MutAIverse: an AI-powered, mechanism-backed platform for discovering novel DNA adducts and their precursor genotoxins

Abstract

Genotoxin exposure leads to DNA adduct formation, potentially causing mutations if unrepaired. Current DNA adductomics platforms or analytical workflows are limited by incomplete spectral libraries, reliance on experimentally validated adducts, limited cellular contexts, and inefficient computational methodologies. We introduce MutAIverse, an advanced AI-driven DNA adductomics analysis platform that overcomes these limitations by leveraging intracellular mechanistic modeling of genotoxin bioactivation to construct a comprehensive DNA adduct library and offers advanced spectral mapping. MutAIverse integrates experimentally validated and chemically valid putative DNA adducts, enabling interpretable retrograde tracking to parental genotoxins. Validation of MutAIverse against experimental MS/MS datasets demonstrated the detection of both known and novel adducts. Furthermore, application to in-house generated DNA adductomics data from tissue biopsies of smokeless tobacco-induced head and neck cancer patients revealed selective enrichment of both novel and known DNA adducts; a subset of them was validated using MS/MS analysis. Collectively, MutAIverse provides a robust, end-to-end, and interpretable platform for advanced DNA adductomics analysis.

Scientific Contribution

This study introduces MutAIverse, a generative AI‑powered platform that constructs a comprehensive, mechanism‑backed DNA adduct library from genotoxin metabolism simulations, enables interpretable retrograde tracing to parental genotoxins, and provides an end‑to‑end analytical workflow. By outperforming existing spectral libraries in coverage and experimental validation across multiple datasets, MutAIverse offers a robust tool for advancing DNA adductomics in toxicology and cancer research.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13321-026-01238-x.

Introduction

Despite its resilience as the master blueprint of cellular architecture, deoxyribonucleic acid (DNA) remains susceptible to damage from exogenous chemical genotoxins [1]. Although cellular DNA repair mechanisms effectively recognize and repair most DNA damage, occasional failures can lead to mutations during replication or transcription [2, 3]. Electrophilic genotoxins can directly or indirectly modify nitrogenous bases by forming covalent bonds, altering their chemical properties [4]. Other genotoxins, with neutral or weak electrophilic properties, are transformed by cellular enzymes into reactive oxygen and nitrogen species (ROS/RNS) [5]. These highly reactive species can then bind to nitrogenous bases, disrupting the integrity and function of DNA. The vast array of chemicals in the environment, combined with the global burden of cancer, highlights the urgent need for advanced biomonitoring techniques like DNA adductomics [6, 7]. Importantly, DNA adductomics is a powerful and rapidly evolving technique that has successfully linked specific environmental genotoxins to particular cancers. For example, exposure to polycyclic aromatic hydrocarbons (PAHs) is associated with lung cancer, aromatic amines with bladder cancer, and Aflatoxin B1 with liver cancer [8, 9]. These findings demonstrate the potential of DNA adductomics to identify carcinogens, elucidate their mechanisms of action, and improve our understanding of cancer development.

To date, various methods have been used to biomonitor DNA adducts in biospecimens, including ³²P-post labeling, immunoassays, fluorescence labeling, and mass spectrometry-based approaches [5]. Advancements in nanopore sequencing, especially Oxford Nanopore Technology (ONT) [5, 10], allow for the detection and characterization of DNA adducts at single-nucleobase resolution, preserving their native forms and positional context [10]. High-Resolution Mass Spectrometry (HRMS) has significantly advanced the field by offering superior sensitivity and accuracy in detecting and characterizing a wide range of DNA adducts, enabling researchers to identify specific chemical exposures and their biological consequences [4, 5, 11]. Due to the vast diversity of carcinogens and genotoxins and the resulting complexity of DNA adducts, advanced analytical techniques are essential for their purification, identification, and characterization [12, 13]. While DNA adducts vary in stability, some are readily detected in biospecimens, while others are unstable, decomposing spontaneously at body temperature or undergoing rapid repair [14]. For example, alkylation of purine bases often destabilizes the glycosidic bond, leading to apurinic sites. In contrast, pyrimidine adducts tend to be more stable, making them preferable biomarkers due to their abundance and ease of analysis [5]. Mass spectrometry (MS) has become a crucial tool for identifying DNA adducts, but its efficacy is hampered by several challenges [15, 16]. First, existing DNA adduct-specific spectral libraries are limited in scope. Widely used databases like the Human Metabolome Database (HMDB) [17], METLIN [18], MassBank [19], and the Kyoto Encyclopedia of Genes and Genomes (KEGG) [20] are not designed for DNA adductomics and contain few DNA adduct entries. While dedicated resources like DNA Adduct Portal[21] and The DNA adduct database [15] have emerged, identifying novel adducts remains difficult. This highlights the need for a comprehensive DNA adduct spectral library to improve detection capabilities and advance our understanding of adduct structural diversity [16]. Second, standard DNA adductome analysis lacks robust computational methods to connect both known and unknown adducts to their parental genotoxins [4]. This limitation, coupled with the restricted scope of existing MS spectral libraries, hinders the identification of potential adducts. Finally, traditional screening methods for MS-based DNA adductomics datasets often lack accuracy and explanatory power.

To address these challenges, we introduce MutAIverse, an advanced AI-driven platform that overcomes the limitations of conventional approaches. Unlike traditional methods that rely on limited experimental data, MutAIverse utilizes intracellular simulation modeling of genotoxin metabolism and bioactivation, incorporating factors like oxidative stress and electrophilicity. This approach generated a comprehensive, end-to-end spectral mapping workflow for thorough identification of DNA adducts in biospecimens. We rigorously validated the MutAIverse spectral library and its computational workflow using diverse public and in-house DNA adductomics datasets. Taken together, MutAIverse is a standalone Python package that provides an extensive MS/MS-based DNA adduct spectral library, streamlined analytical workflows, and sophisticated computational tools for accurately identifying DNA adducts and their parental genotoxins.

Results

In silico modeling of genotoxin metabolism reveals a vast landscape of potential DNA adducts

Accurate and comprehensive identification of DNA adducts from MS/MS spectra requires a robust reference library with extensive coverage. Existing libraries, limited by reliance on experimentally derived spectra, contain fewer than 400 entries. In contrast, computational approaches, especially those using generative AI, can leverage DNA adduct-associated chemical properties to predict novel DNA adducts. However, these approaches are limited by the scarcity of training data (fewer than 400 known adducts), hindering their ability to capture the full chemical diversity of DNA adducts. To overcome this limitation, we developed a novel approach independent of experimentally validated DNA adducts. We introduce MutAIverse, an AI-powered, mechanism-based platform that provides a comprehensive workflow for identifying both known and novel DNA adducts from MS/MS spectra (Fig. 1a). MutAIverse operates on the principle that modeling the bioactivation and biotransformation of potential adduct-forming genotoxins enables the generation of chemically valid DNA adducts. Briefly, chemical genotoxins can either directly form DNA adducts or undergo intracellular phase 1 and 2 metabolism, producing reactive or non-reactive intermediates. Reactive intermediates, such as electrophiles, can then form covalent bonds with DNA bases, resulting in adducts. This comprehensive modeling approach enables MutAIverse to accurately simulate diverse pathways of DNA adduct formation and identify a wide range of adducts. We began by compiling a list of 3,288 experimentally validated genotoxins, assuming they all contribute to DNA adduct formation, either directly or indirectly (Fig. 1b, Supplementary Table 1). Notably, a systematic comparison with IARC Group 1 carcinogens (Supplementary Table 2) indicates that the chemically defined subset (~ 21 out of 42 compounds) is covered by MutAIverse. Importantly, this partial overlap aligns with the mechanistic distinction between genotoxic and non-genotoxic carcinogens. Because non-genotoxic carcinogens drive neoplastic transformation through alternative, non-DNA-reactive mechanisms such as epigenetic dysregulation, sustained cellular proliferation, or chronic inflammation [22] they do not form covalent DNA adducts and therefore fall outside the scope. Any missing or partially represented adducts are explicitly documented and primarily reflect limitations in available mechanistic knowledge rather than model omission. Chemical composition analysis revealed that these genotoxins span diverse chemical classes, primarily benzenoids or organoheterocyclic compounds, reflecting significant chemical diversity (Fig. 1c, Supplementary Table 3). We then used BioTransformer 3.0 [23] to simulate the intracellular metabolism of these genotoxins, predicting the resulting metabolites or fragments via phase 1 and phase 2 enzyme rules. Biotransformation simulations showed that 92.2% (3,033) of the genotoxins were metabolized in silico. Further analysis revealed that Non-biotransformed (NBT) genotoxins occupied a similar chemical space as biotransformed (BT) genotoxins (Fig. 1d), except for enrichment in ether functional groups (Fig. 1e). Neither group showed a distinct subclass preference (Fig. 1c).

Fig. 1.

