 were assigned for testing. Examination of the resulting split confirmed that subclass and adduct distributions in the test set closely mirrored those in the training set (Tables S3–S6). Of the 9,724 compounds without a subclass label, 5,510 were split according to class and 20 according to superclass, and the remaining 4,594 compounds lacked both class and superclass assignments. This stratification strategy preserved chemical and adduct diversity within the test set, while ensuring that the training set retained multiple adduct-specific measurements for a given compound.

To assess whether the held-out test compounds fell within the multivariate chemical space represented by the training data, we performed a leverage-based applicability-domain analysis using Morgan fingerprint substructure counts (Figure A). Leverage provides a measure of how atypical an observation’s combination of predictor values is relative to the training feature space. Compounds exceeding the warning threshold h* = 3p/n, where p is the number of fingerprint substructures and n is the number of training compounds, were considered outside the densely sampled training domain and therefore more representative of extrapolation. Under this threshold (h* = 0.265), 1.7% of training compounds and 7.5% of test compounds exceeded the leverage cutoff, indicating that most held-out compounds fell within the structural domain represented by the training set.

3.

3

Open in a new tab

(A) Applicability domain test showing the number of test-set compounds falling within the chemical space the model was trained on; the dashed black line represents the leverage threshold. (B) Histogram (bin size = 10 Å2) of CCS values in CCSBase2 train (blue) and test (orange) sets. Uniform Manifold Approximation and Projection (UMAP) projection of the global training (C) and test (D) set computed using the full descriptor space. Points colored by subclass; black points are compounds not within the top 20 subclasses.

Figure B shows the corresponding CCS bin distributions for the train and test sets. Interestingly, although CCS was not explicitly used to stratify the split, the two distributions were highly similar, suggesting that stratification by chemical subclass and adduct was sufficient to preserve the overall CCS distribution. We further assessed the chemical distinctiveness between the training and test sets using Uniform Manifold Approximation and Projection (UMAP), a nonlinear dimensionality reduction technique capable of preserving local structure in high-dimensional chemical space. The UMAP representation was generated from molecular and adduct descriptors, including exact molecular mass, signed adduct mass, OHE adduct identity, and Morgan count fingerprints. The embedding was fit exclusively on the training set (Figure C) and subsequently applied to the test set (Figure D), allowing both datasets to be visualized within the same chemical-space representation.

Model Benchmarking and Performance

After training the XGBoost regressor model, we benchmarked CCSBase2 against several established machine learning and deep learning approaches, including SigmaCCS2, GraphCCS, DeepCCS, the original CCSBase, linear regression (LR), Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP) models (Table ). All three baseline models (linear regression, SVR, and MLP) were trained using a feature set containing experimental mass, OHE adduct identity, and Molecular Quantum Numbers (MQNs) computed from SMILES via RDKit. The LR model was trained using scikit-learn's default hyperparameters. SVR used a radial basis function (RBF) kernel (C = 100, gamma = "scale"), while the MLP consisted of two hidden layers of sizes 128 and 64, respectively, with all remaining hyperparameters, including the optimizer, learning rate, L2 regularization strength, convergence tolerance, and learning rate decay, retained at scikit-learn's default settings. To ensure a fair and consistent comparison, SigmaCCS2, GraphCCS, DeepCCS, and CCSBase were retrained using the CCSBase2 training set, and all originally published hyperparameters were preserved during retraining.

3. Benchmark Performance Metrics for CCSBase2 and Comparator Models, Evaluated on the Same Held-out Test Set (n = 12,783) .

model # MAE (Å2) MDAE (Å2) RMSE (Å2) MRE (%) MDRE (%) <1% (%) <2% (%) <3% (%) <5% (%)
CCSBase2 12,802 3.33 2.33 4.89 1.73 1.23 42.3 69.5 83.6 95.0
CCSBase 12,802 4.65 3.45 7.69 2.39 1.82 30.7 53.4 70.5 90.0
DeepCCS 12,796 4.16 3.17 5.70 2.18 1.66 31.9 57.3 74.6 92.0
GraphCCS 12,787 3.41 2.30 5.09 1.79 1.20 42.9 69.0 83.0 94.0
SigmaCCS2 12,791 3.86 3.05 5.30 2.03 1.59 32.5 60.9 79.5 93.6
LR 12,802 5.53 4.57 7.15 2.89 2.41 22.6 42.6 60.0 83.6
SVR 12,802 4.48 3.34 6.30 2.33 1.76 31.5 54.9 71.5 89.8
MLP 12,802 4.71 3.67 6.33 2.45 1.92 28.2 51.6 68.9 88.9
Open in a new tab
a

MAE, mean absolute error; MDAE, median absolute error; RMSE, root mean squared error; MRE, mean relative error, calculated as the mean of |predicted–experimental|/ experimental across all test compounds; MDRE, median relative error; <1%, <2%, <3%, and <5% denote the percentage of test compounds whose predicted CCS fell within the indicated relative error of the experimental value; #, number of test compounds evaluated for each model (models with # < 12,802 failed to generate predictions for a small number of test compounds due to featurization or SMILES-parsing limitations).

AllCCS2 could not be easily benchmarked, as the model is not currently open source, and user-submitted queries are limited on the AllCCS web application. Efforts were made to include CCS Predictor 2.0 and DarkChem in our model evaluation but were ultimately omitted due to experiencing long training times on Google Colab. CCS Predictor 2.0 uses Mordred Fingerprints, a descriptor-based fingerprint that implements 1500+ 2D, and optionally 3D, molecular descriptors. Generating features for the full CCSBase2 dataset exceeded the available CPU RAM, which is a bottleneck introduced by Mordred Fingerprints’ CPU-only computation requirements. DarkChem is also limited to performing inference given only SMILES as opposed to SMILES and adducts.

Although Table reports performance on a single fixed test partition, these estimates may be influenced by sampling variability in the compounds assigned to the test set. Therefore, to assess the robustness of model comparisons, we resampled the held-out test observations with replacement and recalculated MAE from the corresponding model predictions for each bootstrap sample, generating empirical MAE distributions and 95% confidence intervals (CI) (Figure A). CCSBase2 yielded a bootstrap MAE 95% CI of 3.27–3.39 Å2, which did not overlap with those of CCSBase (4.54–4.77 Å2), DeepCCS (4.10–4.23 Å2), SigmaCCS2 (3.80–3.93 Å2), or the three baseline models (4.40–5.62 Å2 across LR, SVR, and MLP). The separation of these bootstrap distributions indicates that the observed performance differences are robust to resampling of the held-out test observations. In contrast, the GraphCCS interval (3.35–3.48 Å2) overlapped that of CCSBase2, indicating that the small difference between their point estimates was not clearly distinguishable under this resampling analysis.

