The central data resource is same-cell paired multi-omics. Technologies including sci-CAR, SNARE-seq, SHARE-seq, Paired-seq and 10x Multiome measure gene expression and chromatin accessibility in the same cell, allowing candidate regulatory elements to be connected to transcriptional output. Their development changed the inference problem: accessibility indicates regulatory potential, whereas expression reflects the current result of regulation, and the relationship is neither instantaneous nor deterministic. A promoter may become accessible before its gene is induced, while an occupied enhancer may not alter steady-state expression. Matched-but-unpaired measurements offer a less direct alternative, and deep generative systems such as scMVP and scMOG can impute one modality from the other. These strategies expand usable data, but they introduce anchoring assumptions and model-dependent artifacts that must be tested rather than treated as equivalent to simultaneous observation.

Among interpretable approaches, motif and co-accessibility methods remain important. SCENIC combines co-expression modules with transcription-factor motif enrichment to score regulons, while SCENIC+ extends this logic to paired multi-omics by proposing enhancer-driven networks linking regulatory elements, transcription factors and target genes. Other systems infer links through different definitions of regulatory potential. scREG combines expression and accessibility to construct subpopulation-specific cis-regulatory networks, while FigR identifies domains of regulatory chromatin and links them to genes in stimulated immune cells. DIRECT-NET uses gradient boosting to nominate cis-regulatory elements and their targets, including within subpopulations or dynamic states. MICA adds mutual information and accessibility refinement for settings with few cells, capturing non-monotonic dependencies that simple linear correlation can miss. These methods offer biological interpretability, but their dependence on sequence motifs, local co-variation and selected priors limits their ability to represent nonlinear or global regulation.

Deep integration and graph learning are designed for the scale, sparsity and complexity of single-cell multi-omics. LINGER combines paired-cell data with atlas-scale external expression information and transcription-factor motif priors through manifold regularization, then transfers inferred networks to expression-only samples to estimate transcription-factor activity. scMultiomeGRN uses neighbor aggregation and cross-modal attention to construct cell-specific networks. DeepMAPS represents genes, peaks and cells as a heterogeneous graph and applies a graph transformer to learn local and global relationships. GLUE performs graph-linked unified embedding for multi-omics integration at scales reaching millions of cells. Other approaches add directed priors, global autoencoders or spatial associations to graph attention. These architectures can exploit topology beyond local correlation, but their results depend heavily on graph construction, prior quality, annotations and benchmark labels, meaning a larger model is not automatically a more reliable regulatory model.

Because tumor states can be continuous and plastic, several methods explicitly model dynamics. Dictys combines multi-omics footprinting, stochastic process networks and probabilistic models to infer context-dependent networks along developmental continua, while inferCSN uses pseudotime and accounts for uneven cell distributions when estimating regulatory relationships. scTIGER uses attention-based convolutional networks to detect condition-specific rewiring, and scMTNI applies multi-task learning to lineages and cell types. A separate family tests candidate causal hypotheses through virtual perturbation. CellOracle removes or alters inferred transcription-factor relationships to simulate loss of function; Pando combines multi-omics, binding-site predictions and organoid experiments; and GenKI and scTenifoldKnk generate virtual knockouts from wild-type data. These simulations can prioritize regulators, but a predicted knockout response remains a hypothesis until reproduced by CRISPR knockout, CRISPR interference, Perturb-seq or another orthogonal experiment.

Benchmarking exposes a gap between network reconstruction and biological usefulness. Common edge metrics, including area under the receiver-operating-characteristic curve, area under the precision-recall curve, early precision, odds ratio and statistical confidence, are affected by class imbalance, incomplete labels, threshold choices and the limited coverage of reference networks. Resources such as ENCODE and ChIP-seq provide valuable binding evidence but remain incomplete and biased toward particular cell types, while DREAM and BEELINE benchmarks were built largely for earlier inference settings. Perturb-seq and single-cell CRISPR screens add functional measurements but cover restricted genes, states and contexts. The review therefore recommends donor- or cell-type-aware test splits, reporting of label coverage and observable metric ceilings, tests against simple co-expression and gene-level baselines, negative controls, subsampling checks, hyperparameter sensitivity analysis and code availability.

Cancer applications show why cell-state specificity may matter. In a mouse model of skin squamous-cell carcinoma, paired single-cell RNA and chromatin-accessibility profiles separated epithelial-mesenchymal-transition states, while CRISPR/Cas9 loss-of-function experiments supported Klf5 and Pitx1 as regulators of early EMT and metastasis and Nfatc1 and Creb3l1 as later-stage regulators. In melanoma, quantitative state diagrams derived from regulatory networks identified several resistance routes, including dedifferentiation and a hyperpigmented state, and proposed combination strategies that reduced predicted occupancy by resistant states. The latter remained primarily computational, whereas the squamous-cell-carcinoma findings included functional perturbation. This contrast illustrates why a network can be useful for generating mechanistic hypotheses even when its proposed interventions have not yet been tested experimentally.