Workflow and analysis of MutAIverse-generated DNA adducts. a Schematic diagram illustrating the MutAIverse workflow for generating synthetic DNA adducts using generative AI. Phase 1, phase 2, and enzyme commission (EC)-based biotransformations were performed with BioTransformer 3.0. Biotransformed molecules were screened for electrophilicity and oxidative damage. Positive candidates were used to create vocabulary and train Graph Neural Network (GNN) models, which were then employed to generate an in silico DNA adductome. GNN models were trained on both genotoxins and their biotransformation products. Predictions of in silico ESI + MS/MS spectra at collision energies of 10 eV (E0), 20 eV (E1), and 40 eV (E2) were generated using CFM ID 4.0 to build a spectral library. b Schematic representation of DNA adduct formation through electrophilic attack of genotoxins on DNA. Circos plot depicts the number of genotoxins sourced from the Tox21 database and literature, along with the distribution of ClassyFire chemical superclasses for genotoxic molecules. Genotoxins were classified into hierarchical chemical categories using the ClassyFire API. c Radar plot illustrating the proportion of biotransformed (BT) and non-biotransformed (NBT) genotoxins across different chemical superclasses. d Scatter plot displaying Principal Component Analysis (PCA) of the Extended Connectivity Fingerprints (ECFP6) for biotransformed (BT) and non-biotransformed (NBT) genotoxins. ECFP6 fingerprints were generated from canonical SMILES representations using RDKit. e Heatmap depicting the functional groups present in biotransformed (BT) and non-biotransformed (NBT) genotoxins. Datasets were converted to SDFset format and analyzed using ChemmineR to calculate functional group counts, which were then visualized in a heatmap. f Sunburst plot showing the counts of biotransformation products across iterations (k = 1-3). g Heatmap displaying the log-transformed frequencies of the top 10 biotransformation reactions across iterations (k = 1-3). h Schematic representation of the prediction process for electrophilicity and oxidative damage of biotransformed genotoxins using Metabokiller. Molecules with probabilities > 50% were classified as positive for these properties. Venn diagram shows the overlap of biotransformed genotoxin products classified as positive for electrophilicity (111,300), oxidative damage (9608), or both (161,022), which were used for GNN training. Molecules negative (44,420) for both properties were excluded. i Scatter plot displaying the frequency distribution of vocabulary substructures. The x-axis shows the atom counts of substructures. Vocabulary was built from common substructures present in biotransformed genotoxins, with a frequency threshold set to 10 occurrences. j Bar plot showing the absolute number of DNA adduct molecules generated for each seed deoxyribonucleoside (dA, dT, dG, and dC). k Ridge plot displaying the distribution of Synthetic Accessibility scores (SA score) for generated DNA adducts. The red line represents the median SA score for experimentally validated DNA adducts (EVA), and the black line represents the median SA score for DrugBank molecules. SA scores were computed using the pyTDC package. l Line plot showing the optimization of SA scores for generated DNA adduct molecules (at 0-biotransformation) for each seed (dA, dT, dG, dC) over generation numbers. MIMOSA was used to generate molecules by optimizing SA score, with a population size of 800. m Bar plot depicting the absolute numbers of DNA adduct molecules in MutAIverse categorized by seed molecule (dA, dT, dG, and dC) after filtering out invalid adducts. n Scatter plot illustrating the PCA projection of generated DNA adducts in bioactivity space, with bioactivity descriptors

To account for the possibility that metabolized genotoxin products may undergo additional rounds of biotransformation, given their accessibility to the cellular metabolism machinery, we performed iterative biotransformation simulations. In these simulations, the biotransformation cycle parameter, k, was varied from 0 (no biotransformation) to 3 (three rounds of biotransformation). These additional steps ensured that most, if not all, possibilities of genotoxin processing within the cell were considered. As expected, the number of metabolized products increased with higher k values, ranging from 25,274 (k = 1) to 217,434 (k = 3) (Fig. 1f). Further analysis revealed that the common biotransformations included hydroxylation, desaturation, reduction, oxidation, and epoxidation, among others (Fig. 1g).

Next, to identify reactive genotoxins and their metabolites, we used two previously published machine-learning models [22]. These models predict the potential of compounds or metabolites to induce oxidative stress or exhibit electrophilic properties. Screening the biotransformed genotoxin fragments (metabolized products) using these models revealed that 86.38% of them contribute to either oxidative stress or electrophilic properties (reactive metabolites) (Fig. 1h). Consequently, we selected the union of these reactive metabolites for downstream analysis under the assumption that both could equally contribute towards DNA adduct formation. Once compiled, filtered, and selected, these metabolized, potentially electrophilic, and oxidative stress-inducing genotoxin fragments were used to build a vocabulary for the downstream generation of DNA adducts. This vocabulary included abundant atoms like carbon, oxygen, chlorine, and nitrogen, as well as simple organic compounds like benzene (Fig. 1i; Supplementary Table 4). Finally, to generate a comprehensive DNA adduct library, we used the MultI-constraint MOlecule SAmpling (MIMOSA) framework [24]. MIMOSA leverages input molecules as initial guesses and samples new molecules from the target distribution, which, in our case, are reactive genotoxin metabolized products (electrophilic and capable of inducing oxidative stress). The framework pretrains two property-agnostic Graph Neural Networks (GNNs) for molecule topology and substructure-type prediction. Starting with DNA-specific deoxyribonucleosides: deoxyadenosine (dA), deoxycytosine (dC), deoxyguanine (dG), and deoxythymine (dT) as seed molecules, we employed the aforementioned metabolized genotoxin fragment-based vocabulary within the MIMOSA framework (Fig. 1j). The entire process of generating novel DNA adducts was iteratively performed using the metabolized genotoxin-generated vocabularies obtained from each biotransformation cycle parameter (k = 0, 1, 2, and 3). The impact of k-biotransformation on GNN model performance was measured by assessing the validation loss at various levels (k = 0-3) (Supplementary Fig. 1a). More precisely, the validation loss was largest in the case of k = 0, but it significantly decreased with biotransformation (k ≥ 1) (Supplementary Fig. 1a). The increase in k values resulted in a decrease in the variability of model performance, leading to increased learning of the GNN model, as suggested by the narrower confidence intervals (Supplementary Fig. 1a). The generated DNA adducts were synthesized with the optimization of the Synthetic Accessibility (SA score) metric, which assesses the feasibility of chemical synthesis. The generated adducts exhibited a broad spectrum of SA scores, reflecting the varied complexity and synthesizability of the molecules (Fig. 1kl and Supplementary Fig. 1b-d). This approach yielded 303,377 unique DNA adducts, with individual contributions from each deoxyribonucleoside as follows: 94,978 from dA, 65,928 from dT, 66,838 from dG, and 75,633 from dC (Fig. 1m). Clustering analysis of MutAIverse adducts in bioactivity and spectral space resulted in distinct clusters showing the diversity of MutAIverse-generated DNA adducts (Fig. 1n; Supplementary Fig. 1e). Finally, to confirm the specificity of our approach, we performed similar simulations with non-genotoxic compounds (k = 0 to 3) and compared the resulting fragments to those generated from genotoxins using machine learning. Our results suggest that with an increasing value of k, the AUC-ROC values also increased with the same hyperparameters, suggesting differences in the chemical compositions of fragments obtained for genotoxins and non-genotoxins (Supplementary Fig. 1f-i). In summary, by simulating the intracellular metabolism of genotoxins and leveraging a novel AI-driven framework, we have generated a comprehensive and chemically valid DNA adductome.

MutAIverse generates novel and known DNA adducts while preserving chemical properties

