Prioritizing DNA methylation biomarkers using graph neural networks and explainable AI

DNA methylation is a significant epigenetic modifi- cation involving the addition of a methyl group to the position 5′ of the cytosine residues. The modification is responsible for disease progression, immune response, and outcomes in diseases such as breast cancer (BC) and acute myeloid leukemia (AML). Illumina’s HumanMethylation450 BeadChip (450K) and EPIC BeadChip (850K) methylation arrays are heavily used for such cancer studies to determine differentially expressed and dif- ferentially methylated genomic regions. Many of these are biomarkers used effectively for exploring therapeutic targets. Several studies report a few potential biomarkers, but the enor- mous numbers of largely unexplored probe-level (CpG sites) methylation signals may contain additional significant biomark- ers. To prioritise the under-explored and disease-specific CpG sites from DNA methylation arrays and potentially uncover novel biomarkers, we present the novel approach GraphMeX- plain, a graph neural network (GNN)-based approach with ex- plainable AI module. The underlying graph neural network is a principal neighbourhood aggregation (PNA). The approach uses the biomarkers reported in recent studies to rank biomark- ers from the unexplored set. A similarity graph between CpG sites (known and unexplored sets) is constructed using DNA methylation β values from arrays, producing an interaction graph. Biomarkers from recent studies are used as seeds and from the unexplored CpG sites, highly-variable ones (exclud- ing the seeds) are selected that vary significantly between con- ditions (BC patients and normal controls for breast cancer ar- rays). Using the combination of seed and highly-variable CpG sites, a positive-unlabeled approach, network-informed adap- tive positive-unlabeled learning (NIAPU), is utilized to assign a set of soft labels to unknown CpG sites such as likely posi- tive, weakly negative, likely negative, and reliable negative in the descending order of likelihood of CpG sites being poten- tial biomarkers. The graph neural network, a multi-layer PNA, refines the soft label assignments and achieves a high F1 clas- sification score (weighted) of 0.93 for BC and 0.91 for AML. The most likely set of CpG sites, classified under "likely pos- itive", are further explored using GNNExplainer, an explain- able AI approach. Subgraphs for likely positive CpG sites, pre- dicted with high probabilities, are computed and their prox- imities to the original seed CpG sites are analysed. The CpG sites which are predicted as likely positives have close interac- tions to the seeds. The top likely positive CpG site for BC is cg13265740 (C6orf115) where gene C6orf115 is strongly associ- ated with BC. For AML, the top likely positive predicted CpG site is cg23281527 (KLHDC7A) where gene KLHDC7A plays a strong role in the mechanism of AML. A high percentage of these likely positive predicted CpG sites for both BC and AML, which remained unseen by the GNN model during training, are highly relevant to them and can serve as potential therapeutic targets and prognostic values. Contact: [email protected]

DNA Methylation; Breast Cancer; Acute Myeloid Leukemia; Graph neural networks; Explainable AI; 1: Background A. Introduction. DNA methylation is a stable modification that provides ideal conditions for epigenetic studies. CpG sites are the locations of many of these modifications and are in close proximity to gene promoters and regulatory el- ements (1). To enable large-scale computation of methyla- tion patterns, Illumina (2) introduced the Infinium Human- Methylation microarray platform, a high-throughput technol- ogy for genome-wide analysis. The initial version, the Hu- manMethylation27 (27K) BeadChip, targeted 27,578 CpG sites associated with 14,495 gene promoters which was fol- lowed by the more comprehensive HumanMethylation450 (450K) BeadChip assessing methylation at 482,421 CpG sites. The DNA-methylation arrays (Illumina Infinium 450K, EPIC/850k, and the newer EPIC-v2/935k) quantify cytosine methylation at CpG sites on single-base resolutions using bisulfite-converted DNA fused to probe sets ( 3–6). Array-based DNA methylation has been pivotal to study and analyse epigenetic factors for cancers importantly because a) it comprehensively captures widespread disease-specific hypo- and hyper-methylation genomic regions, b) it sup- ports differential methylation analyses, and c) it corelates with gene expression for disease-specific marker discovery. Several established methods (for example quantile and func- tional normalizations) address disease’s global hypo- and hyper-methylation patterns, and joint analysis with RNA- sequencing data yielded precise DNA methylation markers for cervical cancer ( 4, 7, 8). Although the Illumina 450K ar- ray measures approximately 485K CpG sites, only a small subset has been annotated from the large landscape with dis- eases. Epigenome-wide association studies (EW AS) knowl- edge bases aggregate CpGtrait hits across studies at a large scale and map to tenshundreds of thousands of unique CpGs. Anup Kumar et al. | bio Rχiv | January 26, 2026 | 1–14 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint However, it captures only a minority of the entire 450K de- sign revealing that most CpG sites remain comparatively un- studied ( 9). Many cancer studies typically distill methyla- tion signals from hundreds of thousands of probes down to dozens. A pediatric AML study analyzed 450K data and reduced the number of relevant features to 1300 CpG sites stratifying AML subtypes (10). Another AML study ini- tially found only 1243 CpG sites that were significantly dif- ferent between two clusters of AML patients. Further, the analysis found only 10 CpG sites from 649 exhibiting sig- nificant methylation while PRDM16 gene being highly ex- pressed ( 11). For hereditary BC and ovarian cancer, a re- cent study ( 12) analysed 450K data of more than 230 BC patients and 150 healthy humans to find 17 and 27 CpGs in EPCAM and RAD51C promoter regions hypermethylated, respectively and 4 CpG sites associated with a higher risk of BC. Annotated CpG sites in several studies are concen- trated on a relatively smaller slice of available sites. To allow a larger number of CpG sites within analysis for enhancing their attributions to specific cancers, we propose, GraphM- eXplain, a GNN-based approach to make use of the existing annotations of CpG sites for specific diseases such as BC and AML (13, 14) to prioritize additional biomarkers from the unstudied CpG sites. B. State of the art approaches. Microarray data such as 450K and 850K have been extensively used in the fields of cancer prognosis to measure DNA methylation and gene ex- pression patterns at CpG sites. GeneMANIA (15) creates a graph of gene interactions to predict diffuse large b-cell lym- phoma (DLBCL) and prostate cancer (PC). Spectral cluster- ing is used as a feature selection approach in conjunction with GNN which improves the precision of disease classi- fication to 10.90% and 16.22% in DLBCL and PC, respec- tively. The nodes (genes) in the interaction graph are rep- resented by the gene expression from microarrays. A combi- nation of GeneMANIA, feature selection, spectral clustering, and GNN