4.

4

Open in a new tab

(A) Mean absolute error (MAE) confidence intervals for benchmarked models in-distribution test set predictions when bootstrapped with replacement 1,000 times. (B) MAE confidence intervals for benchmarked model’s out-of-distribution test set prediction when bootstrapped with replacement 1000 times.

GraphCCS is a deep graph convolutional neural network that predicts CCS by converting SMILES strings with specific adduct types into molecular and adduct graphs, enabling explicit learning of atom (nodes) and bond (edges) structural relationships. We believe the strong performance of CCSBase2 and GraphCCS suggests that molecular representations retaining local structural and topological information are particularly informative for CCS prediction, consistent with previous fingerprint- and graph-based CCS prediction studies. − In CCSBase2, this capability is captured through Morgan count fingerprints, which encode atom-centered substructures within a defined radius by assessing the presence of local chemical features. We found Morgan count fingerprints to be among the most influential contributors to the performance of our model. Prior to adopting Morgan count fingerprints, we evaluated a broad set of descriptors from RDKIT's rdMolDescriptors library; however, model performance plateaued at a mean relative error of 1.95%. These results suggest that local substructural representations provide information not fully captured by global molecular descriptors.

To our knowledge, existing CCS prediction models do not explicitly encode signed adduct mass as a standalone feature; rather, these approaches represent ion identity using the observed ionic m/z, which reflects the mas of the neutral molecular plus any adduct. ,, CCSBase2 introduces signed adduct mass as an additional feature, calculated by subtracting the SMILES-derived exact molecular mass from the experimentally measured ion mass, analogous to the mass-validation procedure used during database curation (Figure ). Both exact molecular mass and signed adduct mass were then incorporated as independent features for model training.

To determine whether signed adduct mass provided information beyond OHE adduct identity alone, we compared otherwise identical models trained with and without this feature. Predictions were matched across models by SMILES, adduct, and experimental CCS value to ensure that repeated measurements of the same compound-adduct pair were paired correctly. Inclusion of signed adduct mass produced a statistically significant reduction in median relative error, from 1.246% to 1.231% (Wilcoxon statistic = 39,748,978, p = 3.343 × 10‑3).

Because the aggregate performance metrics are dominated by the most heavily represented adducts ([M + H]+, [M – H]−, and [M + Na]+), they may obscure performance differences among less common adduct types. We therefore stratified predictive error by adduct type and examined median relative error as a function of training-set representation for each adduct present in the test set (Figure S3). As expected, prediction error generally decreases with increasing numbers of training observations, with the largest errors occurring primarily among sparsely represented adducts. Predictions for the multiply charged ions, such as [M + 3H]3+, exhibit particularly large prediction errors, likely due to the large impact of charge repulsion on gas-phase conformations. Notably, however, several low-abundance adducts exhibited median relative errors comparable to those of the most highly represented adducts. We suspect this pattern reflects the contribution of molecular structure to CCS prediction: because Morgan count fingerprints encode structural information independently of adduct identity, predictions for rare adducts can still draw on structural patterns learned from compounds represented elsewhere in the training set. We note, however, that performance estimates for adducts represented by very few test observations should be interpreted cautiously because of their greater sampling variability.

Performance on the held-out test set was compared with 5-fold cross-validation results to assess the consistency of model performance across different data partitions (Table ). The held-out test metrics closely matched the corresponding cross-validation estimates, providing additional evidence that the fixed test partition was not biased relative to the training data. Predictive performance was also evaluated separately for each of the 36 adduct types represented in the test set, with adduct-specific error metrics reported in Table S7.

4. Comparison of 5-Fold Cross-Validation Performance with the Held-out Test Set Evaluation .

metric value description
MAE (Å2) 3.33 MAE (test)
CV MAE (Å2) 3.36 ±0.02 5-fold CV mean ± STD
MDAE (Å2) 2.33 median absolute error (test)
CV MDAE (Å2) 2.31 ±0.01 5-fold CV mean ± STD
RMSE (Å2) 4.89 RMSE (test)
CV RMSE (Å2) 5.09 ± 0.12 5-fold CV mean ± STD
MRE (%) 1.73 MRE (test)
CV MRE (%) 1.74 ± 0.01 5-fold CV mean ± STD
MDRE (%) 1.23 MDRE (test)
CV MDRE (%) 1.20 ± 0.00 5-fold CV mean ±STD
R2 0.977 coefficient of determination (test)
CV R2 0.974 ± 0.002 5-fold CV mean ± STD
Open in a new tab
a

MAE (mean absolute error) and RMSE (root mean squared error) are reported in 2 and reflect absolute prediction error; RMSE weights larger errors more heavily than MAE, making it more sensitive to adducts with a small number of poorly predicted outliers. MDAE (median absolute error) reflects typical error while being robust to such outliers. MRE (mean relative error) and MDRE (median relative error) report the mean and median relative error (%), normalizing error by each compound's CCS value and thereby enabling comparison across adducts spanning different absolute CCS ranges.

We next examined the performance of other CCS prediction models to elucidate factors contributing to differences in predictive accuracy. SigmaCCS2, which employs a graph neural network framework, relies on explicit conformer generation to approximate molecular topology. In practice, CCS prediction typically assumes that the lowest-energy conformer dominates; however, SigmaCCS2 adopts a simplified strategy by selecting the first generated conformer rather than performing systematic ranking. Even after energy optimization, multiple candidate conformers can share comparable energies, and identifying the specific conformer sampled during IM-MS remains nontrivial. As an illustrative example, Betaine (C­[N+]­(C)­(C)­CC­(=O)­[O−]) has all 50 conformers converging to the same energy minimum after generating 50 candidates using RDKIT’s AllChem.MMFFOptimizeMoleculeConfs method. We therefore hypothesize that uncertainty regarding the gas-phase conformer sampled during IM-MS may contribute to prediction error when conformer selection is based on a single generated structure.