Recent efforts to compile a dedicated database of DNA adducts, accompanied by the construction of a standard reference MS/MS spectral library, have identified 279 experimentally validated (referred to as EVA, Experimentally Validated Adducts) and 303 chemically plausible DNA adducts (referred to as SA, Suspected Adducts) [15] (Fig. 2a; Supplementary Table 5). Next, we investigated whether the DNA adductome generated using the MutAIverse framework includes previously characterized EVA and SA DNA adducts. We first compared physicochemical properties (atom count, molecular weight, LogP, TPSA, and Kier-Hall Index [25]) between MutAIverse, EVA, and SA adducts (Fig. 2b). Overlapping distributions across these properties demonstrate that our mechanism-based approach captures the chemical space of known DNA adducts. We also compared topological features (κ1, κ2, κ3, and LabuteASA) between MutAIverse, EVA, and SA adducts. Similar results for these topological features further validate that our framework preserves key chemical and structural features of DNA adducts (Fig. 2c, Supplementary Fig. 2b). Furthermore, TMAP [26] clustering of MutAIverse adducts revealed distinct clusters based on properties like LogP, molecular weight, and ring count (Supplementary Fig. 2a). This analysis highlighted variations in hydrophobicity, size, and complexity, providing insights into the distribution and relationships among the adducts. To evaluate the level of structural diversity, we also performed Tanimoto similarity (TS) and Overlap coefficient analyses between the fingerprints computed for the MutAIverse adducts and those in the EVA and SA datasets. The maximum TS scores ranged from 0.4 to 1, indicating a considerable degree of structural similarity between the libraries (Fig. 2d, e; Supplementary Fig. 2f, g). In contrast, the minimum TS scores ranged from 0.0 to 0.3, reflecting a broad range of structural diversity among the chemically valid DNA adducts generated by MutAIverse. These analyses demonstrate that our framework captures a wide range of structural variations within the DNA adductome. Next, we conducted a comparative analysis of the DNA adducts from MutAIverse, EVA, and SA datasets at the spectral peak level. This involved generating in silico ESI (+) MS/MS spectral peaks for the DNA adducts precursor [M + H] + using the CFM-ID [27, 28] computational framework at three distinct collision energy settings, denoted as 10 eV (E0), 20 eV (E1), and 40 eV (E2), corresponding to different stages of energy application and the resulting fragmentation patterns. Subsequently, we employed the matchms [29] Python library to perform pairwise comparisons of these spectral peaks, utilizing the cosine similarity metric to quantify the degree of correspondence between MutAIverse and EVA or SA DNA adducts (Fig. 2f-h; Supplementary Table 6). In the comparison between MutAIverse-generated adducts and EVA, we observed a pronounced peak in the cosine similarity distribution, with the mean (95% confidence interval) ranging from 0.94 (0.93, 0.95) (Fig. 2g). Detailed analysis indicated that the highest cosine similarity values were recorded at collision energy level E0, followed by E1 and E2 (Fig. 2g; Supplementary Fig. 2c). A comparable trend was observed in the analysis of MutAIverse-generated adducts versus SA adducts (Fig. 2h; Supplementary Fig. 2d), confirming the robustness of our approach across different datasets and energy levels. Finally, we compared the distribution of spectral features of the DNA adducts present in MutAIverse, EVA, and SA at these three collision energy levels and observed comparable spectral entropies between the groups (Fig. 2i). Entropy similarity analysis revealed high entropy similarity for EVA and SA adducts compared to the MutAIverse library (Supplementary Fig. 2 h). Notably, in these analyses, several MutAIverse adducts with cosine similarity = 1 were detected during the comparative analysis with EVA (Fig. 2j). To ensure robust validation, predicted spectra were benchmarked against experimentally acquired MS/MS spectra [21], replacing earlier in silico‑in silico comparisons. The resulting cosine similarity distributions (Supplementary Fig. 2i) demonstrate realistic agreement consistent with experimental variability. In summary, a thorough analysis of MutAIverse-generated DNA adducts, compared with EVA and SA, demonstrated congruent distributions across the chemical, topological, and entropy metrics. This consistency underscores the fidelity of MutAIverse in replicating the chemical space and structural characteristics of known DNA adducts.

Fig. 2.

Comparative analysis of DNA adducts across MutAIverse, EVA, and SA. a Sunburst plot representing the absolute number of experimentally validated DNA adducts (EVA) and suspected DNA adducts (SA) classified by deoxyribonucleoside bases (dA, dG, dT, dC, dU, dI). The data include EVA (279 adducts) and SA (303 adducts) compiled from the DNA adductomics database. b Violin and boxplots showing the distribution of atom count, molecular weight, LogP (partition coefficient), Total Polar Surface Area (TPSA), and Hall-Kier alpha (connectivity index) across MutAIverse, EVA, and SA. Properties were calculated from SMILES representations using RDKit. c Violin and boxplots illustrating the distribution of κ1 (first order κ shape index), κ2 (second order κ shape index), κ3 (third order κ shape index), and LabuteASA (surface area) across MutAIverse, EVA, and SA. Properties were computed from SMILES representations using RDKit. d Boxplot showing the maximum, minimum, and median Tanimoto similarity scores between EVA and MutAIverse DNA adducts. e Boxplot showing the maximum, minimum, and median Tanimoto similarity scores between SA and MutAIverse DNA adducts. f Schematic representation of in silico ESI (+) MS/MS spectra prediction for EVA and SA. The MS/MS spectra of EVA and SA were compared against the MutAIverse spectral library, with cosine scores computed. g Density plot (left) showing the kernel density estimate of maximum cosine scores for MutAIverse DNA adducts when MS/MS spectra of EVA were searched against the MutAIverse spectral library. Heatmap displaying the maximum cosine scores across different collision energies (E0, E1, E2) for these searches. Cosine scores were computed using the matchms cosine greedy algorithm with a default m/z tolerance of 0.1 Da. h The density plot (left) depicts the kernel density estimate of maximum cosine scores for MutAIverse DNA adducts when MS/MS spectra of SA were searched against the MutAIverse spectral library. Heatmap displaying the maximum cosine scores across different collision energies (E0, E1, E2) for these searches. i Boxplot representing the distribution of spectral entropy across MutAIverse, EVA, and SA at collision energies E0, E1, and E2. Spectral entropy was calculated from normalized spectra using the Flash entropy package. j Examples of MutAIverse DNA adducts with a cosine score of 1 when compared with EVA

MutAIverse framework enables precise DNA adduct-genotoxin linking

While various analytical methods excel at identifying novel DNA adducts in biospecimens, linking these adducts to their parent genotoxins remains a significant challenge. To address this, we employed an in silico approach that models the reversal of genotoxin bioactivation and biotransformation within cells, thereby developing methods for retrograde tracing DNA adducts back to their potential parental genotoxins. We named this sub-module as AdductLinker due to its functional importance. AdductLinker employs a multi-step computational framework: it begins by isolating the nitrogenous base from the DNA adduct to extract the genotoxin-contributing fragment via a process called splicing (Fig. 3a, b). It then calculates the structural fingerprints of this fragment and all genotoxin intermediates and uses k-nearest neighbor (kNN) clustering to identify the nearest neighboring intermediate and their corresponding genotoxins (Fig. 3a). Finally, AdductLinker determines the maximum common substructure (MCS) among these nearest neighbors, which is used to trace back through the genotoxin-intermediates-metabolized fragment tree, thereby identifying potential parental genotoxins (Fig. 3a; Supplementary Fig. 3a). AdductLinker outputs the queried DNA adduct, its upstream metabolized products, and the parental genotoxin, along with a confidence score. While this approach significantly accelerates the process of linking DNA adducts to their parent genotoxins, it is limited by its reliance on the genotoxins used to construct MutAIverse and the prerequisite that not all DNA adducts undergo splicing (Fig. 3c-e; Supplementary Fig. 3b). We validated and tested AdductLinker using bona fide DNA adduct-genotoxin pairs, including AFB1-N7-dG-Aflatoxin, N2-DB[a,h]ADE-dG-Dibenzo[a,h]anthracene, N2-DB[a,l]PDE-dG-Dibenzo[a,l]pyrene, N6-(5-Me-CDE)-dG-5-methyl-chrysene, 2-AF-dG-2-Aminofluorene and Benzo[a]pyrene-dG-Benzo[a]pyrene (Fig. 3f; Supplementary Fig. 3f). Our results suggest that the AdductLinker sub-module of MutAIverse successfully backtracked both the upstream metabolized products and the parental genotoxins, demonstrating its efficacy in identifying potential genotoxin origins for a given DNA adduct.

Fig. 3.

Architecture and validation of the AdductLinker module for backtracing parent genotoxins. a Schematic diagram illustrating the AdductLinker module's architecture for backtracing the parent genotoxin of a DNA adduct. During preprocessing, the nitrogenous base is removed from the input DNA adduct using Maximum Common Substructure (MCS) analysis, retaining only the modified segment. ECFP6 fingerprints of the modified segment are computed. During backtracing, k-Nearest Neighbors (k-NN) of the fingerprints are searched within the MutAIverse object (a dictionary of phase 1 and phase 2 genotoxin reactions), and the MCS of the nearest neighbor is identified in phase 1 and phase 2 reaction data, leading to the identification of the parent genotoxin. The output includes structures of the DNA adduct, intermediate metabolite, parent genotoxin, and a confidence score (probability). b Stacked bar plot depicting the absolute number and proportion of phase 1 and phase 2 biotransformation products of genotoxins across k-biotransformation iterations (k = 1-3). c Donut plot representing the proportion of spliced and unspliced DNA adducts across four nitrogenous bases for EVA adducts. d Boxplot showing the distribution of atom count for the modified segments of spliced EVA DNA adducts across four nitrogenous bases. e t-SNE projection displaying the chemical space projection of modified segments from spliced DNA adducts. Morgan fingerprints (n = 2068, radius = 3) of these segments were computed using RDKit. f Backtracing examples demonstrating the identification of parent genotoxins for three specific DNA adducts: 1) 7AFB1-dG (AFB1-N7-dG), 2) N2-Dibenz[a,h]anthracene-diolepoxide-dG, and 3) N2-Dibenzo[a,l]pyrene-diol epoxide. g Schematic diagram showing the external validation of MutAIverse by searching MS/MS adductomics data from prostate tissues against the MutAIverse library for DNA adduct detection using the mapper module. h Boxplot illustrating the distribution of maximum cosine scores for EVA and MutAIverse DNA adducts found in prostate cancer tissue adductomics samples. i Backtracing example demonstrating the identification of the parent genotoxin for the EVA DNA adduct PhIP-dG and 7-2(Carboxy-ethyl)-dG detected in prostate cancer tissue adductomics samples. j Kernel density estimate depicting the maximum cosine scores for MutAIverse DNA adducts found in prostate cancer tissue adductomics samples, dotted lines indicate scores greater than the reported EVA (0.84). k Pie chart illustrating the number of EVA and MutAIverse adducts found in prostate tissue of prostate cancer patients above cosine score of 0.9. l Schematic diagram showing the external validation of MutAIverse by searching MS/MS adductomics data from rat liver tissue exposed to genotoxicants/carcinogens against the MutAIverse library for DNA adduct detection using the mapper module. m Boxplot showing the distribution of maximum cosine scores for EVA and MutAIverse DNA adducts found in rat liver tissue adductomics samples exposed to genotoxicants/carcinogens. n Backtracing example demonstrating the identification of the parent genotoxin for the EVA DNA adducts 8-(N'-MeIQx)-dG and N2-MeIQx-dG detected in rat liver tissue adductomics samples exposed to genotoxicants/carcinogens. o Kernel density estimate depicting the maximum cosine scores for MutAIverse DNA adducts found in rat liver tissue adductomics samples, dotted lines indicate scores greater than the reported EVA (0.95). p Bar plot depicting the number of EVA and MutAIverse adducts found in carcinogen-treated and control rat liver tissue samples above a cosine score of 0.9