methods is used to predict four diseases (DLBCL, leukemia, prostate cancer, ALL) (16). Both approaches use only one variant of GNN for analysis and use a small subset of features to distinguish between diseases. MOGONET ( 17) introduces the integration of multi-omics data from mRNA and microRNA expression and DNA methylation patterns to predict diseases using combined features in the view corre- lation discovery network (VDCN). This study utilizes data from Alzheimer’s disease, low-grade glioma, kidney cancer, and breast invasive carcinoma to predict their states using multi-omics data by training three graph convolutional net- works and VDCN. However, the study shows biomarkers at the gene level and is limited to using only graph convolu- tions. In the multi-level attention graph neural network ( 18) (MLA-GNN), similarity between co-expression of genes is used to create an interaction graph, and nodes are represented by their expressions. Combining the interaction graph and node representation, graph attention network is trained for disease (glioma grading and COVID-19 diagnosis) and sur- vival prediction (glioma). In GraphAge (19), complete set of DNA methylation data is converted into CpGCpG inter- action graph. The nodes (CpG) are initialized with a com- bination of DNA methylation values, CpG-island informa- tion, distance from the transcription start sites (TSS) and base pair positions. The graph edges integrate co-methylation with same-gene and same-chromosome indicators. A PNA- based GNN used in GraphAge outperformed AltumAge (20), DeepMAge (21), and Horvath ( 22) and post-hoc explana- tions (GNNExplainer (23)) are aggregated into methylation- regulated networks (MRNs) to support pathway explanations (cardiac muscle contraction pathway hypermethylates with age leading to reduced gene expression. However, GraphAge is framed as a regression method trained only on healthy samples and lacks analysis on biomarker discovery. Our ap- proach, GraphMeXplain, addresses these gaps for biomarker discovery by (i) building disease-specific interaction graphs directly from 450K BC and AML cohorts, (ii) assigning soft labels for unknown CpGs using positiveunlabeled (PU) learn- ing (NIAPU), (iii) benchmarking five diverse GNN archi- tectures, and (iv) validating predictions of novel biomarkers with prediction explanations based on graph connectivities. 2: Method A. Data collection. Datasets are downloaded from NCBI (25) for BC (26, 27) and AML (28, 29). For BC, DNA methy- lation 850K dataset ( 26) is available for 50 BC patients and 30 normal controls. For AML, the DNA methylation 450K dataset ( 28) is available for 30 patients divided into two co- horts according to when methylating levels are measured - day 0 and day 8 (after receiving the decitabine (DAC) drug which acts as a demethylating agent). B. Data preprocessing. The DNA methylation arrays for BC and AML contain β values of methylation levels for CpG sites. Each such measurement is created with a p-value which is used for filtering out (p-value >= 0.5) insignificant CpG sites. Furthermore, those CpG sites associated with single-nucleotide polymorphisms (SNPs) are filtered out to reduce the impact of genetic variation on DNA methylation levels. Additionally, gender-specific methylation bias is re- moved by excluding CpG sites (associated to the X and Y chromosomes). After preprocessing, CpG sites are annotated with their respective genes leading to an updated site name as "probe_gene". Probe and CpG site are interchangeably used in the following sections. The respective prior stud- ies for BC and AML analysed the Illumina arrays used in this approach and proposed a few biomarkers for both dis- eases. These known biomarkers are called as "seed" in this approach. To explore and rank additional novel biomarkers from the arrays, the CpG sites (excluding the seed CpG sites) are then evaluated for their variabilities between the condi- tions - for BC, between the DNA methylation levels of pa- tients and normal controls and for AML, DNA methylation levels of patients between day 0 and day 8 (after the admin- istration of DAC drug). 10000 CpG sites are chosen for fur- ther detection of biomarkers which show the highest varia- tion between respective pairs of conditions. These potential 10000 CpG sites are then combined with the seeds to create a (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint C Positive-unlabeled learning Feature 1Feature 2 Sample 1 ... ... Sample 2 ... ... ... ... ... Sample M ... ... Feature N ... ... ... ... d. NeDBIT features (degree, NetRank, Info diffusion, heat diffusion ...) e. Combined DNA methylation and NeDBIT features f. Train graph neural network using interaction graph and combined features g. Visualisations: model convergence and explainable AI graph Network-informed adaptive positive-unlabeled learning (NIAPU) a. DNA methylation beta values c. Annotated soft labels by NIAPU (positive, likely positive, weakly negative, likely negative, reliable negative) b. Interaction graph computed using similarity matrix between features Fig. 1. CpG site prioritization workflow: a) shows the tabular representation of DNA methylation β values. Features (columns) are CpG sites and samples (rows) are patients. Each cell contains DNA methylation β value ranging between 0 and 1. b) an interaction graph is created by computing k-nearest neighbours for each CpG site and then filtering those neighbours based on Pearson correlation threshold. This results in varying sets of neighbours for each site which are interpreted as interacting sites in a graph similar to protein-protein interaction networks. c) using network-informed adaptive positive-unlabeled learning (24), nodes (CpG sites) in the interaction graph, created in previous step, are softly annotated with labels. The nodes which are known to be implicated for BC and AML are labeled as positive. The next set of nodes which are closer to this positive set are labeled as likely positive. It serves as the most promising set of nodes to prioritize novel CpG sites. The other sets of nodes which are farther from the positive set are labeled as weakly negative, likely negative and reliably negative. d) NIAPU also computes graph topology-based features for each node. e) The graph topology-based features are concatenated with DNA methylation β values (from a) to have comprehensive sets of features for each node. f) utilising interaction graph from b), soft labels from c) and features from e), 5 different graph neural networks (GNNs) are trained and compared to find the best performing network. g) multiple visualizations are used to analyse the performance downstream prediction and prioritisation, and prediction explanation tasks. new dataset for both diseases separately. The unknown ones are then further analysed by PU approach to learn soft labels which is described in the next section. C. Positive-unlabeled learning. Positive-unlabeled (PU) learning (30, 31) is an approach designed for cases where only a subset of samples is labeled as positive, while the bulk of the samples are unlabeled and may contain hidden positives. Treating the samples as negatives is a hard as- sumption. Recently published approach, network-informed adaptive positive-unlabeled (NIAPU) (24), utilises the PU methodology for disease gene discovery