SIMD: Synergistic integration mutualistic platform based on single-cell and proteotranscriptomics for drug repositioning

Abstract

This study aimed to introduce a synergistic integration platform based on single-cell and proteotranscriptomics for drug repositioning (SIMD) to repurpose breast cancer drug candidates by considering the bidirectional nature of molecular dynamics in protein transcriptomics and tumor heterogeneity at the single-cell level. SMID consists of two core methods: the anti-correlated proteotranscriptome perturbation score (ACPS), which measures the negative correlation between perturbagen-induced expression signatures and multi-omic expression profiles from patient-derived breast cancer samples; and perturbagen repositioning scoring and scRNA-seq data (PPNE), which ranks perturbagens exhibiting high ACPS scores and inverse transcriptomic associations with neoplastic epithelial (NE) cells identified through single-cell RNA sequencing data. We validated the top five SIMD-prioritized perturbagens by assessing cell viability in five breast cancer cell lines. Many perturbagens ranked highest by the ACPS have been ascertained in clinical trials for breast cancer or other cancer treatments. Validation revealed a substantial reduction in cell viability for the majority of perturbagens in most cell types. Our result observed significantly overexpressed phosphorylation sites within the AKT1 and PI3K genes, which are buparlisib target genes, suggesting that buparlisib may exert its effects across a spectrum of breast cancer types, including triple-negative breast cancer, by modulating molecular mechanisms within the PI3K pathway.

Introduction

Breast cancer exhibits pronounced cellular and molecular heterogeneity, which contributes to its diverse clinical presentations and unpredictable drug responses^1–3. Drug repositioning, which involves exploring the therapeutic potential of clinically approved and off-patent compounds, has emerged as a cost-effective and time-efficient strategy in drug development⁴. This approach has been widely adopted in diverse studies aimed at applying drug repositioning across various disease models, including anticancer therapies (e.g., breast cancer⁵), immunotherapy targets⁴, and infectious diseases, such as COVID-19⁶.

Large-scale pharmacogenomic initiatives have systematically profiled the responses of genetically diverse human cancer cell lines to hundreds of therapeutic agents and integrated these data with multi-omics annotations, thereby advancing the identification of predictive biomarkers and uncovering previously unrecognized therapeutic efficacies via computational drug repositioning approaches. The GDSC database provides open-access drug response data from ~75,000 experiments involving 138 anticancer agents and ~700 genetically characterized cancer cell lines⁷. The CCLE offers comprehensive multi-omics data for 1457 human cancer cell lines⁸, and the DMAP project contributes foundational resources by extensively mapping gene dependencies in molecularly annotated cancer models using high-throughput RNA interference and CRISPR-Cas9–based functional genomics approaches⁹. The LINCS L1000 project profiles transcriptomic responses to >20,000 perturbagens across ~1000 cell lines, serving as a critical resource for drug repositioning¹⁰. This dataset has since enabled a wide range of successful studies, leading to the discovery of novel drug–disease associations and therapeutic opportunities. For instance, Mohammadi et al. proposed BRD-A06641369 and BRD-K97926541 as immunomodulatory agents that downregulate B2M and secretory leukocyte protease inhibitor (SLPI) in breast cancer epithelial cells based on single-cell transcriptomic data¹¹. Other studies have reported KM-00927 and BRD-K75081836 as novel histone deacetylase inhibitors, and mitomycin C as a topoisomerase IIB inhibitor¹¹ In a separate analysis, six candidate compounds, including the pan-Akt inhibitor, patasertib (GDC-0068), were identified for their potential to enhance the efficacy of anticancer peptides (ACPs)⁴.

Despite the availability of large-scale datasets and advanced analytical tools, most breast cancer drug candidates have failed to show meaningful clinical benefits, owing to the molecular heterogeneity of the disease and its negative impact on treatment response^12,13.