To test this sub-module further in analyzing experimental MS/MS data from biospecimens, we utilized the recently published datasets from Walmsley and colleagues [30]. The dataset contained tumor-adjacent prostate tissue from 4 prostate cancer patients. In the original study, the authors identified 361 putative and six bona fide DNA adducts, including dG-C8-PhIP, N2-Dimethyldioxane-dG, 4-OH-2-heptenal-dG, (E)-4-hydroxy-2-nonenal-dC, (E)-4-oxonon-2-enal-dC, and M1AA-dA using a standard DNA adduct-specific MS library which contained 349 adducts of 279 unique masses [30]. We analyzed this dataset with the MutAIverse computational framework while selecting the matchms-based cosine similarity calculations (Fig. 3g). Methodologically, mzML files of MS/MS data was processed using the mapper module of MutAIverse. Spectral matching was carried out with the MutAIverse mapper module using the matchms cosine greedy algorithm (m/z tolerance 0.1 Da), complemented by approximate nearest neighbor search via Hierarchical Navigable Small World (HNSW) graph indexing. This approach successfully recapitulated multiple known and novel DNA adducts, including the ones reported in the original study. The six of the bona fide adducts identified were Phenylimidazo[4,5-b]pyridine-dG, N2-Dimethyldioxane-dG, E-4-Oxonon-2-enal-dC, (E)-4-Hydroxy-2-nonenal-dC, 3-(2′-Deoxy-β-D-ribofuranosyl)-7-methyl-8-formyl[2,1-i]pyrimidopurine-dA, and 4-Hydroxy-2-heptenal-dG. Notably, Phenylimidazo[4,5-b]pyridine-dG was a reported DNA adduct and was mapped to an experimentally validated adduct (EVA) of the MutAIverse with a high cosine similarity of 0.84. Furthermore, the AdductLinker sub-module mapped this adduct to parental genotoxin phenylimidazo[4,5-b]pyridine with confidence scores of 100% (Fig. 3i). Similarly, in the case of N2-Dimethyldioxane-dG, another EVA, which did not match with any known spectra in the MutAIverse library but was backtracked to multiple genotoxins such as (1R,6S)-7-oxabicyclo[4.1.0]heptane and 2-methyl-2-(oxiran-2-yl)oxirane with confidence scores of 87.5%. Similar analysis with other bona fide adducts led to their identification in the MutAIverse library with varied cosine similarities, as well as parental genotoxins. As stated above, we also observed multiple novel DNA adducts exhibiting a CS higher than that of PhIP (0.84) (Fig. 3h-k; Supplementary Fig. 3 g-j). Implementation of AdductLinker with these novel DNA adducts (CS > 0.9) suggested (2R,6R)-2,6-dimethylmorpholine, 2-[2-(1H-indol-3-yl)ethylamino]oxane-3,4,5-triol and 2-methylidenehexanal as top three potential parental genotoxins.

Next, we also used an independent, controlled rodent-based dataset [30]. In this dataset, rats were exposed to a cocktail of genotoxic mutagens (B[a]P, 4-ABP, PhIP, AαC, and MeIQx), and post-exposure, DNA adductomics was performed on genomic DNA from liver tissue using wSIM-MS/MS mass spectrometry (Fig. 3l). Next, we mapped the experimentally elucidated MS/MS peaks onto a standard DNA adduct library containing EVA and compared it to the subset of the MutAIverse library containing only in silico-generated adducts (Fig. 3m; Supplementary Fig. 3 k-m). Comparison to EVA library reproduced six bona fide DNA adducts: phenylimidazo[4,5-b]pyridine-dG, dG-C8-2-Amino-9H-pyrido[2,3-b]indole, dG-C8-3,8-dimethylimidazo[4,5-f]quinoxalin-2-amine, dG-C8-4-aminobiphenyl, N-(2′-deoxyadenosin-8yl)-4-aminobiphenyl, N-(2′deoxyguanosin-N2-yl)-4-aminobiphenyl as mentioned in the original study. Notably, Phenylimidazo[4,5-b]pyridine-dG, an EVA, demonstrated a cosine similarity of 0.76 and was backtracked to phenylimidazo[4,5-b]pyridine by AdductLinker sub-module as a potential parental genotoxin with confidence scores of 100 (Fig. 3i). The adduct dG-C8-2-Amino-9H-pyrido[2,3-b]indole showed a cosine similarity of 0.90 and was matched to the genotoxin 2-Amino-9H-pyrido[2,3-b]indole (AαC) with confidence scores of 100 (Fig. 3n). A similar analysis with other bona fide adducts resulted in their identification within the MutAIverse library, exhibiting varied cosine similarities and corresponding parental genotoxins. Similar to the prostate cancer dataset, we also observed multiple novel DNA adducts when mapping with the MutAIverse library (Fig. 3o, p). Taken together, all these results collectively underscore the robustness, comprehensiveness, and accuracy of the MutAIverse workflow in effectively identifying and validating known and novel DNA adducts within complex MS/MS datasets.

MutAIverse identified novel DNA adducts in smokeless tobacco-exposed head and neck cancer