The comparatively weaker performance of DeepCCS may reflect underfitting under its originally published training configuration. DeepCCS was trained for 150 epochs using a batch size of 2, which is small relative to the 53,351-observation CCSBase2 training set and to batch sizes of approximately 16–64 commonly used in related neural network studies. , While small batch sizes can sometimes improve generalization in low-data regimes, they are slow to converge and are affected by gradient noise.

The performance limitations of our original CCSBase model are likely due to its clustering-based modeling strategy. CCSBase applies K-Means clustering to the full dataset and trains an independent support vector regression (SVR) model within each cluster. However, in high-dimensional feature spaces, K-Means assigns every datapoint to a cluster, even when it lies far from any cluster centroid. Because CCSBase relies on MQNs, a relatively low-dimensional (42-feature) representation, chemically dissimilar compounds can be grouped together. This results in SVR models trained on heterogeneous and limited data, reducing their capacity to capture compound-specific structure-CCS relationships.

As comparatively simple models, LR, SVR, and MLP they capture broad, coarse-grained patterns in the data but struggle to represent the subtle structural nuances that can substantially shift CCS, particularly when trained on limited data for a given compound subclass. We use these models' performance as a baseline, demonstrating that CCSBase2 has learned finer-grained structural patterns that translate into stronger predictive performance.

To further assess model generalizability, all models were evaluated using an independently curated out-of-distribution (OOD) test set. Candidate OOD compounds were assembled by merging the publicly available training datasets associated with SigmaCCS2, GraphCCS, and Deep CCS, whose original training distributions differ from CCSBase2. We first removed compounds with SMILES already present in CCSBase2 and subsequently excluded compounds with a nearest-neighbor Tanimoto similarity greater than 0.7 to any compounds in the CCSBase2 training set. This procedure yielded a structurally distinct OOD dataset containing 5,827 measurements representing 3,154 unique SMILES (Table S8 andFigure S4). Benchmarking results are summarized in Table .

5. Benchmark Performance Metrics for CCSBase2 and Comparator Models, Evaluated on the Same OOD Test Set (n = 5,827) .

model # MAE (Å2) MDAE (Å2) RMSE (Å2) MRE (%) MDRE (%) <1% (%) <2% (%) <3% (%) <5% (%)
CCSBase2 5,827 5.06 3.45 15.3 2.57 1.89 28.5 52.4 69.9 89.5
CCSBase 5,827 5.45 3.45 16.6 2.73 1.96 26.6 50.8 68.4 87.1
DeepCCS 5,789 5.42 3.62 15.8 2.75 1.99 26.8 50.2 67.6 87.6
GraphCCS 5,825 5.00 2.98 17.0 2.52 1.64 33.1 57.1 73.0 88.9
SigmaCCS2 5,827 4.90 3.36 14.7 2.53 1.83 29.0 53.2 71.0 89.4
LR 5,827 6.40 4.31 20.2 3.22 2.34 22.2 43.7 60.6 82.2
SVR 5,827 5.73 3.51 18.3 2.80 1.96 28.5 50.7 68.2 87.3
MLP 5,827 6.07 3.87 16.2 3.10 2.11 25.0 47.9 64.8 85.7
Open in a new tab
a

MAE, mean absolute error; MDAE, median absolute error; RMSE, root mean squared error; MRE, mean relative error, calculated as the mean of |predicted – experimental|/experimental across all test compounds; MDRE, median relative error; <1%, <2%, <3%, and <5% denote the percentage of test compounds whose predicted CCS fell within the indicated relative error of the experimental value; #, number of test compounds evaluated for each model (models with # <5,827 failed to generate predictions for a small number of test compounds due to featurization or SMILES-parsing limitations).

The success of CCSBase2 and GraphCCS in leveraging local atom–bond features is a key finding that offers clear guidance for future CCS prediction research, indicating which feature representations are most effective and enabling the development of specialized models that better capture local chemical structure. While CCSBase2 and GraphCCS achieved similar prediction performance using our unified benchmark dataset, CCSBase2 offers several practical advantages. First, CCSBase2 achieves strong in-distribution performance using a comparatively lightweight ML architecture. The model can be trained on an Apple Silicon M4 Pro CPU in approximately 5 minutes without hyperparameter optimization, whereas retraining GraphCCS using its published hyperparameters required approximately 9.5 hours on an NVIDIA T4 GPU. CCSBase2 also generated predictions for 12,802 compounds in less than 2 seconds on the M4 Pro CPU, compared with approximately 9.5 minutes for GraphCCS on the T4 GPU (Table S9). The longer GraphCCS inference time reflects, in part, the additional preprocessing required to construct graph-based molecular representations. Second, CCSBase2 generated predictions for all 12,802 test-set observations; GraphCCS, in contrast, generated predictions for 12,787 due to preprocessing and molecular featurization exceptions encountered in the upstream graph-construction workflow. These failures may be resolved in future implementations through improved input handling and updates to RDKit. Third, the XGBoost framework supports direct feature-level interpretation. To facilitate chemical interpretation of the Morgan count fingerprint features, we developed a web application that visualizes the atomic environments corresponding to individual fingerprint features for each compound (https://ccsbase2-fingerprint-explorer.streamlit.app/), allowing individual structural contributions to CCS predictions to be examined in chemical context (Figure S5). Finally, the CCSBase2 workflow includes a complementary subclass classifier that can infer missing chemical subclass annotations when ClassyFire does not provide assignments.

The benchmarking process we underwent in this work is critical for unbiased model evaluation and fairness. Each model was retrained on the same dataset, requiring the setup of separate Google Colab environments, modification of training pipelines to accommodate expanded adduct coverage, and custom handling of exceptions during graph and conformer generation for complex SMILES strings. Many ML fields benefit from static, community-accepted datasets, such as GLUE, SQuAD, and MMLU, that are intentionally excluded from training and used solely for model evaluation. Such datasets enable direct, reproducible, and transparent comparison across methods. To support the establishment of a similar standard within the CCS prediction community, we are publicly releasing CCSBase2, including fixed training and test set files (https://github.com/libinxulab/ccsbase2). We anticipate that this resource will facilitate reproducible benchmarking and accelerate the development of highly accurate CCS prediction models.