Resistance, lineage plasticity and tumor classification provide further applications. In patient-derived glioma stem-like cells, inferred network topology distinguished trajectories through resistant and sensitive intermediate states under drug treatment, and simulations guided siRNA or combination approaches that increased killing in the preclinical system. A TCF4-dependent network in melanoma was linked to immunotherapy resistance and supported by functional perturbation. Single-cell and bulk integration distinguished two epithelial tumor states in colorectal cancer, while approaches in hepatocellular carcinoma identified immune-activated, metabolic and proliferative terminal states; the proliferative signature was associated with worse overall survival but had not been prospectively validated. Other studies resolved age-related regulatory signatures in acute myeloid leukemia, progression-associated epigenetic programs in lung adenocarcinoma, a BAP1-associated interferon program in clear-cell renal cell carcinoma, FOXA1-driven lineage transformation in prostate cancer and carcinogenesis-associated plasticity in gastric-cell isthmus states.

Networks are also being extended beyond malignant cells. Multi-omics in triple-negative breast cancer identified a myofibro-inflammatory program in cancer-associated fibroblasts, indicating that stromal states are maintained by coordinated regulation. Future spatial and multi-omics methods will need to separate cell-intrinsic networks from ligand-receptor effects and incorporate cell neighborhoods and physical proximity. Larger ambitions include adding DNA methylation, histone modifications, spatial information and temporal measurements, combining foundation models with virtual perturbation and CRISPR screens, and building benchmarks across tumor types and platforms. Foundation models such as GET, EpiAgent, RegFormer and ChromBERT may improve transfer across sparse cell states, but attention-derived edges can add little beyond gene-level statistics, and internal feature relations do not establish biological causation. The review is a structured expert synthesis rather than a systematic review or meta-analysis, and the field’s decisive next step is not simply more elaborate architectures but shared benchmarks, uncertainty estimates, transparent reporting and causal validation in organoids, primary samples and clinically annotated cohorts.

Subject of Research: AI-assisted reconstruction and validation of cell-state-specific gene regulatory networks in cancer

Article Title: AI and computational methods for reconstructing cell-state-specific gene regulatory networks from single-cell transcriptome and chromatin accessibility data in cancer: methods, applications, and perspectives

Article References: Wang, J., & Chen, F. (2026). AI and computational methods for reconstructing cell-state-specific gene regulatory networks from single-cell transcriptome and chromatin accessibility data in cancer: methods, applications, and perspectives. Frontiers in Genetics, 17, Article 1982712. https://doi.org/10.3389/fgene.2026.1982712

Image Credits: AI Generated

DOI: 10.3389/fgene.2026.1982712

Keywords: single-cell multi-omics, gene regulatory networks, cancer cell states, artificial intelligence, chromatin accessibility, transcriptomics, causal validation, drug resistance, tumor heterogeneity, virtual perturbation, computational, methods

The motivation behind the work lies in a paradox that has emerged from large-scale plasma proteomics studies. Over the past several years, researchers measuring thousands of proteins in blood samples from tens of thousands of individuals have shown that these molecular fingerprints can estimate chronological age with striking accuracy. Yet a protein that predicts age is not necessarily a protein that drives ageing. Correlation, as the field has repeatedly learned, does not equal causation, and the leap from a predictive biomarker to a druggable target requires layers of evidence that simple statistical associations cannot supply. PEN was designed specifically to make that leap more systematic, transparent, and reproducible.

Architecturally, PEN unfolds in six sequential stages, each of which adds a distinct form of biological reasoning. The pipeline begins with a plasma proteome input, typically a matrix of protein abundance measurements across many individuals. In the second stage, a multilayer perceptron, a classic type of deep neural network, is trained to predict chronological age from the protein measurements. In the demonstration study, the team applied this stage to data from 44,179 UK Biobank participants profiled for approximately 2,920 plasma proteins. The model achieved a test coefficient of determination of 0.8723, a Pearson correlation of 0.934, and a mean absolute error of just 2.30 years, confirming that the proteomic age clock was performing at a level comparable to the best published proteomic predictors.

The third stage is where PEN begins to move beyond prediction. Using a technique called Integrated Gradients, an attribution method originally developed for interpreting deep learning models, the researchers identified which proteins contributed most strongly to the age predictions and characterized the functional modules those proteins participate in. Integrated Gradients works by accumulating the gradients of the model’s output with respect to its inputs as features are gradually shifted from a baseline to the actual value, yielding a per-protein importance score. This step converts the opaque neural network into a ranked list of diagnostic proteins, providing the raw material for the target discovery stages that follow.