Identifying novel DNA adducts with MutAIverse advances our understanding of adduct formation and holds promise for clinical translation. This could facilitate the development of new diagnostic biomarkers and therapies. As proof of concept, we analyzed head and neck tumor biopsies and adjacent non-malignant tissues from patients with a history of smokeless tobacco (SLT) use (Fig. 4a). The study included a balanced representation of both sexes (Fig. 4b; Supplementary Table 7). We performed high-resolution MS-based DNA adductomics on genomic DNA from these samples (refer to data availability section for HRMS data), comparing results between tumor and adjacent non-malignant tissues (Fig. 4a). We identified a greater number of DNA adducts with higher cosine similarity (CS) using the MutAIverse library compared to the EVA library (Fig. 4c). This suggests that MutAIverse covers a broader chemical space of DNA adducts. Even with a stringent CS threshold (> 0.9), MutAIverse identified significantly more potential adducts (Fig. 4c; Supplementary Fig. 4a,b). This trend was consistent across both tumor and adjacent normal tissues (Fig. 4c). Given the patients' clinical histories of smokeless tobacco use, we employed the AdductLinker module of MutAIverse to unveil the potential parent genotoxins or carcinogens. By projecting the tumor-specific DNA adducts identified through MutAIverse and performing retrograde tracking with the AdductLinker sub-module, we pinpointed several notable adduct-genotoxin pairs, such as 8-(N’-4-Chloroaniline)-dA-4-chlorobenzene-1,2-diamine and N6-Dibenz[a,h]anthracene-diolepoxide-dA, and 3,4-dihydronaphtho[1,2-b]phenanthrene-3,4-diol. Other alkylated DNA adducts such as 3-methyl-dA, N6-methyl-dA, and 3-ethyl-dA, along with etheno adducts 3,N4-etheno-dC were also detected. These findings suggest that the genotoxic agents identified are likely implicated in malignancy, as the adducts 3-methyl-dA and N6-Methyl-dA were specific to tumor tissues and detected in 18.51% of the samples (Fig. 4d). The spectral search of EVA against the MutAIverse library revealed a substantial degree of similarity (Fig. 4e), with an average cosine similarity of 0.94. MutAIverse DNA adducts were detected across both tumor and tumor-adjacent tissues (referred to as normal), while 31 MutAIverse DNA adducts were exclusive to tumor tissue (Fig. 4f, g). Notably, the structures shown are MutAIverse candidate structures, chemically plausible adducts generated in silico that matched the experimental spectra with cosine similarity ≥ 0.9. These candidates require orthogonal experimental confirmation. Chemical similarity analysis revealed the structural divergence between EVA and adducts created by MutAIverse (Fig. 4h, i), indicating that MutAIverse covers a wide range of chemical variations between EVA and MutAIverse adducts detected in tumor and tumor-adjacent samples. The AdductLinker predicted Ketene as the parental genotoxin of the MutAIverse adduct with m/z 266.0739 with a confidence score of 100%. The observed neutral loss of 114 Da from the precursor m/z 266 to the fragment m/z 152 corresponds to a modified deoxyribose that undergoes charge‑directed rearrangement during CID, losing a neutral fragment of C₄H₆O₄ rather than the canonical intact 2′‑deoxyribose loss of 116 Da. This process yields a resonance‑stabilized pyrimidinium ion at m/z 152. Backtracing with AdductLinker, the MutAIverse adduct with m/z 286.1033 to hydroperoxyethane, and the MutAIverse adduct (m/z 546.2135) with Benzo[c]phenanthrene 3,4-dihydrodiol-1,2 epoxide (B(c)Phde), a well-known tobacco-related carcinogen [31] (Fig. 4j). The intensity of both validated adducts with m/z 266.0739 (MutAIverse adduct) and m/z 377.1868 (N4-Pyridyl-oxobutyl-dC) was detected in both tumor and tumor-adjacent cells (Supplementary Fig. 4c,d). Correlation analysis of the intensity of m/z 266.07 with the age of patients across tumor and tumor-adjacent normal cells showed a weak positive correlation, whereas the intensity of m/z 377.1868 showed a negative correlation in the case of tumor-adjacent normal cells (Supplementary Fig. 4c). A previous study involved urinary DNA adductomics analysis using high-resolution mass spectrometry on smokeless tobacco users, non-users, and head and neck cancer patients [32]. The method was validated with synthetic DNA adduct standards, including NDAMPP, 6O-MG, 8-NG, 7-MG, 3-MG, and 8-OHdG, by matching retention times in the samples. Additional DNA adducts were identified using a database library. Statistically significant DNA adducts were confirmed using LC-ESI-QTOF-MS/MS fragmentation for detailed structural analysis. To eliminate the possibility of detecting deoxyribonucleosides in the samples, an in silico MS/MS analysis of deoxyadenosine (dA), deoxythymidine (dT), deoxyguanosine (dG), and deoxycytidine (dC) was performed and compared with MutAIverse library. The comparative analysis revealed that MutAIverse adducts exhibited higher cosine similarity than the dA, dT, dC, and dG, further supporting the specificity of the identified adducts (Supplementary Fig. 4f).

Fig. 4.

Screening and analysis of head and neck cancer DNA adductomics samples from SLT-exposed patients using the MutAIverse spectral library. a Schematic illustration showing the screening of adductomics samples from head and neck cancer tissue of patients exposed to smokeless tobacco (SLT) against the MutAIverse spectral library using the mapper module. b Bar plot on the left displays the absolute counts of male and female head and neck cancer patients. Boxplot on the right depicts age distribution across male and female patients. Donut plot showing the absolute counts of tumor samples and normal samples extracted adjacent to tumor tissue. Scatter plot representing the PCA projection of tumor and normal samples, with PCA analysis conducted on raw metabolomics samples using MetaboAnalyst 6.0. c Boxplot illustrating the distribution of maximum cosine scores for EVA and MutAIverse DNA adducts in adductomics samples from SLT-exposed head and neck cancer patients. d Venn diagram highlighting the EVA DNA adducts found in tumor and normal samples, with a cosine score threshold of  > 0.9. Structures of 3-methyl-dA and N⁶ methyl-dA EVA adducts found exclusively in tumor tissue of head and neck cancer patients. e Stemplot illustrates the maximum cosine score of detected EVA adducts in the MutAIverse library. f Venn diagram showing the number of MutAIverse adducts found in tumor and normal samples, with a cosine score threshold of > 0.9. g Structures of a few MutAIverse adducts were detected exclusively in tumor tissue and in both tumor and tumor-adjacent samples. h Scatter plot illustrating the PCA projection of tumor-associated EVA and MutAIverse adducts in chemical space. i Heatmap showing the pairwise tanimoto similarity comparison between the tumor associated EVA and MutAIverse adducts. j Backtracing of MutAIverse DNA adducts m/z 266.0739, m/z 286.1033 and m/z 546.2135 via adduct linker. k Flow diagram showing precursor m/z search of MutAIverse DNA adduct hits (cosine score > 0.9) in adductomics samples of SLT-exposed head and neck cancer tissue, followed by MS/MS fragmentation of the detected precursor m/z. l MS spectra of precursor m/z ion [M + H] + of m/z 377.1868 and 266.0763 obtained by HRMS (error < 10 ppm) m Experimental MS/MS spectra of the detected tobacco exposure-associated EVA DNA adduct with m/z 377.1868 (error < 10 ppm) (green) and the MutAIverse DNA adduct with m/z 266.0739 (error < 10 ppm) (blue). n Predicted fragmentation pattern of the detected EVA (green) and MutAIverse (blue) DNA adducts based on the experimental MS/MS spectra

Although MutAIverse offers a comprehensive collection of genotoxin-guided and chemically valid DNA adducts, some of these adducts, while computationally and conceptually plausible, may not always be present in real-world biological scenarios. To validate the existence of novel DNA adducts identified by MutAIverse in these samples, we conducted rigorous, high-resolution MS analysis on two DNA adducts, designated N4-Pyridyl-oxobutyl-dC (N4-POB-dC) with m/z 377.1824 which is formed by the exposure of tobacco-specific nitrosamines [11] and a novel DNA adduct with m/z 266.0739 and MS/MS analysis of these compounds were done using low-resolution mass spectrometry (Fig. 4k-m). The experimentally observed precursor ion at m/z 266.0739 is in agreement with the in silico CFM-ID predicted value of m/z 266.0771. Some of the predicted fragments have mass error > 10 ppm due to their low abundance (Fig. 4n). Detailed examination of the MS/MS spectra confirmed the presence of the novel DNA adduct in 14.81% of tumor samples and 27.27% of tumor-adjacent control tissues. This validation not only demonstrates the efficacy of MutAIverse in uncovering novel DNA adducts but also underscores its potential as a tool for biomonitoring clinically relevant, adducts-based biomarkers associated with smokeless tobacco-related head and neck cancers. Notably, spectral mirror plots comparing the experimental MS/MS spectra against the predicted spectra for these novel adducts are provided in Supplementary Fig. 4 g, further substantiating the annotations.

Discussion

The identification and characterization of DNA adducts are crucial for understanding their roles in carcinogenesis and mutagenesis. Traditional methods, reliant on experimentally validated data, are limited to cataloging fewer than 400 DNA adducts, which restricts the exploration of their full chemical diversity. Computational approaches, particularly those utilizing generative AI frameworks, offer promising solutions for broadening the scope of DNA adduct identification. In this study, we introduce MutAIverse, a novel mechanism-driven generative AI framework designed to (a) generate all possible chemically valid DNA adducts from experimentally validated genotoxins, (b) provide a comprehensive MS spectral library and associated toolkits for mapping DNA adductomics datasets, and (c) support the computational prediction of parental genotoxins from detected DNA adducts. Unlike traditional methods that rely on limited experimentally validated adducts, MutAIverse employs a novel simulation-based approach to generate a more extensive array of DNA adducts.

The chemical diversity of DNA adducts generated by MutAIverse is contingent upon the selection of genotoxins used in the initial setup. Although our study incorporates an updated list of experimentally validated genotoxins, the framework is adaptable and can accommodate newly identified genotoxins as they become available. Incorporating novel genotoxins could further expand the DNA adduct library and enhance the linkage of newly generated DNA adducts to their respective parental genotoxins. Additionally, the selection of biotransformation prediction tools plays a critical role in determining the chemical diversity of the DNA adducts generated by MutAIverse. We employed BioTransformer 3.0 [23], a sophisticated tool that integrates both phase 1 and phase 2 metabolism predictions with EC-based metabolism and machine learning techniques. This choice contrasts with other tools such as SyGMa [33], which relies on predefined reaction data and rule-based methods, or GLORYx [34], which combines rule-based approaches with machine learning. Given that genotoxins can contribute to adduct formation either directly or following biotransformation, our approach included multiple iterations of biotransformation simulations to capture a comprehensive range of possible adducts. To prioritize metabolized genotoxin fragments based on their potential for adduct formation, we utilized predictive models for electrophilicity and oxidative stress. However, additional factors could be considered to refine the adduct library further, including other reactivity types, lipophilicity, intracellular stability of genotoxin or its metabolized fragments, bioavailability, and the stability of formed adducts. Integrating these parameters could enhance the accuracy and relevance of the DNA adductome generated by MutAIverse.