Conclusion

We demonstrate that CCSBase2 achieves accurate CCS prediction by leveraging features that describe local atom-bond relationships rather than global molecular structure. In particular, the use of Morgan count fingerprints, combined with a novel adduct mass descriptor, enhances feature separability and contributes to improved CCS prediction accuracy. We further show that classical ML models can achieve performance comparable to deep neural networks while maintaining fair generalization to out-of-distribution data, without substantial loss of accuracy. Importantly, CCSBase2 is less prone to running into algorithmic constraints during featurization, which can prevent CCS inference for specific compounds. Beyond model comparison, we introduce CCSBase2 as a unified, publicly available evaluation dataset with fixed training and test splits to enable fair, reproducible evaluation of future CCS prediction models. In future work, we aim to leverage these insights into local atom-bond representations to further advance neural network architectures, exploring hybrid models that integrate graph-based learning with Morgan count fingerprints to improve robustness and predictive performance.

Supplementary Material

ac6c03711_si_001.pdf (983.6KB, pdf)

Acknowledgments

We acknowledge Dylan H. Ross (now at Pacific Northwest National Laboratory), Jang Ho Cho, and Pengyu Chen for their contributions to the development of the original CCSBase model. This work was supported by the National Institutes of Health (NIH) Grants (R01ES031927 and R01ES035021 to L.X.). R.N. acknowledges the support from the Environmental Pathology and Toxicology Training Grant (NIH T32ES007032) and an NRSA Individual Predoctoral Fellowship (NIH F31ES037537).

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.analchem.6c03711.

  • Subclass classifier optimization and performance; subclass and adduct distributions in CCSBase2 training and test sets; metrics across adducts in test set; computation resource utilization; out of distribution dataset curation; illustration of CCS Fingerprint Explorer web application for the chemical interpretation of Morgan count fingerprint features learned by the XGBoost CCS prediction model; histogram of replicate CCS deviation; superclass and substructure enrichment among removed and high variance replicate measurements (PDF)

A.B. and R.S. contributed equally to this work. L.X. conceptualized the study; all authors were involved in experimental design; A.B. and R.S. contributed to the writing and editing of Python code; A.B., R.S., R.N., and G.R. contributed to the writing of the original manuscript; all authors contributed to the review and editing of the final manuscript.

The authors declare no competing financial interest.