by integrating bio- (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint logical network information to infer candidate genes. The unlabeled genes are not treated as negatives but a mixture of potential positives and true negatives. NIAPU combines two core ideas: a) graph topological features informed by network diffusion and biology (NeDBIT) capturing relation- ships between genes based on the interaction graph and b) adaptive positive-unlabeled (APU) labeling method to assign soft labels to genes (positive, likely positive, weakly nega- tive, likely negative, reliable negative) using a markov diffu- sion process (32). Both components work together enabling NIAPU to propagate information from positive genes to unla- beled ones based on network proximity and similarity within NeDBIT features. XGDAG ( 33) uses it to rank genes for 10 diseases achieving superior performance to state-of-the- art algorithms such as DIAMOnD ( 34) and GUILD (35) in multi-disease evaluations. C.1. Graph construction. Preprocessed 450K and 850K Il- lumina methylation arrays (see Methods section) are used to create separate interaction graphs for BC and AML. The combination of seeds and unknown probes (10000 highly variable probes), represented as "probe_gene", serve as fea- tures containing DNA methylation β values. Pearson corre- lation among the features are computed in an all-vs-all man- ner. Further, k-nearest neighbours are computed for each probe and neighbours of each probe are filtered out by pear- son correlation. A suitable threshold is computed for filter- ing out lower quality neighbours for each probe. Overall, it implies that the positive probes (seeds) show higher correla- tion to positive and potentially positive probes and the po- tentially negative probes show higher interaction among the negative probes and lower to the positive and potentially posi- tive set. Each node in the interaction graph represents a probe (CpG site) annotated with its gene ("probe_gene") and each retained pair of probe interactions represents an edge in the interaction graph (36). C.2. Feature computation. NIAPU computes two categories of NeDBIT features - a) graph topology-based (netshort, netring, degree and ring) and b) diffusion based (heatdiff and infodiff ( 37–39)) features. Computation of these fea- tures requires the interaction graph (see Graph construction section) and the association scores of genes with the disease. GraphMeXplain utilizes both categories of features - a) graph topology-based features measure the significance of nodes in the context of the specific disease by finding their proximi- ties to seeds and b) diffusion-based features model the pro- gression of association scores from seeds across the interac- tion network to gather the strength of associations between probes based on the network connectivity and diffusion dy- namics. The changes in DNA methylation β values of probes between different conditions (BC patients and normal groups for BC and between day 0 and day 8 of the same AML pa- tients) are utilized as their association scores to respective diseases. C.3. Soft labels assignment. APU assigns soft labels to all the probes using the graph topology- and diffusion-based fea- tures computed by NeDBIT. A subset of probes that is most distant from seeds (P), estimated by using the quantile thresh- old on the distances of the NeDBIT-features, are called reli- ably negative (RN) probes. A distance based transition ma- trix is computed using markov diffusion process (40). Fol- lowing the steady state of the diffusion process, the remain- ing unlabeled probes are partitioned into three sets (having similar number of probes) making use of adaptive quantiles - likely positive (LP), weakly negative (WN) and likely nega- tive (LN). LP set of probes are estimated to be the most sim- ilar to seeds. D. Node features and graph neural networks. The cor- responding NeDBIT and DNA methylation β values of probes belonging to 5 soft labels are concatenated to cre- ate combined set of features. With the interaction graph, concatenated features and the labeled probes, graph neural network (GNN)-based deep learning classifiers are utilised to iteratively refine the features and labels. GNN is heav- ily used for tasks such as node and graph classification for graph-based datasets. Gene prioritisation follows node clas- sification task where candidate genes (nodes) are ranked in the order of likelihood to be involved in a disease path- way (33, 41, 42). In our approach, enriching probe repre- sentations using graph neural networks, we benchmark five GNNs which differ in the manner of neighbourhood con- struction for probes. These GNNs include graph convolu- tion network (GCN) (43), GraphSAGE ( 44), graph atten- tion network (GA Tv2) which computes attention over neigh- bours (45), GraphTransformer which estimates global atten- tion with edge encodings for long-range dependencies ( 46) and principal neighbourhood aggregation (PNA) which con- catenates multiple aggregators with degree-scalers to han- dle heavy-tailed degree distributions (47). Each network in- cludes a multi-layer perceptron to project logits of GNNs into 5 dimensional output as classes. Based on the accuracy met- rics such as F1 score computed on the test data, PNA emerges as the best performing architecture out of 5 evaluated GNNs. D.1. Principal neighbourhood aggregation (PNA). Each probe in the interaction graph is represented by a vector of combined NeDBIT and DNA methylation β values. The combined vector for each probe initialises the GNNs for the node classification task which undergo transformation through successive GNN layers. The underlying aggregation approach in the GNN layers distinguishes the type of GNN. Traditional aggregators such as sum or mean are used which limits the expressivity of node representations. Therefore, to overcome this drawback, PNA combines multiple aggregators such as mean, maximum, minimum and standard deviation to compute representations of nodes from their respective neighbourhood. Furthermore, PNA uses degree-scalers to scale the aggregated representations based on the size of the neighbourhood of nodes ( 47). The combination of aggregated and degree-scaled representations provide richer signals to distinguish different neighbourhood structures and feature distributions robustly leading to the achievement of high accuracy in classifying probes. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint E Experimental setup E. Experimental setup. We benchmark five GNNs on the node classification task for two diseases - BC and AML us- ing 850K and 450K Illumina arrays, respectively. Interaction graphs are undirected for both diseases containing millions of edges and approximately 11,000 probes. The entire set of approximately 11,000 probes are divided into training and test data with a ratio of 2:1. The training data is further di- vided into a ratio of 2:1 for the actual training and valida- tion data. For BC, each probe is represented by 86 dimen- sional features - six NeDBIT, 50 BC patients and 30 normal controls. For AML, probes have 46 dimensional features - six NeDBIT, 20 patients whose DNA methylation levels are measured at day 0 and at day 8 after the administration of decitabine (DAC) drug making 20 × 2 = 40 features. Nvidia V100 GPU with 32 GB of memory is utilised for training all five GNNs for comparison separately. To restrict the size of neighbourhood in order to fit them into the GPU memory, neighbourhood sampler is used. It samples a 3-hop neigh- bourhood for each probe - all 1-hop neighbours (immediate neighbours), 50 2-hop neighbours and 25 3-hop neighbours. Each GNN contains four graph-based deep layers for mes- sage passing and aggregation and a linear layer for class pro- jection. The initial learning rate is 5e − 4 which is optimised by the learning rate scheduler. To minimize overfitting and stabilise training, a dropout (0.2), and weight decay (1e − 3) are used. All GNN models are trained for 20 epochs with a mini-batch size of 32. Best trained model for each GNN is retained based on its performance on validation data. Metrics such as precision, recall and F1 scores, are used to compare GNNs. Cross-entropy loss function is used to compute error between ground-truth (estimated by APU) and the predicted labels. The set of LP predicted probes is selected from the prediction made on test data. Using GNNExplainer (23), a subgraph is sampled for the top LP probe to visualize and understand its neighbourhood which explains the reasoning behind the prediction made by the GNN model. 3: Results Five GNNs are trained and their performance on test data are compared on multiple metrics such as F1, precision and re- call. The performance of each GNN is averaged over five experiment runs. V arious visualizations such as heatmap, UMAP , explanation subgraphs, violin and bar plots are used to analyse the performance of GNNs. A. Breast cancer (BC). A.1. Training and test performance. The heatmap in Figure 2 shows a performance comparison of PNA, GraphTrans- former, GraphSage, GCN and GA Tv2 GNN variants. PNA outperforms other GNNs on BC datasets achieving higher F1 score (weighted) of 0.93 in classifying five classes. Ad- ditionally, the performance of PNA is also superior on F1 (macro), F1 (micro), precision and recall metrics to other GNNs. Based on this superior performance, PNA is chosen as the classifier for the BC dataset for further downstream analysis. F1 (macro) F1 (weighted) F1 (micro) Precision Recall PNA GraphTransformer GraphSage GATV2 GCN 0.940 0.927 0.929 0.932 0.929 0.928 0.916 0.916 0.917 0.916 0.917 0.893 0.894 0.893 0.894 0.662 0.803 0.814 0.797 0.814 0.623 0.754 0.764 0.757 0.764 Models vs Metrics 0.0 0.2 0.4 0.6 0.8 1.0 Metric score Fig. 2. BC: The heatmap shows the comparison of five GNNs on multiple metrics showing that the PNA outperforms other GNNs on all compared metrics. A.2. Uniform Manifold Approximation and Projection (UMAP) embeddings. UMAP (48) learns a low-dimensional projection suitable for visualizing clusters in a dataset. UMAP is used to reduce the dimensions of raw NeDBIT and DNA methylation feature representations and the cor- responding trained representations (predicted by PNA) to two dimensions for each probe. The soft labels (provided by APU) and predicted labels (predicted by PNA) are used to annotate the lower dimensional points in Figures 3 and 4, respectively. In Figures 3 and 4, each two dimensional point refers to a probe. Figure 3 shows raw probe features (concatenated NeDBIT and β values of DNA methylation). The seeds (blue) are separated from other probes, LP (orange), WN (green), LN (red) and RN (purple), showing little separability. Contrastingly, in the UMAP of the probe representations (Figure 4), extracted from last but one layer of the PNA model, not only seeds (blue) show good separability but other classes are also bound to clear clusters. The clear separability of classes shows that the PNA model transforms the raw feature representations of probes for their robust classification. 0 5 10 15 20 UMAP1 5 0 5 10 15 20UMAP2 UMAP of raw features (NeDBIT + DNA Methylation): PNA Label 0 1 2 3 4 Fig. 3. BC: UMAP diagram shows untrained (raw) features for test probes reduced to 2 dimensions. Only the positive cluster (blue) is distinct whereas other clusters belonging to different classes show high overlap. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint 10 5 0 5 10 15 UMAP1 10 0 10 20 UMAP2 Embeddings UMAP of last PNAConv4 layer: PNA Class 0 1 2 3 4 Fig. 4. BC: UMAP diagram shows trained features for test probes. The 64 dimen- sional representations, computed from the last but one PNA layer, are reduced to 2 dimensions by UMAP . In comparison to Figure 3, all the clusters representing soft labels of probes show clear separability showcasing robust refinement of untrained features by PNA and leading to high classification accuracy . A.3. Top likely predicted probes. Likely positive (LP) set of probes may contain potentially positive probes signifi- cant for BC. Classes are predicted by using a trained PNA model on test data and probes with LP class are separated for further analysis. Top 20 LP probes are extracted based on their prediction probabilities and sorted by the differ- ence (shown in absolute numbers) of DNA methylation β values between two groups (BC patients and normal con- trols) in descending order. The top LP predicted probes show variability across both group and may contain novel biomarkers for BC. Furthermore, the top predicted probe is "cg13265740_C6orf115" where gene "C6orf115" (officially known as ABRACL) (49) promotes breast cancer progres- sion by elevating cell proliferation, invasion and migration. The second top LP predicted probe is "cg06645081_PDE8A" where gene "PDE8A" has been shown to be a key regula- tor of T-cell and breast cancer cell motility (50). Overall, out of the top 20 LP predicted probes and their associated genes, 12 genes are known to be associated with BC includ- ing a probe "cg17914584_BRCA2" located on BRCA2 gene, widely known for BC (51). A.4. Explainable AI. Understanding the predictions of AI models has been a challenge as the models are too complex and contain too large number of parameters to easily explain their predictions. Explainable AI methods are increasingly becoming necessary to reason with models predictions ( 52). We have used GNNExplainer to extract a subgraph for the top predicted node to understand and show the evidence for PNA model’s prediction of "cg13265740_C6orf115" probe (red) as LP . Figure 6 shows its subgraph where the probe is in close proximity of five positive nodes (green) explaining reason behind its prediction as a potential candidate for BC (53). The proximal nodes (blue) are also predicted as LP . Additionally, Figure 7 shows strong pearson correlation of "cg13265740_C6orf115" probe with positive probes further providing the evidence of it being a potential positive probe for BC. B. Acute myeloid leukemia (AML). B.1. Training and test performance. The PNA-based graph neural network architecture performed the best on 450K AML Illumina arrays achieving 0.91 F1 score (weighted) in classifying five classes (Figure 8). The model is used to predict classes of test probes, and the top-20 candidates pre- dicted from the LP class are analyzed. Figure 11 shows the top-20 LP predicted candidates that show differences in the methylation levels measured on day 0 and day 8 (after DAC drug treatment) between the same set of patients. The figure shows variability in DNA methylation levels across two time points. In contrast, Figure 12 