MutAIverse, while offering significant advancements in DNA adduct identification, also has several limitations that must be acknowledged. Firstly, the framework’s reliance on an initially selected set of genotoxins may result in an incomplete representation of both the number and chemical diversity of generated DNA adducts. Specifically, many chemical compounds that are not genotoxic in their native state can become genotoxins through bioactivation by cellular enzymes. Currently, MutAIverse does not model this process, which limits its ability to predict adducts formed from such transformed compounds. Integrating a genotoxicity prediction model upstream could address this issue, but it may also impact the subsequent performance and usability of the AdductLinker sub-module. Secondly, the biotransformation predictions incorporated in MutAIverse are predominantly based on human-specific metabolic pathways. This approach does not account for biotransformations that occur due to the influence of the microbiome, which can significantly affect the formation and diversity of DNA adducts, as per recent reports [35, 36]. Incorporating microbiome-based biotransformation predictions could substantially broaden the scope and chemical diversity of the generated DNA adduct library. Thirdly, the current computational framework for generating DNA adducts has been optimized using the Synthetic Accessibility (SA score) score alone. While this parameter is useful, the addition of other metrics, such as LogP (partition coefficient) and other relevant chemical properties, could further enhance the accuracy and breadth of the DNA adduct library. Parallel optimization using these additional parameters could yield a more comprehensive representation of potential DNA adducts. Fourthly, we utilized CFM-ID [27, 28] for the in silico generation of ESI MS/MS spectral libraries. As a predictive tool, CFM-ID may introduce certain biases, which could affect the accuracy of the mapping algorithms and the quality of the downstream results. Lastly, although MutAIverse facilitates the identification of novel DNA adducts, their functional relevance remains to be determined. It is essential to assess whether these adducts act as transcription or replication blockers or how they contribute to mutagenesis. These levels of predictions are not presently offered by MutAIverse. All these potential biases underscore the need for further refinement and validation of the predictive models used in MutAIverse. Addressing these limitations by integrating additional models and optimizing parameters could enhance the capabilities and applications of MutAIverse, providing a more robust tool for DNA adductomics research.

Despite these limitations, MutAIverse presents several notable advancements and stands as the first proof of concept for employing a generative AI and metabolism simulation-based strategy to generate a comprehensive array of potential DNA adducts from selected genotoxins. The framework's innovative approach marks a significant departure from traditional methods and demonstrates the potential of integrating AI with biochemical simulations. It is strongly recommended to use MutAIverse in conjunction with established MS spectral libraries, such as those provided by the National Institute of Standards and Technology (NIST) [37], The Human Metabolome Database (HMDB) [17], METLIN [18], MassBank [19], or the Kyoto Encyclopedia of Genes and Genomes (KEGG) [20]. Additionally, adduct-specific libraries, including DNA Adduct Portal[21] and The DNA Adduct Database [15], should be considered. Integrating these spectral libraries with MutAIverse could greatly enhance the accuracy and comprehensiveness of DNA adduct identification.

Methods

Datasets compilation

A comprehensive dataset of genotoxins, consisting of 3,288 compounds, was assembled from multiple sources, including peer-reviewed literature and established databases such as Tox21 [38]. For each genotoxin, Simplified Molecular-Input Line-Entry System (SMILES) notations were retrieved through the PubChem Identifier Exchange Service. The complete list of genotoxins, along with their sources and chemical classifications, is provided in Supplementary Table 1 and 3. Notably, non‑genotoxins used for the comparative analysis in Supplementary Fig. 1f-i were curated from Mittal et al. 2022 [22]. To ensure consistency and accuracy in molecular representation, the retrieved SMILES notations were standardized to canonical SMILES using OpenBabel version 3.1.1 [39]. This standardization process is crucial for maintaining uniformity in subsequent computational analyses. For the DNA adducts, data were sourced from the DNA adductomics database [15], which includes a curated collection of 279 experimentally validated DNA adducts (EVA) and 303 suspected DNA adducts (SA). Similar standardization procedures for chemical representation were applied to these adducts to ensure consistency and comparability in the analysis.

Chemical taxonomy classification

ClassyFire [40] was used to analyze the chemical diversity of genotoxin molecules through a structured classification approach. Initially, SMILES notations for the genotoxins were converted into Structure Data File (SDF) format using OpenBabel to ensure uniformity in data representation. The SDF files were then processed by ClassyFire, a web server that classifies chemical entities based on a comprehensive chemical ontology. The classification process involved several stages: (1) preprocessing to ensure data consistency and readiness, (2) feature extraction to identify key structural attributes, (3) rule-based category assignment to classify molecules into appropriate taxonomic categories, and (4) selection of the direct parent category to provide hierarchical context. The output of ClassyFire was generated in SDF format, facilitating subsequent analysis and integration of the classification results into the broader dataset. This method enabled a thorough exploration of the chemical diversity among the genotoxins.

In silico biotransformation of genotoxins

The compiled genotoxin molecules were subjected to in silico biotransformation using BioTransformer 3.0 [23] to predict phase 1 (P1), phase 2 (P2), and enzyme commission (EC)-based metabolism. BioTransformer employs a hybrid approach that integrates rule-based and machine-learning methodologies to predict the metabolic pathways of small molecules. For each genotoxin, predictions were made for P1, P2, and EC-based pathways independently across three iterations. The ‘cypmode’ parameter was set to "combined," which incorporates both rule-based and machine-learning techniques. The iterative biotransformation process involves applying the selected in silico reactions for a specified number of cycles; for example, products from an initial reaction serve as substrates for subsequent biotransformations in two iterations. SMILES strings  of the genotoxins were used as input for the BioTransformer, and the results were generated in SDF format, which facilitated further analysis and integration of the predicted biotransformation products into the dataset.

Systematic characterization of genotoxins

The systematic characterization of genotoxins within chemical space was performed to elucidate their chemical properties. Extended Connectivity Fingerprints (ECFP6) [41] were computed for both biotransformed and non-biotransformed genotoxin groups using the RDKit library [42]. The fingerprints had a dimensionality of 2048 with a radius of 3. Principal Component Analysis (PCA) was then conducted to reduce the dimensionality of the dataset, utilizing the decomposition. PCA function from the Scikit-learn library [43] to extract three principal components. The PCA results were visualized using the Matplotlib library, which facilitated the identification of clustering patterns and relationships within the data. This visualization provided insights into the distribution of genotoxins in chemical space, highlighting both biotransformed and non-biotransformed compounds and offering a deeper understanding of their chemical properties and potential interactions.

Functional group analysis

To identify and categorize the predominant functional groups within biotransformed and non-biotransformed genotoxins and to enhance our understanding of their chemical composition, an in-depth functional group analysis was performed using the ChemmineR library [44] (v3.46.0). Initially, the genotoxin datasets were converted to the SDFset format, which facilitated the extraction of functional group counts into a matrix. The functional group data were then scaled and visualized using heatmaps to illustrate the distribution and relative abundance of these groups across the datasets. This comprehensive analysis provided detailed insights into the chemical diversity and functional group composition of the genotoxins, offering a deeper understanding of their biochemical properties and interactions.

Electrophilicity and oxidative stress prediction

Electrophilic potential and oxidative damage are key indicators of genotoxicity. Electrophilic molecules are more likely to attack nucleophilic sites on DNA, leading to DNA adduct formation. To evaluate these properties in biotransformed genotoxic metabolites, we employed the electrophilicity and oxidative stress prediction models within the Metabokiller package [22]. Metabokiller is an ensemble carcinogenicity prediction tool that quantitatively assesses carcinogenic risk by evaluating various factors, including electrophilicity, cell proliferation potential, oxidative stress, genomic instability, epigenetic alterations, and anti-apoptotic responses. The electrophilicity and oxidative stress models in Metabokiller utilize a multilayer perceptron algorithm. For this analysis, we set the probability threshold for both electrophilicity and oxidative stress at 0.5. Canonical SMILES representations of biotransformed genotoxin products were used as input for Metabokiller, allowing us to systematically predict and categorize their electrophilic and oxidative stress potential. This approach provided valuable insights into the genotoxic properties and potential biological impacts of the metabolites.

