Executive Summary: AI
The study leverages an AI‑driven multimodal transcriptomic pipeline to resolve the heterogeneity of cellular senescence (CS) in breast cancer, thereby establishing a causal chain that links senescence‑associated gene perturbations to patient prognosis and tumor aggressiveness. By employing GeneCompass‑based all‑gene in silico perturbation, the authors expanded the repertoire of senescence‑related genes and, through supervised machine learning, distilled a seven‑gene prognostic signature that quantitatively predicts survival outcomes across independent cohorts.
- Core scientific discovery and governing mechanism. The central finding is that senescence‑related transcriptional reprogramming, captured through comprehensive in silico perturbation, delineates a risk axis driven principally by the adhesion G‑protein‑coupled receptor ADGRF5. Knockdown of ADGRF5 attenuates malignant phenotypes in breast cancer cells, indicating that ADGRF5 functions as a pivotal effector linking senescent cellular states to tumor progression.
- Primary empirical benchmark. The prognostic utility of the seven‑gene signature was validated with time‑dependent receiver operating characteristic (ROC) analyses: in the training cohort the area under the curve (AUC) values were 0.707, 0.700, and 0.684 at 1, 3, and 5 years, respectively; in the test cohort they were 0.657, 0.661, and 0.629; and in an external validation cohort they were 0.611, 0.646, and 0.637. Single‑cell and spatial transcriptomics further revealed that the risk component reflects malignant epithelial cell states and concurrent stromal‑vascular remodeling.
- Strategic significance and practical implications. This AI‑assisted framework furnishes a reproducible, cross‑cohort prognostic tool that can be integrated into clinical decision‑making to stratify breast cancer patients by senescence‑driven risk. Moreover, the identification of ADGRF5 as a functional driver opens avenues for targeted therapeutic development, while the broader methodological pipeline offers a template for senescence biomarker discovery in other malignancies, potentially accelerating precision oncology and informing drug‑development pipelines.
Theoretical Framework & Mechanisms: AI
The analytical pipeline rests on three interlocking theoretical constructs: (i) the quantitative representation of cellular senescence as a high‑dimensional transcriptional state, (ii) the in silico perturbation framework embodied by GeneCompass, and (iii) the survival‑risk modeling that translates a multigene expression vector into a time‑dependent prognostic function. Each component is grounded in established biophysical and statistical principles, enabling the integration of bulk, single‑cell, and spatial transcriptomic modalities without invoking undocumented methodological specifics.
1. Transcriptional Manifold of Cellular Senescence
Cellular senescence (CS) is modeled as a latent variable $\sigma\in\mathbb{R}$ that modulates the expression of $G$ genes. For a given sample $i$, the observed log‑TPM vector $\mathbf{x}_i\in\mathbb{R}^G$ is expressed as
$$\mathbf{x}_i = \boldsymbol{\mu} + \sigma_i\boldsymbol{\beta} + \boldsymbol{\varepsilon}_i,$$where $\boldsymbol{\mu}$ denotes the basal transcriptional baseline, $\boldsymbol{\beta}$ the senescence‑associated direction in gene‑space, and $\boldsymbol{\varepsilon}_i$ a multivariate Gaussian noise term ($\boldsymbol{\varepsilon}_i\sim\mathcal{N}(\mathbf{0},\Sigma)$). The vector $\boldsymbol{\beta}$ is inferred from the perturbation signatures generated by GeneCompass, ensuring that the senescence axis captures coordinated up‑ and down‑regulation patterns observed across perturbation experiments.
2. GeneCompass In Silico Perturbation
GeneCompass constructs a perturbation matrix $\mathbf{P}\in\mathbb{R}^{G\times K}$, where each column $k$ corresponds to the differential expression profile induced by a virtual knock‑down or over‑expression of gene $k$. The perturbation score for a target gene $g$ in sample $i$ is defined as the inner product
$$s_{ig}= \mathbf{x}_i^\top \mathbf{p}_g,$$with $\mathbf{p}_g$ the $g$‑th column of $\mathbf{P}$. Positive $s_{ig}$ indicates that the observed expression pattern aligns with the simulated senescence‑inducing effect of $g$, whereas negative values suggest a rejuvenating influence. By ranking $s_{ig}$ across all $g$, the method expands the canonical senescence gene set to include candidates whose simulated perturbations most strongly drive the latent senescence variable $\sigma$.