shows the top-20 reliably neg- ative predicted candidates having virtually no difference in DNA methylation levels between day 0 and day proving that they are not promising candidates for further analysis. B.2. Top likely predicted genes. Likely positive (LP) set of probes may contain potentially positive probes significant for AML. Classes are predicted by using a trained PNA model on test data and probes with LP class are separated for fur- ther analysis. Top 20 LP probes are extracted based on their prediction probabilities and sorted by the difference of DNA methylation β values between two groups (AML pa- tients at day 0 and day 8 (after DAC drug treatment)) in descending order. The top LP predicted probes show vari- ability across both group and may contain novel biomark- ers for AML (Figure 11). The top LP predicted probe "cg23281527_KLHDC7A" is shown to be hypomethylated in AML subtype ALL ( 54). The second LP predicted probe "cg07854132_OVCH1" is not yet shown to be associated to AML or Decitabine drug. The third predicted LP probe is "cg03154665_TTC14". Mutations in the gene "TTC14" have been shown to be associated to AML ( 55). Additionally, (56) shows that "TTC14" gene is commonly upregulated in AML cell line (AML3) using RNA sequencing. Overall, 13 out of top-20 predicted LP probes are associated to AML including "cg22334059_RUNX1T1" where gene "RUNX1T1" ( 57) is strongly associated to AML. B.3. Explainable AI. The subgraph is created using GN- NExplainer (Figure 13) for the top LP predicted probe "cg23281527_KLHDC7A" (red) showing the five neighbour- ing probes (green) as the known positives. Its proxim- ity to known positive probes provides strong evidence for "cg23281527_KLHDC7A" probe to be predicted as LP . Ad- ditionally, its high pearson correlation to the positive probes (Figure 14) explains why it is a potentially positive probe. 4: Discussion In GraphMeXplain, a PNA-based GNN is used to predict and prioritize novel CpG sites from BC and AML DNA methyla- tion arrays. The approach leverages on a few known positive CpG sites from recent studies to find proximities to a set of potential CpG sites that can be used for therapeutic targets and also for further research in BC and AML based stud- ies. An explainable AI approach, GNNExplainer, computes suitable subgraph for CpG sites predicted as likely positives. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint B Acute myeloid leukemia (AML) cg13265740_C6orf115 cg06645081_PDE8A cg08624850_TOR1AIP2 cg08545593_CTSZ cg01313002_VPS41 cg22372654_VPS13B cg10447327_LANCL1 cg05309281_TRIP12 cg16478897_RMI2 cg07999824_MAP3K7CL cg20609803_FCER1G cg07412002_TMEM30A cg07302959_FAM133B cg08924923_ZMYM2 cg02401171_SPTA1 cg23098236_UBE2K cg17914584_BRCA2 cg17865849_GALC cg03751915_RNASEH2B-AS1 cg17212881_IL4R Feature (probe_gene) 0.0 0.2 0.4 0.6 0.8 1.0 1.2-value =0.060 =0.059 =0.057 =0.057 =0.056 =0.056 =0.056 =0.055 =0.055 =0.055 =0.054 =0.054 =0.053 =0.053 =0.053 =0.053 =0.052 =0.052 =0.052 =0.052 Distribution and mean differences per feature ( -values): BC vs Normal Conditions BC Normal Fig. 5. The figure shows DNA methylation β values (on y-axis) across sets of BC patients (in orange) and normal controls (in blue) for the top-20 likely positive (LP) predicted CpG sites (on x-axis). The set of LP CpG sites, unseen during training, are predicted with high probability and out of top-20, 12 are known to be associated with BC. 5984 6849 2692 6278 58016185 1932 6290 5619 5813 6231 Explanation subgraph of seed node: 5619 Nodes 1932:cg17660211_PCED1B 2692:cg09821790_SLC7A6 5619:cg13265740_C6orf115 5801:cg16399365_ZNF238 5813:cg17705041_PCGF3 5984:cg05287415_C1orf115 6185:cg11941546_CLDN7 6231:cg24978679_CIART 6278:cg11413039_PUS3 6290:cg18262079_TJP1 6849:cg10058464_C4orf22 Fig. 6. BC: The figure shows the explanation subgraph, computed by GNNEx- plainer, of the top LP predicted CpG site (red) "cg13265740_C6orf115" (from Fig- ure 5). The CpG sites shown in green are known positive sites (seeds) for BC. The subgraph shows that the top-LP predicted CpG site is closely connected to several known positive CpG sites explaining the reason for the site being pre- dicted as LP . In the figure, "5619" represents the internal index of the top LP node (cg13265740_C6orf115). Internal indices of other probes are shown in the legend along side their names. The explanation subgraph provides insights into the relation- ships of the likely positive CpG sites and their proximities to known positive sites. Additionally, the DNA methylation distributions of these likely positive sites exhibit variabilities for different groups (BC patients and normal controls and DNA methylation levels measured on day 0 and day 8 af- cg04179740_C10orf54 cg03388786_LYN cg18881723_SLAMF1 cg22832271_ARID5B cg26607031_SLC25A33 Seed signals 0.00 0.25 0.50 0.75Pearson correlation Correlation of cg13265740_C6orf115 with seeds Fig. 7. The bar plot shows the pearson correlation of the top-LP predicted CpG site "cg13265740_C6orf115" with the known positive CpG sites for BC (shown in Figure 6). The LP predicted "cg13265740_C6orf115" CpG site has strong correlation with known positive sites further explaining the reason of it being predicted as likely positive. ter DAC treatment). Several likely positive predicted CpG sites have been known to be associated with BC and AML but was unseen during GNN training showcases that the ap- proach is robust and the predictions can be validated with available knowledge of CpG sites and their associated genes. GraphMeXplain also benchmarks five different GNNs and chooses PNA as the best performing one based on the test dataset. One limitation of the approach lies in the graph con- (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint F1 (macro) F1 (weighted) F1 (micro) Precision Recall PNA GraphTransformer GraphSage GATv2 GCN 0.928 0.914 0.914 0.922 0.914 0.919 0.903 0.904 0.904 0.904 0.844 0.816 0.823 0.851 0.823 0.837 0.826 0.826 0.827 0.826 0.815 0.805 0.807 0.815 0.807 Models vs Metrics 0.0 0.2 0.4 0.6 0.8 1.0 Metric score Fig. 8. AML: The heatmap shows the comparison of five GNNs on multiple metrics showing the PNA outperforms other GNNs on all compared metrics. 10 5 0 5 10 15 UMAP1 10 5 0 5 UMAP2 UMAP of raw features (NeDBIT + DNA Methylation): PNA Label 0 1 2 3 4 Fig. 9. AML: UMAP diagram shows untrained (raw) features for test probes reduced to 2 dimensions. Only the positive cluster (blue) is distinct whereas other clusters belonging to different classes show high overlap. struction where Illumina arrays are specific to each disease (850K for BC and 450K for AML). Multi-omics datasets can be explored to support and refine graph edges based on DNA methylation patterns across CpG sites for different groups. Overall, GraphMeXplain showcases huge potential in apply- ing GNN-based learning on tabular datasets in their graphical representation to rank potential biomarkers across 2 diseases. 