References

  1. Mesleh M. F., Hunter J. M., Shvartsburg A. A., Schatz G. C., Jarrold M. F.. Structural Information from Ion Mobility Measurements: Effects of the Long-Range Potential. J. Phys. Chem. 1996;100(40):16082–16086. doi: 10.1021/jp961623v. [DOI] [Google Scholar]
  2. Lai Y., Koelmel J. P., Walker D. I., Price E. J., Papazian S., Manz K. E., Castilla-Fernandez D., Bowden J. A., Nikiforov V., David A., Bessonneau V., Amer B., Seethapathy S., Hu X., Lin E. Z., Jbebli A., McNeil B. R., Barupal D., Cerasa M., Xie H., Kalia V., Nandakumar R., Singh R., Tian Z., Gao P., Zhao Y., Froment J., Rostkowski P., Dubey S., Coufalikova K., Selicova H., Hecht H., Liu S., Udhani H. H., Restituito S., Tchou-Wong K.-M., Lu K., Martin J. W., Warth B., Godri Pollitt K. J., Klanova J., Fiehn O., Metz T. O., Pennell K. D., Jones D. P., Miller G. W.. High-Resolution Mass Spectrometry for Human Exposomics: Expanding Chemical Space Coverage. Environ. Sci. Technol. 2024;58:12784–12822. doi: 10.1021/acs.est.4c01156. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Foster M., Rainey M., Watson C., Dodds J. N., Kirkwood K. I., Fernández F. M., Baker E. S.. Uncovering PFAS and Other Xenobiotics in the Dark Metabolome Using Ion Mobility Spectrometry, Mass Defect Analysis, and Machine Learning. Environ. Sci. Technol. 2022;56(12):9133–9143. doi: 10.1021/acs.est.2c00201. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Metz T. O., Baker E. S., Schymanski E. L., Renslow R. S., Thomas D. G., Causon T. J., Webb I. K., Hann S., Smith R. D., Teeguarden J. G.. Integrating Ion Mobility Spectrometry into Mass Spectrometry-Based Exposome Measurements: What Can It Add and How Far Can It Go? Bioanalysis. 2017;9(1):81–98. doi: 10.4155/bio-2016-0244. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Blaženović I., Kind T., Ji J., Fiehn O.. Software Tools and Approaches for Compound Identification of LC-MS/MS Data in Metabolomics. Metabolites. 2018;8(2):31. doi: 10.3390/metabo8020031. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Zhang X., Gillig K. J., Sun W.. Recent Advances in Ion Mobility Instrumentation for Targeted Analysis. TrAC Trends in Analytical Chemistry. 2026;194:118507. doi: 10.1016/j.trac.2025.118507. [DOI] [Google Scholar]
  7. Fenn L. S., Kliman M., Mahsut A., Zhao S. R., McLean J. A.. Characterizing Ion Mobility-Mass Spectrometry Conformation Space for the Analysis of Complex Biological Samples. Anal Bioanal Chem. 2009;394:235–244. doi: 10.1007/s00216-009-2666-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Stow S. M., Causon T. J., Zheng X., Kurulugama R. T., Mairinger T., May J. C., Rennie E. E., Baker E. S., Smith R. D., McLean J. A., Hann S., Fjeldsted J. C.. An Interlaboratory Evaluation of Drift Tube Ion Mobility–Mass Spectrometry Collision Cross Section Measurements. Anal. Chem. 2017;89(17):9048–9055. doi: 10.1021/acs.analchem.7b01729. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Kartowikromo K. Y., Olajide O. E., Hamid A. M.. Collision Cross Section (CCS) Measurement and Prediction Methods in Omics. J Mass Spectrom. 2023;58(9):e4973. doi: 10.1002/jms.4973. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Hernández-Mesa M., D’Atri V., Barknowitz G., Fanuel M., Pezzatti J., Dreolin N., Ropartz D., Monteau F., Vigneau E., Rudaz S., Stead S., Rogniaux H., Guillarme D., Dervilly G., Le Bizec B.. Interlaboratory and Interplatform Study of Steroids Collision Cross Section by Traveling Wave Ion Mobility Spectrometry. Anal. Chem. 2020;92(7):5013–5022. doi: 10.1021/acs.analchem.9b05247. [DOI] [PubMed] [Google Scholar]
  11. Paglia G., Angel P., Williams J. P., Richardson K., Olivos H. J., Thompson J. W., Menikarachchi L., Lai S., Walsh C., Moseley A., Plumb R. S., Grant D. F., Palsson B. O., Langridge J., Geromanos S., Astarita G.. Ion Mobility-Derived Collision Cross Section As an Additional Measure for Lipid Fingerprinting and Identification. Anal. Chem. 2015;87(2):1137–1144. doi: 10.1021/ac503715v. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Gupta R.. Introduction of 4D-Proteomics: A Breakthrough in Analyzing the Plant Proteomes During Stress Conditions. Physiol Plant. 2026;178(1):e70735. doi: 10.1111/ppl.70735. [DOI] [PubMed] [Google Scholar]
  13. Butalewicz J. P., Brodbelt J. S.. Development and Modification of Ion Mobility Mass Spectrometry Instrumentation for Enhanced Analysis of Biological Molecules. J Am Soc Mass Spectrom. 2025;36(11):2351–2369. doi: 10.1021/jasms.5c00222. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. May J. C., McLean J. A.. Ion Mobility-Mass Spectrometry: Time-Dispersive Instrumentation. Anal. Chem. 2015;87(3):1422–1436. doi: 10.1021/ac504720m. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Pringle S. D., Giles K., Wildgoose J. L., Williams J. P., Slade S. E., Thalassinos K., Bateman R. H., Bowers M. T., Scrivens J. H.. An Investigation of the Mobility Separation of Some Peptide and Protein Ions Using a New Hybrid Quadrupole/Travelling Wave IMS/Oa-ToF Instrument. International Journal of Mass Spectrometry. 2007;261(1):1–12. doi: 10.1016/j.ijms.2006.07.021. [DOI] [Google Scholar]