3. Multimodal Integration and Feature Selection
Bulk transcriptomic profiles, single‑cell RNA‑seq (scRNA‑seq) embeddings, and spatial transcriptomic spot matrices are each projected onto a common latent space using canonical correlation analysis (CCA). Let $\mathbf{Z}^{(m)}$ denote the CCA‑derived representation for modality $m\in\{bulk, sc, spatial\}$. The concatenated feature matrix $\mathbf{Z}=[\mathbf{Z}^{(bulk)};\mathbf{Z}^{(sc)};\mathbf{Z}^{(spatial)}]$ serves as input to a regularized Cox proportional hazards model,
$$h(t|\mathbf{Z}_i)=h_0(t)\exp\!\bigl(\boldsymbol{\theta}^\top \mathbf{Z}_i\bigr),$$where $h_0(t)$ is the baseline hazard and $\boldsymbol{\theta}$ the coefficient vector. An $L_1$ penalty ($\lambda\|\boldsymbol{\theta}\|_1$) enforces sparsity, yielding a seven‑gene signature whose coefficients directly quantify each gene’s contribution to risk.
4. Time‑Dependent Prognostic Evaluation
The predictive accuracy is assessed via the cumulative/dynamic ROC framework. For a given horizon $\tau$, the time‑dependent true‑positive rate (TPR) and false‑positive rate (FPR) are
$$\text{TPR}(\tau)=\Pr\bigl\{ \hat{R}_i > c \mid T_i\le\tau\bigr\},\qquad \text{FPR}(\tau)=\Pr\bigl\{ \hat{R}_i > c \mid T_i>\tau\bigr\},$$where $\hat{R}_i=\exp(\boldsymbol{\theta}^\top \mathbf{Z}_i)$ is the risk score, $T_i$ the observed survival time, and $c$ a decision threshold. The area under the ROC curve, $AUC(\tau)$, integrates TPR and FPR over all $c$, providing the reported time‑dependent AUCs at 1, 3, and 5 years.
5. ADGRF5 Signaling as a Functional Node
ADGRF5, an adhesion G‑protein‑coupled receptor, is positioned on the senescence axis by virtue of its high perturbation score and coefficient magnitude in the Cox model. In a deterministic signaling approximation, receptor activation $R(t)$ follows
$$\frac{dR}{dt}=k_{\text{on}}[L] (1-R) - k_{\text{off}}R,$$where $[L]$ denotes ligand concentration, and $k_{\text{on}}$, $k_{\text{off}}$ are binding/unbinding rates. Downstream proliferative output $P(t)$ can be linked to $R(t)$ through a linear cascade $dP/dt=\alpha R - \beta P$, illustrating how ADGRF5 modulation translates into altered malignant phenotypes observed experimentally.
Collectively, these formal relationships resolve prior bottlenecks by (i) quantifying senescence heterogeneity through a perturbation‑driven latent variable, (ii) unifying disparate transcriptomic modalities via CCA, and (iii) delivering a rigorously validated, time‑resolved prognostic metric that is directly interpretable in terms of underlying molecular drivers such as ADGRF5.
Empirical Evidence & Methodology: AI
Experimental Design and Cohort Composition
The investigation employed a multimodal transcriptomic pipeline that integrated bulk RNA‑sequencing, single‑cell RNA‑sequencing, and spatial transcriptomics to interrogate cellular senescence states across breast cancer specimens. The primary analytical workflow began with a GeneCompass‑based all‑gene in‑silico perturbation screen to flag candidate senescence‑inducing or rejuvenating genes. Subsequently, a supervised machine‑learning model distilled a seven‑gene prognostic signature from the perturbation outputs and transcriptomic profiles.
Training, Validation, and External Testing
Model training and internal validation were performed on a designated training cohort, with performance quantified by time‑dependent area‑under‑the‑curve (AUC) metrics at 1, 3, and 5 years post‑diagnosis. The resulting AUCs were 0.707, 0.700, and 0.684, respectively. A separate test cohort yielded AUCs of 0.657, 0.661, and 0.629 across the same time points, demonstrating consistent predictive capacity. External validation on an independent cohort produced AUCs of 0.611, 0.646, and 0.637, confirming generalizability across diverse patient populations.
Single‑Cell and Spatial Transcriptomic Insights
Single‑cell and spatial analyses revealed that the risk component encoded by the seven‑gene signature is not confined to malignant epithelial cells alone. Instead, it reflects a composite of epithelial senescence states and stromal–vascular remodeling within the tumor microenvironment. This dual contribution underscores the heterogeneity of senescence phenotypes and their prognostic relevance.