5: Availability of supporting source code and requirements Project name: Gene prioritization using Graph neural net- works for Illumina arrays Availability: The source code can be made available upon reasonable request Operating system: Linux Programming languages: Python, XML Licence: MIT License 6: Declarations A. List of abbreviations. BC: Breast Cancer; AML: Acute Myeloid Leukemia; GNN: Graph Neural Network; GCN: 10 5 0 5 10 15 20 UMAP1 10 5 0 5 10 15 20 UMAP2 Embeddings UMAP of last PNAConv4 layer: PNA Class 0 1 2 3 4 Fig. 10. AML: UMAP diagram shows trained features for test probes. The 64 dimen- sional representations, computed from the last but one PNA layer, are reduced to 2 dimensions by UMAP . In comparison to Figure 9, all the clusters representing soft labels of probes show clear separability showcasing robust refinement of untrained features by PNA and leading to high classification accuracy . Graph Convolution Network; PNA: Principal Neighbour Ag- gregation; NIAPU: Network-Informed Adaptive Positive- Unlabeled Learning; APU: Adaptive Positive-Unlabeled Learning; B. Competing interests. The authors declare that they have no competing interests. C. Author Approvals. All authors have seen and approved the manuscript, and it has not been accepted or published elsewhere. D. Authors’ contributions. Authors’ contributions follow the order of names. A.K. designed the experiments, wrote scripts and original draft of the manuscript. T.A.D pro- vided data references, validated ideas and contributed to the manuscript, B.G, H.B. and R.B validated ideas and con- tributed to the manuscript. E. Funding. Prof. Dr. Rolf Backofen received funding from the German Federal Ministry of Education and Re- search (BMBF grant 03ZU1208CA, 03ZU1208DG) nan- odiag BW: Digitaler Nanoporen-Sequenzierer & Marker "In- teractom Profiler". F . Acknowledgements. We thank Nanodiag consortium for their support. 7: References 1. Adedeji O. Adetunji, Henrietta Owusu, Esiosa F . Adewale, Precious Adedayo Adesina, Christian Xedzro, T olulope Peter Saliu, Shahidul Islam, Zhendong Zhu, and Olanrewaju B. Morenikeji. Dna methylation: A key regulator in male and female reproductive outcomes. Life, 15(7), 2025. ISSN 2075-1729. doi: 10.3390/life15071109. 2. Illumina, Inc. Illumina sequencing technology . https://www.illumina.com/, 2025. Accessed: 2025-10-10. 3. Ruth Pidsley , Chloe C. Y . Wong, Manuela Volta, Katie Lunnon, Jonathan Mill, and Leonard C. Schalkwyk. A data-driven approach to preprocessing illumina 450k methy- lation array data. BMC Genomics, 14(1):293, 2013. ISSN 1471-2164. doi: 10.1186/ 1471-2164-14-293. 4. Jean-Philippe Fortin, Aurélie Labbe, Mathieu Lemire, Brent W. Zanke, Thomas J. Hudson, Elana J. Fertig, Celia M. T . Greenwood, and Kasper D. Hansen. Functional normalization of 450k methylation array data improves replication in large cancer studies. Genome Biology, 15(11):503, 2014. ISSN 1474-760X. doi: 10.1186/s13059-014-0503-2. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint F Acknowledgements 5. Ruth Pidsley , Elena Zotenko, Timothy J. Peters, Mitchell G. Lawrence, Gail P . Risbridger, Peter Molloy , Susan Van Djik, Beverly Muhlhausler, Clare Stirzaker, and Susan J. Clark. Critical evaluation of the illumina methylationepic beadchip microarray for whole-genome dna methylation profiling. Genome Biology, 17(1):208, 2016. ISSN 1474-760X. doi: 10. 1186/s13059-016-1066-1. 6. Dominic Guanzon, Jason P . Ross, Chenkai Ma, Oliver Berry , and Yi Jin Liew. Comparing methylation levels assayed in gc-rich regions with current and emerging methods. BMC Genomics, 25(1):741, 2024. ISSN 1471-2164. doi: 10.1186/s12864-024-10605-7 . 7. Xiaotu Ma, Yi-Wei Wang, Michael Q Zhang, and Adi F Gazdar. Dna methylation data analysis and its application to cancer research. Epigenomics, 5(3):301–316, 2013. doi: 10.2217/epi.13.26. PMID: 23750645. 8. Wanxue Xu, Mengyao Xu, Longlong Wang, Wei Zhou, Rong Xiang, Yi Shi, Y unshan Zhang, and Y ongjun Piao. Integrative analysis of dna methylation and gene expression identified cervical cancer-specific diagnostic biomarkers. Signal T ransduction and T argeted Therapy, 4(1):55, 2019. ISSN 2059-3635. doi: 10.1038/s41392-019-0081-6. 9. Thomas Battram, Paul Y ousefi, Gemma Crawford, Claire Prince, Mahsa Sheikhali Babaei, Gemma Sharp, Charlie Hatcher, María Jesús Vega-Salas, Sahar Khodabakhsh, Oliver Whitehurst, Ryan Langdon, Luke Mahoney , Hannah R. Elliott, Giulia Mancano, Matthew A. Lee, Sarah H. Watkins, Abigail C. Lay , Gibran Hemani, T om R. Gaunt, Caroline L. Relton, James R. Staley , and Matthew Suderman. The ewas catalog: a database of epigenome-wide association studies. Wellcome Open Research, 7:41, 2022. doi: 10.12688/wellcomeopenres.17598.2. l ’ 2022 Battram T et al. No competing interests were disclosed. 10. Olga Krali, Josefine Palle, Christofer L. Bäcklin, Jonas Abrahamsson, Ulrika Norén- Nyström, Henrik Hasle, Kirsi Jahnukainen, Ólafur Gísli Jónsson, Randi Hovland, Bir- gitte Lausen, Rolf Larsson, Lars Palmqvist, Anna Staffas, Bernward Zeller, and Jes- sica Nordlund. Dna methylation signatures predict cytogenetic subtype and outcome in pediatric acute myeloid leukemia (aml). Genes, 12(6), 2021. ISSN 2073-4425. doi: 10.3390/genes12060895. 11. Genki Y amato, T omoko Kawai, Norio Shiba, Junji Ikeda, Y usuke Hara, Kentaro Ohki, Shin- Ichi Tsujimoto, T aeko Kaburagi, Kenichi Y oshida, Y uichi Shiraishi, Satoru Miyano, Nobu- taka Kiyokawa, Daisuke T omizawa, Akira Shimada, Manabu Sotomatsu, Hirokazu Arakawa, Souichi Adachi, T akashi T aga, Keizo Horibe, Seishi Ogawa, Kenichiro Hata, and Y asuhide Hayashi. Genome-wide dna methylation analysis in pediatric acute myeloid leukemia. Blood Advances, 6(11):3207–3219, 2022. ISSN 2473-9529. doi: https://doi.org/10.1182/ bloodadvances.2021005381. 12. Miguel Ruiz-De La Cruz, Héctor Martínez-Gregorio, Clara Estela Díaz-Velásquez, Fernando Ambriz-Barrera, Norma Gabriela Resendiz-Flores, Rina Gitler-Weingarten, María Patricia Rojo-Castillo, Didier Pradda, Javier Oliver, Sandra Perdomo, Eva María Gómez-García, Aldo Hugo De La Cruz-Montoya, Luis Ignacio T errazas, Gabriela T orres-Mejía, Fidel de la Cruz Hernández-Hernández, and Felipe Vaca-Paniagua. Methylation marks in blood dna reveal breast cancer risk in patients fulfilling hereditary disease criteria. npj Precision On- cology, 8(1):136, 2024. ISSN 2397-768X. doi: 10.1038/s41698-024-00611-z. 13. Tiantian Wang, Peilong Li, Qiuchen Qi, Shujun Zhang, Y an Xie, Jing Wang, Shibiao Liu, Suhong Ma, Shijun Li, Tingting Gong, Huiting Xu, Mengqiu Xiong, Guanghua Li, Chongge Y ou, Zhaofan Luo, Juan Li, Lutao Du, and Chuanxin Wang. A multiplex blood-based as- say targeting dna methylation in pbmcs enables early detection of breast cancer. Nature Communications, 14(1):4724, 2023. ISSN 2041-1723. doi: 10.1038/s41467-023-40389-5 . 14. Gabriele Greve, Geoffroy Andrieux, Pascal Schlosser, Nadja Blagitko-Dorfs, Usama-Ur Rehman, T obias Ma, Dietmar Pfeifer, Gerhard Heil, Andreas Neubauer, Jürgen Krauter, Michael Heuser, Helmut R. Salih, Konstanze Döhner, Hartmut Döhner, Björn Hackanson, Melanie Boerries, and Michael Lübbert. In vivo kinetics of early , non-random methylome and transcriptome changes induced by dna-hypomethylating treatment in primary aml blasts. Leukemia, 37(5):1018–1027, 2023. ISSN 1476-5551. doi: 10.1038/s41375-023-01876-2. 15. Kun Y u, Weidong Xie, Linjie Wang, Shoujia Zhang, and Wei Li. Determination of biomarkers from microarray data using graph neural network and spectral clustering. Scientific Reports, 11(1):23828, 2021. ISSN 2045-2322. doi: 10.1038/s41598-021-03316-6 . 