  16. Lu G., Xu S., Huang P., Li L.. Enhancing Lipidomics With High-Resolution Ion Mobility-Mass Spectrometry. Proteomics. 2026;26:e70026. doi: 10.1002/pmic.70026. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Dodds J. N., Baker E. S.. Ion Mobility Spectrometry: Fundamental Concepts, Instrumentation, Applications, and the Road Ahead. J Am Soc Mass Spectrom. 2019;30(11):2185–2195. doi: 10.1007/s13361-019-02288-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Hinnenkamp V., Klein J., Meckelmann S. W., Balsaa P., Schmidt T. C., Schmitz O. J.. Comparison of CCS Values Determined by Traveling Wave Ion Mobility Mass Spectrometry and Drift Tube Ion Mobility Mass Spectrometry. Anal. Chem. 2018;90(20):12042–12050. doi: 10.1021/acs.analchem.8b02711. [DOI] [PubMed] [Google Scholar]
  19. Li X., Wang H., Jiang M., Ding M., Xu X., Xu B., Zou Y., Yu Y., Yang W.. Collision Cross Section Prediction Based on Machine Learning. Molecules. 2023;28(10):4050. doi: 10.3390/molecules28104050. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Ross D. H., Cho J. H., Xu L.. Breaking Down Structural Diversity for Comprehensive Prediction of Ion-Neutral Collision Cross Sections. Anal. Chem. 2020;92(6):4548–4557. doi: 10.1021/acs.analchem.9b05772. [DOI] [PubMed] [Google Scholar]
  21. Zhou Z., Luo M., Chen X., Yin Y., Xiong X., Wang R., Zhu Z.-J.. Ion Mobility Collision Cross-Section Atlas for Known and Unknown Metabolite Annotation in Untargeted Metabolomics. Nat Commun. 2020;11:4334. doi: 10.1038/s41467-020-18171-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Rainey M. A., Watson C. A., Asef C. K., Foster M. R., Baker E. S., Fernández F. M.. CCS Predictor 2.0: An Open-Source Jupyter Notebook Tool for Filtering Out False Positives in Metabolomics. Anal. Chem. 2022;94(50):17456–17466. doi: 10.1021/acs.analchem.2c03491. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Zhang H., Luo M., Wang H., Ren F., Yin Y., Zhu Z.-J.. AllCCS2: Curation of Ion Mobility Collision Cross-Section Atlas for Small Molecules Using Comprehensive Molecular Representations. Anal. Chem. 2023;95(37):13913–13921. doi: 10.1021/acs.analchem.3c02267. [DOI] [PubMed] [Google Scholar]
  24. Guo R., Zhang Y., Liao Y., Yang Q., Xie T., Fan X., Lin Z., Chen Y., Lu H., Zhang Z.. Highly Accurate and Large-Scale Collision Cross Sections Prediction with Graph Neural Networks. Commun Chem. 2023;6:139. doi: 10.1038/s42004-023-00939-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Liao Y., Chen R., Xie T., Yu C., Wang J., Zhang Z., Lu H.. SigmaCCS2: Collision Cross Section Prediction Using Molecular Topological and Geometric Information with Graph Neural Network. Chemometrics and Intelligent Laboratory Systems. 2026;268:105588. doi: 10.1016/j.chemolab.2025.105588. [DOI] [Google Scholar]
  26. Xie T., Yang Q., Sun J., Zhang H., Wang Y., Zhang Z., Lu H.. Large-Scale Prediction of Collision Cross-Section with Very Deep Graph Convolutional Network for Small Molecule Identification. Chemometrics and Intelligent Laboratory Systems. 2024;252:105177. doi: 10.1016/j.chemolab.2024.105177. [DOI] [Google Scholar]
  27. Plante P.-L., Francovic-Fontaine É., May J. C., McLean J. A., Baker E. S., Laviolette F., Marchand M., Corbeil J.. Predicting Ion Mobility Collision Cross-Sections Using a Deep Neural Network: DeepCCS. Anal. Chem. 2019;91(8):5191–5199. doi: 10.1021/acs.analchem.8b05821. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Colby S. M., Nunẽz J. R., Hodas N. O., Corley C. D., Renslow R. R.. Deep Learning to Generate in Silico Chemical Property Libraries and Candidate Molecules for Small Molecule Identification in Complex Samples. Anal. Chem. 2020;92(2):1720–1729. doi: 10.1021/acs.analchem.9b02348. [DOI] [PubMed] [Google Scholar]
  29. Baker E. S., Hoang C., Uritboonthai W., Heyman H. M., Pratt B., MacCoss M., MacLean B., Plumb R., Aisporna A., Siuzdak G.. METLIN-CCS: An Ion Mobility Spectrometry Collision Cross Section Database. Nat Methods. 2023;20(12):1836–1837. doi: 10.1038/s41592-023-02078-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Chen, T. ; Guestrin, C. . XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016; p 785–794. 10.1145/2939672.2939785 [DOI] [Google Scholar]
  31. Zheng X., Aly N. A., Zhou Y., Dupuis K. T., Bilbao A., Paurus V. L., Orton D. J., Wilson R., Payne S. H., Smith R. D., Baker E. S.. A Structural Examination and Collision Cross Section Database for over 500 Metabolites and Xenobiotics Using Drift Tube Ion Mobility Spectrometry. Chem. Sci. 2017;8(11):7724–7736. doi: 10.1039/C7SC03464D. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Song X.-C., Canellas E., Dreolin N., Goshawk J., Nerin C.. A Collision Cross Section Database for Extractables and Leachables from Food Contact Materials. J. Agric. Food Chem. 2022;70(14):4457–4466. doi: 10.1021/acs.jafc.2c00724. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Zhu Z., Xie C., Lin S., Xu X., Niu S., Guo L., Dong J.. Large Language Model-Empowered Compound Collision Cross-Section Prediction. Anal. Chem. 2025;97(36):19791–19800. doi: 10.1021/acs.analchem.5c03492. [DOI] [PubMed] [Google Scholar]
  34. Broeckling C. D., Ganna A., Layer M., Brown K., Sutton B., Ingelsson E., Peers G., Prenni J. E.. Enabling Efficient and Confident Annotation of LC–MS Metabolomics Data through MS1 Spectrum and Time Prediction. Anal. Chem. 2016;88(18):9226–9234. doi: 10.1021/acs.analchem.6b02479. [DOI] [PubMed] [Google Scholar]
  35. Kim S., Chen J., Cheng T., Gindulyte A., He J., He S., Li Q., Shoemaker B. A., Thiessen P. A., Yu B., Zaslavsky L., Zhang J., Bolton E. E.. PubChem 2025 Update. Nucleic Acids Research. 2025;53(D1):D1516–D1525. doi: 10.1093/nar/gkae1059. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Hines K. M., Ross D. H., Davidson K. L., Bush M. F., Xu L.. Large-Scale Structural Characterization of Drug and Drug-Like Compounds by High-Throughput Ion Mobility-Mass Spectrometry. Anal. Chem. 2017;89(17):9023–9030. doi: 10.1021/acs.analchem.7b01709. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. The RDKit Documentation–The RDKit 2026.03.5 D; Greg Landrum and Other RDKit Contributors, 2026; https://www.rdkit.org/docs/ (accessed 2026-08-17). [Google Scholar]