De novo generation of DNA adductome using MIMOSA

To facilitate the de novo generation of a DNA adductome from genotoxins and their biotransformed products (iterations k = 1-3), we employed the Graph Neural Network (GNN) framework of MIMOSA [24] (Multi-constraint Molecule Sampling for Molecule Optimization). The initial step in MIMOSA involves generating a vocabulary of molecular substructures, including atoms and rings, from the input molecules. This vocabulary was created by identifying frequent substructures present in both the original genotoxins (iteration k = 0) and their biotransformed products (iterations k = 1-3). Substructures appearing in more than ten molecules were included in the vocabulary. For molecular dataset preparation, SMILES notations were loaded from a text file and divided into training (90%) and validation (10%) sets. Graph Convolutional Network (GCN) model was constructed with 50 input features per node, 100 hidden units, and three graph convolutional layers. The training procedure involved five epochs, with validation checks performed every 5000 iterations. The trained model was saved, and the entire process was executed on a CPU. MIMOSA employs three fundamental molecular operations such as add, replace, and delete, within a Monte Carlo Markov Chain sampling optimization framework. To generate synthetic DNA adducts, deoxyribonucleosides (dA, dT, dG, and dC) were used as seed molecules. The generation process involved a population size of 800 across 1200 generations, enabling a thorough exploration of potential DNA adducts derived from both genotoxins and their biotransformed products. Generated DNA adduct structures were evaluated for chemical validity using RDKit [42] by converting SMILES representations into molecular objects. Structures that failed SMILES-to-molecule conversion or exhibited valence errors were considered chemically invalid and subsequently removed. The remaining chemically valid structures were retained for further analysis.

Bioactivity assessment of in silico-generated DNA adducts

To evaluate the bioactivity of in silico-generated DNA adducts, we computed bioactivity descriptors using the Signaturizer package [45] version 1.1.1. Principal Component Analysis (PCA) was then performed using the decomposition. PCA function from the Scikit-learn library [43] to reduce dimensionality. We focused on three principal components to simplify the analysis. The resulting PCA coordinates were visualized with Matplotlib to assess clustering patterns within the bioactivity space of the generated adducts. This analysis provided insights into the bioactivity profiles of the DNA adducts, highlighting distinct clusters that correspond to varying levels of bioactivity.

Prediction of ESI MS/MS spectra for DNA adducts

To predict electrospray ionization (ESI) (+) MS/MS spectra for generated DNA adducts, including precursor [M + H] + ions, Experimentally Validated Adducts (EVA), and Suspected Adducts (SA), we utilized the CFMID 4.0 Docker image [27, 28]. Predictions were made across collision energies of 10 eV (E0), 20 eV (E1), and 40 eV (E2) using standard parameters provided by the tool. CFMID 4.0 employs a two-step sequential fragment transition model to predict MS/MS spectra, incorporating ring cleavage modeling to improve feature consistency. The tool generates fragmentation graphs where ion fragments are represented as child nodes while neutral fragments are excluded. An adjacency matrix is used to depict molecular topology, and heuristic labeling is applied to fragments. CFMID 4.0 integrates SMIRKS-based fragmentation rules for modular compounds, utilizing a hybrid approach that combines Mass Spectra Rule-Based (MSRB) and machine learning-based MS/MS predictor (MSML) methodologies. This approach enhances the accuracy and efficiency of spectral predictions, covering a broad spectrum of chemical groups and adduct forms. Moreover, the predicted spectra were validated against the experimental MS/MS library of DNA adducts reported by Walmsley et al. (2024) [21] to obtain realistic similarity metrics.

Characterization of physicochemical and topological properties of DNA adducts

The physicochemical and topological properties of DNA adducts generated by MutAIverse, along with EVA and SA, were systematically characterized using a comprehensive suite of molecular descriptors. Key properties analyzed included atom count, molecular weight, LogP (octanol-water partition coefficient), topological polar surface area (TPSA), Kier-Hall index, κ indices (κ 1, κ 2, and κ 3), and Labute’s Approximate Surface Area (ASA). These descriptors were computed using the RDKit library [42], providing detailed insights into the molecular features and behaviors of the adducts. This thorough characterization facilitates an in-depth analysis of their chemical properties and potential biological interactions.

Validation of MutAIverse-generated DNA adducts

To validate the DNA adducts generated by MutAIverse, derived exclusively from genotoxins and their biotransformed products, we evaluated their similarity to EVA and SA. This was achieved by comparing the MS/MS spectra of EVA and SA against the MutAIverse spectral library. Cosine similarity scores were calculated using the matchms [29] cosine greedy algorithm, with an m/z tolerance set to 0.1 Da. This analysis assessed the congruence between the generated DNA adducts and those experimentally observed, thereby confirming the robustness and accuracy of the MutAIverse framework.

Spectral complexity analysis

A comparative analysis of the spectral complexity of MutAIverse-generated DNA adducts, EVA, and SA was conducted using spectral entropy. Spectral entropy quantifies the information content of a spectrum, reflecting its complexity as described by Shannon entropy in information theory. It is computed based on the number of peaks weighted by their intensity, with typical values ranging from 0 to 5 for small molecule spectra [46]. Spectral entropy analysis was performed using the ms-entropy package [47], version 1.1.3. The MS spectra intensities were normalized such that their sum equaled 1, utilizing the clean spectrum parameter. Entropy similarity was calculated between the MS/MS spectra of EVA and MutAIverse, as well as between SA and MutAIverse, with a default tolerance parameter of 0.05 Da. For each query spectrum from EVA and SA, the maximum entropy similarity was recorded. This approach provided a detailed comparison of spectral properties, validating the fidelity of MutAIverse in accurately representing experimentally observed DNA adduct spectra.

Spectral matching with the mapper module

The mapper module is designed to facilitate the spectral matching of MS/MS data from adductomics samples against the MutAIverse spectral library, as well as the EVA and SA libraries. It employs two complementary approaches for assessing spectra similarity: peak matching combined with cosine similarity and Approximate Nearest Neighbour (ANN) search using Hierarchical Navigable Small World (HNSW) graph indexing [48]. In the first approach, spectral data preprocessing involves normalizing the raw spectra from both the adductomics file and the spectral libraries. Following normalization, the query spectra are compared to the library spectra through peak matching and cosine similarity [29] assessments. This method identifies potential matches by evaluating the similarity of spectral peaks. The second approach involves encoding both the query and library spectra into spectral embeddings using a pre-trained word2vec model from FastEI [49]. These embeddings are indexed in the HNSW graph, which enables efficient vector search. The vector search identifies the nearest neighbor or most similar spectrum to the query spectra within the spectral embeddings. The output of the spectral matching process includes the top nearest neighbor result, along with density plots and histogram visualizations. This integrated framework combines traditional peak matching with cosine similarity and advanced vector-search techniques, ensuring both rapid and accurate identification of spectral matches (Supplementary Fig. 2e).

Visualization using TMAP

TMAP (Tree-based Minimum Spanning Tree Analysis) provides an advanced approach for visualizing high-dimensional datasets by mapping them into a two-dimensional minimum-spanning tree. This method is noted for its superior preservation of both local and global structures compared to traditional dimensionality reduction techniques like t-SNE and UMAP [26]. TMAP is particularly effective for analyzing extensive chemical datasets, offering clarity and precision in visualization.

In our analysis, MinHash fingerprints (MHFP6, with a dimensionality of 2068) [50] were computed from the SMILES representations of MutAIverse DNA adducts using the MHFP encoder package. These fingerprints encapsulate the chemical features of the adducts. The TMAP visualization was then created using the tmap package, highlighting various chemical properties such as LogP (octanol-water partition coefficient), molecular weight, and ring counts. This approach facilitated a comprehensive and interpretable analysis of the chemical dataset, revealing insights into the distribution and relationships of the DNA adducts.

AdductLinker module for parent genotoxin identification

The AdductLinker module is designed to trace the parent genotoxin molecule associated with a detected DNA adduct. The process begins with preprocessing, where the modified segment of the DNA adduct is isolated by extracting the nitrogenous base through maximum common substructure analysis. This step focuses exclusively on the modified portion of the deoxyribonucleoside. After preprocessing, the module generates Extended Connectivity Fingerprints (ECFP6) [41] for the extracted modified segment. During the backtracing phase, the k-Nearest Neighbors (k-NN) search algorithm is employed to identify fingerprints in the MutAIverse object that are similar to those of the modified segment. The MutAIverse object contains comprehensive storage of phase 1 and phase 2 reaction data for a wide range of genotoxins. The module then determines the Maximum Common Substructure (MCS) [51] of the nearest neighbor found in the phase 1 and phase 2 reaction data. The MCS is used to retrieve the parent genotoxin molecule from the database. The final output includes detailed information such as the structures of the identified DNA adduct and intermediate metabolite, as well as the parent genotoxin molecule. Additionally, a confidence score is provided, reflecting the probability or degree of confidence in the association between the DNA adduct and its genotoxic parent. This robust backtracing capability offers valuable insights into the biochemical pathways and interactions that lead to DNA adduct formation.

MutAIverse Python package

The MutAIverse package offers a comprehensive solution for adductomics research by integrating several key functionalities. It enables the screening of spectral data from adductomics samples against spectral libraries of EVA, SA, and MutAIverse-generated DNA adducts through its mapper module. Following this, the package facilitates the backtracing of parent genotoxins using its AdductLinker module. This integrated approach streamlines the process of identifying and characterizing DNA adducts and their parent genotoxins, providing a robust toolset for adductomics research.