16. Weidong Xie, Wei Li, Shoujia Zhang, Linjie Wang, Jinzhu Y ang, and Dazhe Zhao. A novel biomarker selection method combining graph neural network and gene relationships applied to microarray data. BMC Bioinformatics, 23(1):303, 2022. ISSN 1471-2105. doi: 10.1186/ s12859-022-04848-y. l ’ 2022 The Author(s). 17. T ongxin Wang, Wei Shao, Zhi Huang, Haixu T ang, Jie Zhang, Zhengming Ding, and Kun Huang. Mogonet integrates multi-omics data using graph convolutional networks allow- ing patient classification and biomarker identification. Nature Communications, 12(1):3445, 2021. ISSN 2041-1723. doi: 10.1038/s41467-021-23774-w. 18. Xiaohan Xing, Fan Y ang, Hang Li, Jun Zhang, Y u Zhao, Mingxuan Gao, Junzhou Huang, and Jianhua Y ao. Multi-level attention graph neural network based on co-expression gene modules for disease diagnosis and prognosis. Bioinformatics, 38(8):2178–2186, 02 2022. ISSN 1367-4803. doi: 10.1093/bioinformatics/btac088. 19. Saleh Sakib Ahmed, Nahian Shabab, Abul Hassan Samee, and M. Sohel Rahman. Graphage: Unleashing the power of graph neural network to decode epigenetic aging. PNAS Nexus, 4(6):pgaf177, Jun 2025. ISSN 2752-6542. doi: 10.1093/pnasnexus/pgaf177. l ’ The Author(s) 2025. Published by Oxford University Press on behalf of National Academy of Sciences. 20. Lucas Paulo de Lima Camillo, Louis R. Lapierre, and Ritambhara Singh. A pan-tissue dna-methylation epigenetic clock based on deep learning. npj Aging, 8(1):4, 2022. ISSN 2731-6068. doi: 10.1038/s41514-022-00085-y . 21. Fedor Galkin, Polina Mamoshina, Kirill Kochetov, Denis Sidorenko, and Alex Zhavoronkov. Deepmage: A methylation aging clock developed with deep learning. Aging and Disease, 12(5):1252–1262, 2021. ISSN 2152-5250. doi: 10.14336/AD.2020.1202. l ’ 2021 Galkin et al. Deep Longevity and Insilico Medicine are for-profit organizations developing artificial intelligence solutions for aging research, drug discovery , and longevity medicine. A patent has been applied for the described model and accompanying method. 22. Steve Horvath. Dna methylation age of human tissues and cell types. Genome Biology, 14 (10):R115, 2013. ISSN 1474-760X. doi: 10.1186/gb-2013-14-10-r115. Received June 10, 2013; Accepted October 4, 2013; Published October 21, 2013. 23. Rex Ying, Dylan Bourgeois, Jiaxuan Y ou, Marinka Zitnik, and Jure Leskovec. GNN ex- plainer: A tool for post-hoc explanation of graph neural networks. CoRR, abs/1903.03894, 2019. 24. Paola Stolfi, Andrea Mastropietro, Giuseppe Pasculli, Paolo Tieri, and Davide Vergni. Niapu: network-informed adaptive positive-unlabeled learning for disease gene identification. Bioin- formatics, 39(2):btac848, 02 2023. ISSN 1367-4811. doi: 10.1093/bioinformatics/btac848. 25. National Center for Biotechnology Information. National center for biotechnology informa- tion. https://www.ncbi.nlm.nih.gov/, 2025. Accessed: 2025-10-10. 26. Tiantian Wang and C. Wang. Epigenome analysis of normal and breast cancer (bc) sam- ples. NCBI GEO Series accession GSE237036, 2023. 27. Inc. Illumina. Illumina infinium humanmethylation850 methylationepic. GEO Platform ac- cession GPL21145, 2015. 28. Geoffroy Andrieux, G. Greve, N. Blagitko-Dorfs, M. Boerries, and M. Lübbert. Methylation data of serially sorted primary aml patient blasts prior and after treatment with the dnmt inhibitor decitabine (dac). Gene Expression Omnibus, 2023. https://www.ncbi.nlm. nih.gov/geo/query/acc.cgi?acc=GSE175758. 29. Inc. Illumina. Illumina humanmethylation450 beadchip (humanmethylation450k) platform annotation. Gene Expression Omnibus, 2011. https://www.ncbi.nlm.nih.gov/ geo/query/acc.cgi?acc=GPL13534. 30. Marthinus C. du Plessis, Gang Niu, and Masashi Sugiyama. Analysis of learning from positive and unlabeled data. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems , volume 27. Curran Associates, Inc., 2014. 31. Guangxin Su, Weitong Chen, and Miao Xu. Positive-unlabeled learning from imbalanced data. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 2995–3001. International Joint Conferences on Artificial Intelligence Organization, 8 2021. doi: 10.24963/ijcai.2021/412. Main T rack. 32. T olga Can, Orhan Çamoundefinedlu, and Ambuj K. Singh. Analysis of protein-protein inter- action networks using random walks. In Proceedings of the 5th International Workshop on Bioinformatics, BIOKDD ’05, page 6168, New Y ork, NY , USA, 2005. Association for Com- puting Machinery . ISBN 1595932135. doi: 10.1145/1134030.1134042. 33. Andrea Mastropietro, Gianluca De Carlo, and Aris Anagnostopoulos. Xgdag: explainable genedisease associations via graph neural networks. Bioinformatics, 39(8):btad482, 08 2023. ISSN 1367-4811. doi: 10.1093/bioinformatics/btad482. 34. Saeed D. Ghiassian, Jörg Menche, and Albert-László Barabási. A disease module detection (diamond) algorithm derived from a systematic analysis of connectivity patterns of disease proteins in the human interactome. PLoS Computational Biology , 11(4):e1004120, 2015. doi: 10.1371/journal.pcbi.1004120. 35. Emre Güney and Baldo Oliva. Exploiting protein-protein interaction networks for genome- wide disease-gene prioritization. PLoS ONE, 7(9):e43557, 2012. doi: 10.1371/journal.pone. 0043557. 36. Xuanyu Li, Xuan Zhang, Wenduo He, Deliang Bu, and Sanguo Zhang. Gene expression prediction based on neighbour connection neural network utilizing gene interaction graphs. PLoS One, 18(2):e0281286, February 2023. ISSN 1932-6203. doi: 10.1371/journal.pone. 0281286. Epub 2023 Feb 6. 37. Panisa Janyasupab, Apichat Suratanee, and Kitiporn Plaimas. Network diffusion with cen- trality measures to identify disease-related genes. Mathematical Biosciences and Engineer- ing, 18(3):2909–2929, March 2021. ISSN 1551-0018. doi: 10.3934/mbe.2021147. 38. Daniel E. Carlin, Barry Demchak, David Pratt, Elissa Sage, and T rey Ideker. Network propa- gation in the cytoscape cyberinfrastructure. PLoS Computational Biology, 13(10):e1005598, 2017. doi: 10.1371/journal.pcbi.1005598. 39. Dorothee Nitsch, João P . Gonçalves, Francisco Ojeda, Bart de Moor, and Yves Moreau. Candidate gene prioritization by network analysis of differential expression using machine learning approaches. BMC Bioinformatics, 11:460, 2010. doi: 10.1186/1471-2105-11-460 . 40. Peng Y ang, Xiao-Li Li, Jian-Ping Mei, Chee-Keong Kwoh, and See-Kiong Ng. Positive- unlabeled learning for disease gene identification. Bioinformatics, 28(20):2640–2647, 08 2012. ISSN 1367-4803. doi: 10.1093/bioinformatics/bts504. 41. Xinyu Xi, Jian Li, Jing Jia, et al. A mechanism-informed deep neural network enables prioritization of regulators that drive cell state transitions. Nature Communications, 16:1284, 2025. doi: 10.1038/s41467-025-56475-9. 42. A. Rao, S. Vg, T . Joseph, S. Kotte, N. Sivadasan, and R. Srinivasan. Phenotype-driven gene prioritization for rare diseases using graph convolution on heterogeneous networks. BMC Medical Genomics , 11(1):57, Jul 2018. doi: 10.1186/s12920-018-0372-8 . 43. Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. CoRR, abs/1609.02907, 2016. 44. William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. CoRR, abs/1706.02216, 2017. 45. Shaked Brody , Uri Alon, and Eran Y ahav. How attentive are graph attention networks? CoRR, abs/2105.14491, 2021. 46. Y unsheng Shi, Zhengjie Huang, Wenjin Wang, Hui Zhong, Shikun Feng, and Y u Sun. Masked label prediction: Unified massage passing model for semi-supervised classifica- tion. CoRR, abs/2009.03509, 2020. 