  38. Conroy M. J., Andrews R. M., Andrews S., Cockayne L., Dennis E. A., Fahy E., Gaud C., Griffiths W. J., Jukes G., Kolchin M., Mendivelso K., Lopez-Clavijo A. F., Ready C., Subramaniam S., O’Donnell V. B.. LIPID MAPS: Update to Databases and Tools for the Lipidomics Community. Nucleic Acids Res. 2024;52(D1):D1677–D1682. doi: 10.1093/nar/gkad896. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Djoumbou Feunang Y., Eisner R., Knox C., Chepelev L., Hastings J., Owen G., Fahy E., Steinbeck C., Subramanian S., Bolton E., Greiner R., Wishart D. S.. ClassyFire: Automated Chemical Classification with a Comprehensive, Computable Taxonomy. J Cheminform. 2016;8:61. doi: 10.1186/s13321-016-0174-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Heller S. R., McNaught A., Pletnev I., Stein S., Tchekhovskoi D.. InChI, the IUPAC International Chemical Identifier. J Cheminform. 2015;7:23. doi: 10.1186/s13321-015-0068-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Rogers D., Hahn M.. Extended-Connectivity Fingerprints. J. Chem. Inf. Model. 2010;50(5):742–754. doi: 10.1021/ci100050t. [DOI] [PubMed] [Google Scholar]
  42. Willett P., Barnard J. M., Downs G. M.. Chemical Similarity Searching. J. Chem. Inf. Comput. Sci. 1998;38(6):983–996. doi: 10.1021/ci9800211. [DOI] [Google Scholar]
  43. Gramatica P.. Principles of QSAR Models Validation: Internal and External. QSAR & Combinatorial Science. 2007;26(5):694–701. doi: 10.1002/qsar.200610151. [DOI] [Google Scholar]
  44. Jaworska J., Nikolova-Jeliazkova N., Aldenberg T.. QSAR Applicabilty Domain Estimation by Projection of the Training Set Descriptor Space: A Review. Altern Lab Anim. 2005;33(5):445–459. doi: 10.1177/026119290503300508. [DOI] [PubMed] [Google Scholar]
  45. McInnes, L ; Healy, J ; Melville, J . UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv September 18, 2020; 10.48550/arXiv.1802.03426. [DOI]
  46. Ross D. H., Cho J. H., Xu L.. Breaking Down Structural Diversity for Comprehensive Prediction of Ion-Neutral Collision Cross Sections. Anal. Chem. 2020;92(6):4548–4557. doi: 10.1021/acs.analchem.9b05772. [DOI] [PubMed] [Google Scholar]
  47. Nguyen K. T., Blum L. C., van Deursen R., Reymond J.-L.. Classification of Organic Molecules by Molecular Quantum Numbers. ChemMedChem. 2009;4(11):1803–1805. doi: 10.1002/cmdc.200900317. [DOI] [PubMed] [Google Scholar]
  48. Moriwaki H., Tian Y.-S., Kawashita N., Takagi T.. Mordred: A Molecular Descriptor Calculator. J Cheminform. 2018;10:4. doi: 10.1186/s13321-018-0258-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Yang F., van Herwerden D., Preud’homme H., Samanipour S.. Collision Cross Section Prediction with Molecular Fingerprint Using Machine Learning. Molecules. 2022;27(19):6424. doi: 10.3390/molecules27196424. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Guo R., Zhang Y., Liao Y., Yang Q., Xie T., Fan X., Lin Z., Chen Y., Lu H., Zhang Z.. Highly Accurate and Large-Scale Collision Cross Sections Prediction with Graph Neural Networks. Commun Chem. 2023;6:139. doi: 10.1038/s42004-023-00939-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Wang C., Yuan C., Wang Y., Shi Y., Zhang T., Patti G. J.. Predicting Collision Cross-Section Values for Small Molecules through Chemical Class-Based Multimodal Graph Attention Network. J. Chem. Inf. Model. 2024;64(16):6305–6315. doi: 10.1021/acs.jcim.3c01934. [DOI] [PubMed] [Google Scholar]
  52. De Cripan S. M., Arora T., Olomí A., Canela N., Siuzdak G., Domingo-Almenara X.. Predicting the Predicted: A Comparison of Machine Learning-Based Collision Cross-Section Prediction Models for Small Molecules. Anal. Chem. 2024;96(22):9088–9096. doi: 10.1021/acs.analchem.4c00630. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Wang, A. ; Singh, A. ; Michael, J. ; Hill, F. ; Levy, O. ; Bowman, S. R. . GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding. arXiv 2018; 10.48550/ARXIV.1804.07461. [DOI]
  54. Rajpurkar, P. ; Zhang, J. ; Lopyrev, K. ; Liang, P. . SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing; Association for Computational Linguistics: Austin, Texas, 2016; pp 2383–2392. 10.18653/v1/D16-1264. [DOI] [Google Scholar]
  55. Srivastava, A. ; Rastogi, A. ; Rao, A. ; Shoeb, A. A. M. ; Abid, A. ; Fisch, A. ; Brown, A. R. ; Santoro, A. ; Gupta, A. ; Garriga-Alonso, A. ; Kluska, A. ; Lewkowycz, A. ; Agarwal, A. ; Power, A. ; Ray, A. ; Warstadt, A. ; Kocurek, A. W. ; Safaya, A. ; Tazarv, A. ; Xiang, A. ; Parrish, A. ; Nie, A. ; Hussain, A. ; Askell, A. ; Dsouza, A. ; Slone, A. ; Rahane, A. ; Iyer, A. S. ; Andreassen, A. ; Madotto, A. ; Santilli, A. ; Stuhlmüller, A. ; Dai, A. ; La, A. ; Lampinen, A. ; Zou, A. ; Jiang, A. ; Chen, A. ; Vuong, A. ; Gupta, A. ; Gottardi, A. ; Norelli, A. ; Venkatesh, A. ; Gholamidavoodi, A. ; Tabassum, A. ; Menezes, A. ; Kirubarajan, A. ; Mullokandov, A. ; Sabharwal, A. ; Herrick, A. ; Efrat, A. ; Erdem, A. ; Karakaş, A. ; Roberts, B. R. ; Loe, B. S. ; Zoph, B. ; Bojanowski, B. ; Özyurt, B. ; Hedayatnia, B. ; Neyshabur, B. ; Inden, B. ; Stein, B. ; Ekmekci, B. ; Lin, B. Y. ; Howald, B. ; Orinion, B. ; Diao, C. ; Dour, C. ; Stinson, C. ; Argueta, C. ; Ramírez, C. F. ; Singh, C. ; Rathkopf, C. ; Meng, C. ; Baral, C. ; Wu, C. ; Callison-Burch, C. ; Waites, C. ; Voigt, C. ; Manning, C. D. ; Potts, C. ; Ramirez, C. ; Rivera, C. E. ; Siro, C. ; Raffel, C. ; Ashcraft, C. ; Garbacea, C. ; Sileo, D. ; Garrette, D. ; Hendrycks, D. ; Kilman, D. ; Roth, D. ; Freeman, D. ; Khashabi, D. ; Levy, D. ; González, D. M. ; Perszyk, D. ; Hernandez, D. ; Chen, D. ; Ippolito, D. ; Gilboa, D. ; Dohan, D. ; Drakard, D. ; Jurgens, D. ; Datta, D. ; Ganguli, D. ; Emelin, D. ; Kleyko, D. ; Yuret, D. ; Chen, D. ; Tam, D. ; Hupkes, D. ; Misra, D. ; Buzan, D. ; Mollo, D. C. ; Yang, D. ; Lee, D.