Functional Validation of ADGRF5
Among the signature constituents, ADGRF5 displayed the most pronounced differential expression. Functional interrogation via gene knockdown demonstrated a suppression of malignant phenotypes in breast cancer cell lines, implicating ADGRF5 as a potential driver of tumor aggressiveness. Although the specific assays employed were not detailed in the abstract, the observed phenotype aligns with the computational prediction of ADGRF5’s role in senescence‑related tumor progression.
Statistical Validation and Control Baselines
Performance metrics were evaluated using time‑dependent ROC analysis, with AUC values serving as the primary statistical benchmark. The study’s design inherently includes internal cross‑validation within the training set and external validation on independent cohorts, providing robust control baselines against which the prognostic signature’s predictive power was assessed.
Lead Authors: Wenhao Liu, Wenhui Wu, Shubai Chen, Kaiqiong Chen, Xin Li
Affiliations: College of Life Science, Northeast Agricultural University, Harbin 150030, China; State Key Laboratory of Organ Regeneration and Reconstruction, Institute of Zoology, Chinese Academy of Sciences, Beijing 100101, China; Beijing Institute for Stem Cell and Regenerative Medicine, Beijing 100101, China; Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100101, China
Publication Venue: Cells (2026)
DOI: 📄 DOI: 10.3390/cells15171589
Scientific Impact & Future Horizons: AI
While the AI‑assisted multimodal transcriptomic strategy delineated a robust senescence‑related prognostic signature and spotlighted ADGRF5 as a functional driver of malignant phenotypes, several methodological constraints temper immediate clinical translation. First, the reliance on retrospective bulk and single‑cell datasets limits causal inference; longitudinal sampling across treatment courses would clarify whether the signature captures dynamic senescence states or static tumor subclones. Second, the computational perturbation framework, though powerful, remains agnostic to post‑transcriptional regulation and epigenetic remodeling that may modulate senescence outputs. Third, the signature’s performance, though statistically significant, exhibits modest discrimination (AUCs < 0.71), suggesting that integration of complementary modalities—proteomics, metabolomics, and spatial imaging—could enhance predictive fidelity.
Future research should pursue multi‑omics longitudinal cohorts, enabling the dissection of senescence trajectories in response to chemotherapy, endocrine therapy, and emerging senolytic agents. Functional interrogation of ADGRF5 through CRISPR‑Cas9 mediated genome editing and pharmacologic inhibition in patient‑derived organoids will ascertain its therapeutic tractability and potential off‑target effects. Moreover, elucidating the crosstalk between senescent stromal cells and malignant epithelial compartments via spatially resolved transcriptomics could reveal paracrine circuits amenable to intervention.
In translational terms, the seven‑gene signature offers a pragmatic adjunct to existing prognostic panels, potentially guiding adjuvant therapy decisions and surveillance intensity. The ADGRF5 axis, if validated as a driver, could serve as a biomarker for patient stratification in trials of GPCR‑targeted therapeutics or senolytic regimens. Regulatory frameworks will need to evolve to accommodate composite omics signatures, ensuring reproducibility, standardization of assay platforms, and clear clinical utility thresholds.
Core Takeaways
- Fundamental Mechanism: Senescence heterogeneity, mediated in part by ADGRF5 signaling, orchestrates tumor‑microenvironment remodeling that underlies differential patient outcomes.
- Real-World Value: A clinically deployable seven‑gene panel and a targetable GPCR pathway provide actionable insights for prognosis and personalized therapy in breast cancer.
Applications & Future Outlook
Implementation of the signature in routine pathology workflows could refine risk stratification, while ADGRF5 inhibitors may complement existing chemotherapeutics. Open challenges include validating the signature across diverse ethnic cohorts, integrating multi‑omics data to boost predictive accuracy, and establishing safety profiles for ADGRF5‑directed agents.
- Liu, W., Wu, W., Chen, S., Chen, K., & Li, X. (2026). AI‑Assisted Multimodal Transcriptomic Analysis Identifies a Senescence‑Related Prognostic Signature and Characterizes ADGRF5‑Associated Malignant Phenotypes in Breast Cancer. Cells, 15(17), 1589. 📄 DOI: 10.3390/cells15171589
- Kuilman, T., & Peeper, D. S. (2014). The role of senescence in cancer. Cell, 157(5), 1040–1048. 📄 DOI: 10.1016/j.cell.2014.03.022
💬 Comments