47. Gabriele Corso, Luca Cavalleri, Dominique Beaini, Pietro Liò, and Petar Velickovic. Principal neighbourhood aggregation for graph nets. CoRR, abs/2004.05718, 2020. 48. Leland McInnes, John Healy , and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction, 2020. 49. Jie Li and Hui Chen. Actin-binding Rho activating C-terminal like (ABRACL) transcription- ally regulated by MYB proto-oncogene like 2 (MYBL2) promotes the proliferation, invasion, migration and epithelial-mesenchymal transition of breast cancer cells. Bioengineered, 13 (4):9019–9031, 2022. doi: 10.1080/21655979.2022.2056821. PMID: 35341461. 50. H. Dong, K. P . Claffey , S. Brocke, and P . M. Epstein. Inhibition of breast cancer cell migration by activation of camp signaling. Breast Cancer Research and T reatment, 152(1):17–28, 2015. doi: 10.1007/s10549-015-3445-9. 51. Huaizhi Huang, Chunling Hu, Jie Na, Steven N. Hart, Rohan David Gnanaolivu, Mo- hamed Abozaid, T ara Rao, Y ohannes A. T ecleab, Christine B. Ambrosone, Song Y ao, Amy T rentham-Dietz, A. Heather Eliassen, Lauren R. T eras, Alpa Patel, Christopher A. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint Haiman, Esther M. John, Elena Martinez, James V . Lacey , Dale P . Sandler, Clarice R. Weinberg, Julie R. Palmer, Celine M. Vachon, Janet E. Olson, Kathryn E. Ruddy , Hoda Anton-Culver, Jeffrey N. Weitzel, Peter Kraft, Tina Pesaran, Paulo Cilas Morais Lyra, Rachid Karam, Siddhartha Y adav, Katherine L. Nathanson, Susan M. Domchek, Miguel de la Hoya, Mark Robson, Miika Mehine, Chaitanya Bandlamudi, Diana Mandelker, Alvaro N. A. Mon- teiro, Edwin S. Iversen, Nicholas Boddicker, Wenan Chen, Marcy E. Richardson, Fergus J. Couch, and CARRIERS Consortium. Functional evaluation and clinical classification of BRCA2 variants. Nature, 638(8050):528–537, February 2025. ISSN 1476-4687. doi: 10.1038/s41586-024-08388-8. 52. Amina Adadi and Mohammed Berrada. Peeking inside the black-box: A survey on explain- able artificial intelligence (xai). IEEE Access, 6:52138–52160, 2018. doi: 10.1109/ACCESS. 2018.2870052. 53. Charu Kothari, Alisson Clemenceau, Geneviève Ouellette, Kaoutar Ennour-Idrissi, Annick Michaud, René C-Gaudreault, Caroline Diorio, and Francine Durocher. Tbc1d9: An impor- tant modulator of tumorigenesis in breast cancer. Cancers (Basel), 13(14):3557, Jul 2021. ISSN 2072-6694. doi: 10.3390/cancers13143557. 54. María del Pilar Navarrete-Meneses and Patricia Pérez-Vera. Epigenetic alterations in acute lymphoblastic leukemia. Boletín Médico del Hospital Infantil de México (English Edition), 74 (4):243–264, 2017. ISSN 2444-3409. doi: 10.1016/j.bmhime.2018.01.004. 55. Valeria Visconte, Holly J. Rogers, Jasmeet Singh, Jason Barnard, Manohar Bupathi, Fabiola T raina, John McMahon, Hideki Makishima, Hania Szpurka, Alicja Jankowska, Adeela Jerez, Mikkael A. Sekeres, Y ogen Saunthararajah, Amit S. Advani, Edward Copelan, Haruhiko Koseki, Kenichi Isono, Richard A. Padgett, Sherif Osman, Kaori Koide, Christine O’Keefe, Jaroslaw P . Maciejewski, and Ramon V . Tiu. Sf3b1 haploinsufficiency leads to formation of ring sideroblasts in myelodysplastic syndromes. Blood, 120(16):3173–3186, 2012. doi: 10.1182/blood-2012-05-430876. Epub 2012-07-23. 56. Mohammed Nahari. T argeting apurinic/apyrimidinic endonuclease 1 and 8-oxoguanine dna glycosylase as therapeutic strategy in acute myeloid leukaemia. https://theses. ncl.ac.uk/jspui/bitstream/10443/3896/1/Nahari%20M%202017.pdf, 2025. Accessed: 2026-01-16. 57. Pei Han Y u, Ze Y an Zhang, Y uan Y uan Kang, Ping Huang, Chang Y ang, and Hua Naranman- dura. Acute myeloid leukemia with t(8;21) translocation: Molecular pathogenesis, potential therapeutics and future directions. Biochemical Pharmacology, 233:116774, 2025. ISSN 0006-2952. doi: https://doi.org/10.1016/j.bcp.2025.116774. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint F Acknowledgements cg23281527_KLHDC7A cg07854132_OVCH1 cg03154665_TTC14 cg21481937_PTPRR cg01840128_SLC6A19 cg10656687_DACT2 cg26603598_COX7C cg25299371_KIAA1486 cg00130808_LOC145845 cg14827929_TLR3 cg11498870_DCAF4L2 cg22185977_LPCAT1 cg10533331_RASGRP1 cg18432471_GPR78 cg13008301_CSTF1 cg14983135_UPP1 cg22334059_RUNX1T1 cg02336104_FOXP1 cg08310581_ARHGEF4 cg15954675_SYNPR Feature (probe_gene) 0.0 0.2 0.4 0.6 0.8 1.0 1.2-value =0.219 =0.217 =0.201 =0.198 =0.197 =0.181 =0.181 =0.181 =0.179 =0.170 =0.165 =0.165 =0.161 =0.159 =0.158 =0.157 =0.153 =0.151 =0.150 =0.145 Distribution and mean differences per feature ( -values): Day0 vs Day8 Conditions Day0 Day8 Fig. 11. The figure shows DNA methylation β values (on y-axis) across sets of AML patients at day 0 (in orange) and day 8 (in blue) after adminstering DAC for the top-20 likely positive (LP) predicted CpG sites (on x-axis). The set of LP CpG sites, unseen during training, are predicted with high probability and out of top-20, 13 are known to be associated to AML. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint cg00315391_SCNN1G cg10548492_ACN9 cg04184836_CPEB1 cg14562915_UACA cg04264299_SCARA3 cg03631755_FGF4 cg15760257_SARM1 cg00593962_FZD8 cg03180780_SAMD11 cg24199112_FGF4 cg18539474_MAGI3 cg14235793_KIAA1244 cg21417510_PPM1J cg00225576_PAQR5 cg04912999_PAQR9 cg05358815_SATB2 cg00097177_GOLGA7B cg07303968_HSPA12A cg19722720_WNT2B cg20185017_CACNB2 Feature (probe_gene) 0.0 0.2 0.4 0.6 0.8 1.0 1.2-value =0.015 =0.015 =0.015 =0.015 =0.013 =0.011 =0.011 =0.011 =0.011 =0.011 =0.011 =0.011 =0.011 =0.010 =0.010 =0.010 =0.010 =0.010 =0.010 =0.010 Distribution and mean differences per feature ( -values): Day0 vs Day8 Conditions Day0 Day8 Fig. 12. The figure shows DNA methylation β values (on y-axis) across sets of AML patients at day 0 (in orange) and day 8 (in blue) after adminstering DAC for the top-20 likely (LN) and reliably negative (RN) predicted CpG sites (on x-axis). These CpG sites extremely small difference in methylation levels at day 0 and day 8. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint F Acknowledgements 131 388 551 457 269 505 273 338 345 28 604 Explanation subgraph of seed node: 273 Nodes 28:cg26246840_CD226 131:cg22862734_FREM2 269:cg05592278_LRRC37A3 273:cg23281527_KLHDC7A 338:cg12915585_MYT1L 345:cg26626251_OPCML 388:cg20542619_NFATC1 457:cg03639488_SPOCK1 505:cg05141695_DUSP22 551:cg23497306_PCDHB15 604:cg23689428_LOC84931 Fig. 13. The figure shows the explanation subgraph, computed by GNNExplainer, of the top-LP predicted CpG site "cg23281527_KLHDC7A" (from Figure 11) shown in red. The CpG sites shown in green are known positive sites for AML. The subgraph shows that the top-LP predicted CpG site is closely connected to several known positive CpG sites explaining the site being predicted as likely positive. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint cg22862734_FREM2 cg05592278_LRRC37A3 cg26246840_CD226 cg12915585_MYT1L cg26626251_OPCML Seed signals 0.0 0.5 1.0Pearson correlation Correlation of cg23281527_KLHDC7A with seeds Fig. 14. The bar plot shows the Pearson correlation of the top-LP predicted CpG site "cg23281527_KLHDC7A" with the known positive CpG sites for BC (shown in Figure 6). The LP predicted "cg23281527_KLHDC7A" CpG site has high correlation with known positive sites further explaining the reason of it being predicted as likely positive. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted January 27, 2026. ; https://doi.org/10.64898/2026.01.26.701692doi: bioRxiv preprint