-H. ; Schrader, D. ; Shutova, E. ; Cubuk, E. D. ; Segal, E. ; Hagerman, E. ; Barnes, E. ; Donoway, E. ; Pavlick, E. ; Rodola, E. ; Lam, E. ; Chu, E. ; Tang, E. ; Erdem, E. ; Chang, E. ; Chi, E. A. ; Dyer, E. ; Jerzak, E. ; Kim, E. ; Manyasi, E. E. ; Zheltonozhskii, E. ; Xia, F. ; Siar, F. ; Martínez-Plumed, F. ; Happé, F. ; Chollet, F. ; Rong, F. ; Mishra, G. ; Winata, G. I. ; de Melo, G. ; Kruszewski, G. ; Parascandolo, G. ; Mariani, G. ; Wang, G. ; Jaimovitch-López, G. ; Betz, G. ; Gur-Ari, G. ; Galijasevic, H. ; Kim, H. ; Rashkin, H. ; Hajishirzi, H. ; Mehta, H. ; Bogar, H. ; Shevlin, H. ; Schütze, H. ; Yakura, H. ; Zhang, H. ; Wong, H. M. ; Ng, I. ; Noble, I. ; Jumelet, J. ; Geissinger, J. ; Kernion, J. ; Hilton, J. ; Lee, J. ; Fisac, J. F. ; Simon, J. B. ; Koppel, J. ; Zheng, J. ; Zou, J. ; Kocoń, J. ; Thompson, J. ; Wingfield, J. ; Kaplan, J. ; Radom, J. ; Sohl-Dickstein, J. ; Phang, J. ; Wei, J. ; Yosinski, J. ; Novikova, J. ; Bosscher, J. ; Marsh, J. ; Kim, J. ; Taal, J. ; Engel, J. ; Alabi, J. ; Xu, J. ; Song, J. ; Tang, J. ; Waweru, J. ; Burden, J. ; Miller, J. ; Balis, J. U. ; Batchelder, J. ; Berant, J. ; Frohberg, J. ; Rozen, J. ; Hernandez-Orallo, J. ; Boudeman, J. ; Guerr, J. ; Jones, J. ; Tenenbaum, J. B. ; Rule, J. S. ; Chua, J. ; Kanclerz, K. ; Livescu, K. ; Krauth, K. ; Gopalakrishnan, K. ; Ignatyeva, K. ; Markert, K. ; Dhole, K. D. ; Gimpel, K. ; Omondi, K. ; Mathewson, K. ; Chiafullo, K. ; Shkaruta, K. ; Shridhar, K. ; McDonell, K. ; Richardson, K. ; Reynolds, L. ; Gao, L. ; Zhang, L. ; Dugan, L. ; Qin, L. ; Contreras-Ochando, L. ; Morency, L.-P. ; Moschella, L. ; Lam, L. ; Noble, L. ; Schmidt, L. ; He, L. ; Colón, L. O. ; Metz, L. ; Śenel, L. K. ; Bosma, M. ; Sap, M. ; ter Hoeve, M. ; Farooqi, M. ; Faruqui, M. ; Mazeika, M. ; Baturan, M. ; Marelli, M. ; Maru, M. ; Quintana, M. J. R. ; Tolkiehn, M. ; Giulianelli, M. ; Lewis, M. ; Potthast, M. ; Leavitt, M. L. ; Hagen, M. ; Schubert, M. ; Baitemirova, M. O. ; Arnaud, M. ; McElrath, M. ; Yee, M. A. ; Cohen, M. ; Gu, M. ; Ivanitskiy, M. ; Starritt, M. ; Strube, M. ; Swędrowski, M. ; Bevilacqua, M. ; Yasunaga, M. ; Kale, M. ; Cain, M. ; Xu, M. ; Suzgun, M. ; Walker, M. ; Tiwari, M. ; Bansal, M. ; Aminnaseri, M. ; Geva, M. ; Gheini, M. ; T, M. V. ; Peng, N. ; Chi, N. A. ; Lee, N. ; Krakover, N. G.-A. ; Cameron, N. ; Roberts, N. ; Doiron, N. ; Martinez, N. ; Nangia, N. ; Deckers, N. ; Muennighoff, N. ; Keskar, N. S. ; Iyer, N. S. ; Constant, N. ; Fiedel, N. ; Wen, N. ; Zhang, O. ; Agha, O. ; Elbaghdadi, O. ; Levy, O. ; Evans, O. ; Casares, P. A. M. ; Doshi, P. ; Fung, P. ; Liang, P. P. ; Vicol, P. ; Alipoormolabashi, P. ; Liao, P. ; Liang, P. ; Chang, P. ; Eckersley, P. ; Htut, P. M. ; Hwang, P. ; Miłkowski, P. ; Patil, P. ; Pezeshkpour, P. ; Oli, P. ; Mei, Q. ; Lyu, Q. ; Chen, Q. ; Banjade, R. ; Rudolph, R. E. ; Gabriel, R. ; Habacker, R. ; Risco, R. ; Millière, R. ; Garg, R. ; Barnes, R. ; Saurous, R. A. ; Arakawa, R. ; Raymaekers, R. ; Frank, R. ; Sikand, R. ; Novak, R. ; Sitelew, R. ; LeBras, R. ; Liu, R. ; Jacobs, R. ; Zhang, R. ; Salakhutdinov, R. ; Chi, R. ; Lee, R. ; Stovall, R. ; Teehan, R. ; Yang, R. ; Singh, S. ; Mohammad, S. M. ; Anand, S. ; Dillavou, S. ; Shleifer, S. ; Wiseman, S. ; Gruetter, S. ; Bowman, S. R. ; Schoenholz, S. S. ; Han, S. ; Kwatra, S. ; Rous, S. A. ; Ghazarian, S. ; Ghosh, S. ; Casey, S. ; Bischoff, S. ; Gehrmann, S. ; Schuster, S. ; Sadeghi, S. ; Hamdan, S. ; Zhou, S. ; Srivastava, S. ; Shi, S. ; Singh, S. ; Asaadi, S. ; Gu, S. S. ; Pachchigar, S. ; Toshniwal, S. ; Upadhyay, S. ; Shyamolima; Debnath; Shakeri, S. ; Thormeyer, S. ; Melzi, S. ; Reddy, S. ; Makini, S. P. ; Lee, S.-H. ; Torene, S. ; Hatwar, S. ; Dehaene, S. ; Divic, S. ; Ermon, S. ; Biderman, S. ; Lin, S. ; Prasad, S. ; Piantadosi, S. T. ; Shieber, S. M. ; Misherghi, S. ; Kiritchenko, S. ; Mishra, S. ; Linzen, T. ; Schuster, T. ; Li, T. ; Yu, T. ; Ali, T. ; Hashimoto, T. ; Wu, T.-L. ; Desbordes, T. ; Rothschild, T. ; Phan, T. ; Wang, T. ; Nkinyili, T. ; Schick, T. ; Kornev, T. ; Tunduny, T. ; Gerstenberg, T. ; Chang, T. ; Neeraj, T. ; Khot, T. ; Shultz, T. ; Shaham, U. ; Misra, V. ; Demberg, V. ; Nyamai, V. ; Raunak, V. ; Ramasesh, V. ; Prabhu, V. U. ; Padmakumar, V. ; Srikumar, V. ; Fedus, W. ; Saunders, W. ; Zhang, W. ; Vossen, W. ; Ren, X. ; Tong, X. ; Zhao, X. ; Wu, X. ; Shen, X. ; Yaghoobzadeh, Y. ; Lakretz, Y. ; Song, Y. ; Bahri, Y. ; Choi, Y. ; Yang, Y. ; Hao, Y. ; Chen, Y. ; Belinkov, Y. ; Hou, Y. ; Hou, Y. ; Bai, Y. ; Seid, Z. ; Zhao, Z. ; Wang, Z. ; Wang, Z. J. ; Wang, Z. ; Wu, Z. . Beyond the Imitation Game: Quantifying and Extrapolating the Capabilities of Language Models. arXiv 2022; arXiv:2206.04615 10.48550/ARXIV.2206.04615 [DOI]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

ac6c03711_si_001.pdf (983.6KB, pdf)

Articles from Analytical Chemistry are provided here courtesy of American Chemical Society

RESOURCES