External validation of MutAIverse

External validation of MutAIverse was performed using wSIM-MS/MS (wide selected ion monitoring) data from tumor-adjacent prostate tissue samples of prostate cancer patients (n = 4) (RTPS- 09-T1: patient 1, RTPS- 30-T1: patient 2, RTPS-38-T2: patient 3, and RTPS-41-T2: patient 4) and rat liver tissue samples (control = 3, treated = 3) exposed to carcinogens/genotoxins, including Benzo[a]pyrene (B[a]P), 4-aminobiphenyl (4-ABP), 2-amino-1-methyl-6-phenylimidazo[4,5-b]pyridine (PhIP), 9H-pyrido[2,3-b]indol-2-amine (AαC), and 3,8-dimethylimidazo[4,5-f]quinoxalin-2-amine (MeIQx), as described in the study by Walmsley et al. [30]. Spectral data were processed using the MutAIverse mapper module. Cosine similarity scores were calculated with the matchms cosine greedy algorithm (m/z tolerance 0.1 Da). In parallel, an approximate nearest neighbor search was performed using HNSW graph indexing on spectral embeddings generated by a pre‑trained word2vec model (FastEI). The spectral data from these samples were projected against the MutAIverse, EVA, and SA spectral libraries using the mapper module of the MutAIverse package. Cosine similarity scores were computed to assess the match between experimental spectra and library spectra. The AdductLinker module was then used to trace back the parent genotoxins associated with the detected DNA adducts from the EVA, SA, and MutAIverse-generated adduct libraries. This process validated the MutAIverse framework's accuracy and robustness in identifying and characterizing DNA adducts.

Optimization of DNA isolation and hydrolysis from HNC tissue samples

Tissue biopsy samples were collected from head and neck cancer (HNC) patients with a history of smokeless tobacco consumption, including tumor tissue (n = 27) and tumor-adjacent control tissue (n = 11). Following genomic DNA isolation and enzymatic hydrolysis, the samples underwent untargeted metabolomics analysis. DNA was isolated as given in the protocol of the manufacturer (Nucleospin tissue XS). After the DNA was isolated, purity was checked using a Nanodrop instrument. 1 μg human HNC tissue DNA (in 30 μl water) was digested. In brief, the DNA sample was denatured at 100 °C for 3 min. The sample was cooled rapidly on ice. Later, 3 μL of ammonium acetate (0.1 mol/L, pH 5.3) was added, followed by two units of nuclease P1. The resultant mixture was incubated at 45 °C for 2 h. After, the pH of the mixture was adjusted by adding 4.1 µL of freshly prepared ammonium bicarbonate (1 mol/L), and 0.002 units of venom phosphodiesterase were added. The mixture was incubated at 37 °C for 2 h. After incubating with 0.5 unit of alkaline phosphatase at 37 °C for 1 h, the resultant mixture was evaporated and reconstituted with 40 µL of methanol: water (1:1) and injected into LC-ESI-QTOF-MS for the analysis of DNA adducts. Notably, the analytical framework used for tissue analysis was first validated in our prior work [32], where the same DNA adducts were detected and structurally characterized in urine from smokeless‑tobacco‑treated rats using an Agilent 6495C triple quadrupole mass spectrometer and subsequently in human urine samples using an Agilent 6545 QTOF. Tumor‑adjacent tissue was used as a negative control in the present study.

LC-ESI-QTOF-MS analysis of DNA adducts

The hydrolyzed DNA samples were subjected to liquid chromatography (Agilent 1290 Infinity II HPLC) coupled with high-resolution mass spectrometry (Agilent 6545 LC-ESI-QTOF) in positive ion mode with Agilent dual jet stream electrospray ionization for untargeted determination of DNA adducts in the HNC tissue samples. The mass parameters were set as follows: the source temperature and drying gas flow were maintained at 320 °C and 8 L/min, respectively. The nebulizer pressure was set to 35 psi. The sheath gas temperature and sheath gas flow rate were set to 350 °C and 11 L/min, respectively. The capillary voltage was set to 3500 V, and the nozzle voltage was set to 1000 V. The data was acquired in MS mode with a mass range of 50-1700 m/z. The MS acquisition rate and time were set to 1 spectrum/s and 1000 ms/spectrum, respectively. The MS data was acquired using MassHunter workstation software. The DNA adducts were separated using Agilent Zorbax RRHDC Eclipse plus C18 column (2.1 × 100 mm, 1.8 µm) maintained at 40 °C using mobile phases A: 0.1% formic acid in Milli-Q water and mobile phase B: 0.1% formic acid in acetonitrile. The injection volume and flow rate were set to 10 µL and 0.3 mL/min, respectively. The following gradient elution mode was used for the separation of the compound: 0 min: 90% A, 3 min: 85% A, 5 min:78% A, 10 min: 60% A, 13 min: 50% A, 15 min: 30% A, 18-20 min: 90% A with total run time of 20 min. Notably, the data‑dependent MS/MS (Auto MS/MS) was performed on the QTOF with collision energies ramped from 10 to 40 eV, precursor selection based on abundance, and dynamic exclusion. For low‑abundance precursor ions that did not fragment adequately in the QTOF, targeted MS/MS was performed using an Agilent 6495C triple quadrupole mass spectrometer with collision energies optimized for each precursor, enabling detection of low‑intensity fragment ions.

DNA adductomics on tissues from head and neck cancer patients

Raw data files (.d files) were converted to mzML format (refer Data availability section for HRMS data) and underwent peak centroiding using the msconvert utility of ProteoWizard [52]. These converted files were then searched against the MutAIverse adductome library, as well as the EVA and SA DNA adduct libraries, to compute cosine similarity scores. The AdductLinker module was employed to trace back the parent genotoxins associated with the detected EVA, SA, and MutAIverse-generated DNA adducts. Top hits, defined as those with a cosine similarity greater than 0.9, were identified, and further, their precursor m/z values were confirmed by high-resolution mass spectrometry.

Multivariate analysis of samples

To assess variation across samples based on their spectral composition, we conducted a multivariate analysis using MetaboAnalyst 6.0 [53]. This platform employs peak detection, peak picking, peak alignment, and gap-filling algorithms to identify and standardize features across samples. Principal Component Analysis (PCA) was then performed to explore clustering patterns and visualize the relationships among the samples.

DNA adduct abundance and patient age analysis

Intensity values for m/z 377.18 and 266.07 in tumor and tumor-adjacent control samples were extracted from the feature list generated via LC-MS processing using MetaboAnalyst 6.0 [53]. To assess statistical significance between different sample types, the Mann-Whitney U test was performed. Additionally, to analyze the abundance of m/z 377.18 and 266.07 relative to patient age, Pearson’s correlation coefficient was calculated along with p-values for linear regression analysis.

Statistical analysis

All statistical analyses were performed using SciPy Python package (version 1.14.0). Kolmogorov-Smirnov test was used to test differences in probability distributions for statistical significance. Mann-Whitney U test was performed to compare the medians of two non-parametric distributions. The p-value cutoff for significance was 0.05. Significance levels are: * < 0.05, ** < 0.01, *** < 0.001, and **** < 0.0001.

Author contributions

The study idea was conceived by G.A. Computational and experimental parts of the manuscript were designed and supervised by G.A. and R.M.B., respectively. DNA adductomics was performed by S.J. under the supervision of R.M.B. MutAIverse workflow was developed by S.S. and S.K.M., and the architecture was built by S.S. and S.K.M. S.S. and A.V. developed a generative de novo design workflow. Data analyses were performed by S.S., S.K.M., S.J., Sa.S., and V.G. MutAIverse library was tested by S.S., Sa.S, S.A., A.M., A.S., S.C., S.D., and S.K. Statistical guidance was provided by J.T., A.M., and D.S. Human sample collections were performed by A.R., A.D., K.K., K.D., and A.S., and processed by S.J. and R.M.B. Illustrations were drafted by S.S. and S.K.M. S.S., G.A., and R.M.B. wrote the paper. All authors have read and approved the manuscript.

Data availability

All the datasets used and generated in this manuscript are provided as Supplementary information or on the GitHub repository (https://github.com/the-ahuja-lab/MutAIverse). High-resolution mass spectrometry data, along with the corresponding sample-level information, have been deposited in Zenodo under the 10.5281/zenodo.13377479.

Code availability

The source code for MutAIverse is available on the GitHub repository (https://github.com/the-ahuja-lab/MutAIverse). Python package for MutAIverse is available via pip: https://pypi.org/project/MutAIverse/0.2.4/. We also set up the entire run code on Google Colab: https://colab.research.google.com/drive/1cidZQgZHD0wDcTsRMbO2jY0NFZfKW-Ig?usp=sharing.

Declarations

Ethics approval and consent to participate

The local Medical Ethics Committee at Dr. Bhubaneswar Borooah Cancer Institute, Department of Atomic Energy, Government of India (Reference number: BBCI-TMC/Misc-01/MEC/252/2021) approved the use of head and neck tumor tissues from patients. This approval was in accordance with the guidelines of the Indian Council of Medical Research, New Delhi, Government of India.

Competing interests

The authors declare no competing interests.