Stage four introduces the graph-based core of the framework. The diagnostic proteins are mapped onto a protein-protein interaction network, a vast molecular map in which nodes represent proteins and edges represent known physical or functional interactions. Graph convolutional propagation and diffusion algorithms then spread information across this network, allowing the framework to identify candidate targets that are not merely important on their own but are embedded in neighborhoods of ageing-relevant biology. This network context matters because biological systems rarely act through single molecules; perturbing one node of a densely connected module can ripple through entire pathways. By propagating evidence through the interactome, PEN surfaces candidates that a purely statistical analysis would likely overlook.

The fifth stage applies a discovery-oriented reranking that integrates multiple components of biological evidence, weighing each candidate against additional criteria before a final prioritization is issued. Applied to the UK Biobank plasma proteomics data, this process produced a shortlist of candidate mechanistic target hypotheses that includes FBN1, FBLN5, EDA, RLN3, LHCGR, COL14A1, HLA-E, and TSPAN4. Several of these names will be familiar to students of ageing biology. FBN1 encodes fibrillin-1, a structural component of elastic fibers, and FBLN5 encodes fibulin-5, another extracellular matrix protein essential for elastic fiber assembly, both of which connect plausibly to the vascular stiffening and tissue degeneration characteristic of ageing. HLA-E, an immune regulatory molecule, hints at the inflammatory and immunological dimensions of the ageing process.

The sixth and final stage provides what the authors describe as orthogonal in silico support through virtual perturbation. Here the framework turns to Geneformer, a foundation model of gene regulatory networks trained on large corpora of single-cell transcriptomic data. Geneformer can simulate, computationally, what happens to a cell’s transcriptional state when a particular gene’s activity is reduced or abolished, a process the researchers call virtual perturbation. The team applied this approach to an independent dataset, GSE130973, a publicly available single-cell RNA sequencing dataset of ageing human skin. By asking whether virtual perturbation of each candidate gene shifts cells from an OLD transcriptional state toward a YOUNG one, or vice versa, the framework obtained an independent line of computational evidence that is entirely separate from the plasma proteomics data used to generate the candidates.

The results of this final stage were instructive in both their successes and their limitations. After recalibrating the analysis using an expanded null distribution of 500 random token-valid genes, FBLN5 and HLA-E showed significant positive OLD-to-YOUNG transcriptional state shifts in specific cell populations, strengthening the case that these two proteins merit experimental attention. Other candidates, including several on the original shortlist, displayed weaker or context-dependent effects, a reminder that computational prioritization is a hypothesis-generating exercise rather than a guarantee of biological validity. The authors are explicit on this point: the Geneformer perturbation stage strengthens biological plausibility but does not constitute experimental validation. Laboratory work with cell models, organoids, or animal systems remains essential before any of these candidates can be considered genuine therapeutic targets.

What makes PEN notable is less any single algorithmic component than the discipline of its overall design. Each stage is modular and auditable, and the framework is explicitly intended to be reproducible and generalizable. The authors argue that the same six-stage logic, proteomic input, deep learning prediction, interpretability analysis, graph-based propagation, evidence-weighted reranking, and foundation-model perturbation testing, could be applied to other biomarker-driven target discovery problems far beyond ageing, from cardiometabolic disease to neurodegeneration. The study was conducted using the UK Biobank Resource under Application Number ID200882 and was supported by the Zhangjiang Special Funding Major Project 2024, an initiative titled AI-driven and multi-ancestry evaluated protein therapeutic discovery system, reflecting a broader industrial and academic push to convert population-scale omics data into actionable drug discovery pipelines.

The broader significance of the work lies in its timing. Proteomic age clocks are proliferating, foundation models for biology are maturing rapidly, and the pharmaceutical industry is increasingly willing to pursue ageing itself, or age-related frailty, as a therapeutic domain. What has been missing is a principled way to connect these pieces, to move from a blood test that knows how old you are to a molecule worth targeting in the clinic. PEN offers one template for that connection, complete with built-in humility about what computation can and cannot establish. If candidates such as FBLN5 and HLA-E survive the crucible of experimental validation, the framework that nominated them may well become a standard fixture in the computational geroscience toolkit, and the era in which a blood draw could point directly toward an anti-ageing intervention will have moved one decisive step closer.

Subject of Research: A graph-guided computational framework for prioritizing ageing-associated drug target hypotheses from plasma proteomics data.

Article Title: Protein Expression Net: an integrated graph-guided computational framework and virtual perturbation approach for prioritizing ageing-associated target hypotheses from plasma proteomics

Article References: Wu, W., Jiang, Y., Wei, K., Lin, Z., Gu, R., Wei, S., Wang, Z., Fang, L., Wang, X., Fu, X., Wang, Y., & Pan, L. (2026). Protein Expression Net: an integrated graph-guided computational framework and virtual perturbation approach for prioritizing ageing-associated target hypotheses from plasma proteomics. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06645-3

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06645-3

Keywords: ageing, plasma proteomics, UK Biobank, deep learning, graph neural network, Integrated Gradients, Geneformer, virtual perturbation, target prioritization, BMC Bioinformatics, protein-protein interaction network, biomarkers