To overcome these challenges, research has been conducted in two key areas. First, the rapid advancement of single-cell sequencing has enabled deeper investigation of cancer heterogeneity, resistance, and etiology, leading to drug repositioning efforts that leverage this granularity¹. Studies have shown that rare drug-tolerant cells with distinct transcriptional profiles can coexist within tumors, potentially contributing to resistance^2,3. This approach was exemplified by Mohammadi et al., who used nonmalignant epithelial cells identified via single-cell transcriptomics to reposition drugs for breast cancer immunotherapy¹⁴. Second, disease phenotypes are shaped by complex regulatory mechanisms acting across all central dogma processes, highlighting the need for integrative multi-omics analysis that reflects this molecular complexity¹². Conventional DR strategies have long sought to address a critical limitation: the reliance on single-omic data to identify new indications for approved drugs¹⁵. To overcome the limitations of single-omic approaches, multi-omics integration using CPTAC and TCGA has been used to discover or repurpose anticancer drugs^16,17. However, the advancement of multi-omics-based computational DR approaches remains hindered by critical bottlenecks, including limited availability of proteomic data and the absence of robust integrative methodologies, often leading to reliance on partial omics data or literature-derived auxiliary metadata^{12,13,18–20}.

Leveraging the L1000 resource, we present SIMD, a framework that integrates proteotranscriptomic signaling with single-cell-resolved tumor heterogeneity for anti-breast cancer drug discovery. It incorporates two core modules: ACPS, which quantifies the inverse correlation between perturbagen-induced signatures and omics expression profiles from patient-derived tumor tissues, and perturbagen repositioning scoring and scRNA-seq data (PPNE), which prioritizes high-ACPS perturbagens that also show inverse transcriptomic correlation with neoplastic epithelial (NE) cells from scRNA-seq data. SIMD analysis identified five candidate compounds, including the pan-PI3K inhibitor Buparlisib, which exhibited the strongest negative correlation with NE cells, and we validated their therapeutic potential through experimental assays.

Results

ACPS and validation

To generate 15,207 PSPs, we first merged the proteotranscriptome expression value (PTE) harboring the intrinsic properties of breast cancer from TCGA/CPTAC with the expression data perturbed by perturbagens from L1000 and containing a log₂ fold change of 7159 genes. Next, we calculated the ACPS of 15,207 PSP. The process of calculating the scoring strategy consists of the following three steps (Fig. 1A): A higher ACPS indicates a stronger anti-correlated characteristic of perturbagens, whereas a lower score suggests a weaker anti-correlation. By mapping the major ingredients of 77 prescribed drugs for patients with breast cancer from TCGA to their corresponding CIDs from PubChem, and using the list as the gold standard for AUC calculations, we compared the effectiveness of ACPS across different data types, such as proteomics [P], RNA-seq [R], and proteotranscriptomics [PR] (a combination of both). The average AUC for each p-value threshold was notably higher for the PR data type (0.708), followed by P (0.504), and R (0.670), indicating that our scoring model was better suited for combined data types than for individual data types (Fig. 1B, Data File S1–S6). To consider various aspects, such as chemical structure similarity, protein–protein interactions (PPI), and biological function, while evaluating the ACPS method per data type, we determined that a p-value threshold of 0.04 was the optimal parameter based on the above-mentioned results. To determine the structural similarity between anti-breast cancer drugs and perturbagens, we calculated the proportion of perturbagens with a Tanimoto similarity higher than the cutoff for each category. The LP, LR, LT, HP, HR, and HT groups were assigned 182, 186, 193, 120, 116, and 109 unique CIDs, respectively. Among the designated categories, the HT category showed the highest average proportion of 0.302 at the 0.5–1.0 threshold (Fig. 1C). The target gene list and bioassay data comprising bioactivity and toxicity information for each perturbagen were collected from DrugBank and PubChem, respectively. A t-test was conducted to evaluate the significance of the mean difference in the PPI scores of STRINGDB between the high- and low-perturbagen groups for a given perturbagen as the data type, which exhibited the lowest p-values for all interaction score thresholds (150, 400, 700, and 900) of STRINGDB (Fig. S1). After conducting gene set enrichment analysis (GSEA) using the log₂ fold change of a perturbagen and computing the proportion of BCRG among the significantly enriched gene sets (with a p-value < 0.05) for each class, we observed that the DP data type had an average BCRG inclusion ratio of 0.85. In contrast, the HT dataset had the highest ratio among all gene set categories at 2.80 (Fig. 1D). To validate the credibility of perturbagens from both clinical and biological aspects, a meta-analysis was conducted on 96 of the 109 perturbagens in the HT class for which the target gene information could be retrieved from the ChEMBL database (details in the Methods section). Remarkably, 45.83% of all perturbagens are involved in ongoing cancer-related clinical trials (Fig. 1E), whereas 28.33% are under investigation in clinical trials to assess their efficacy in breast cancer treatment (Fig. 1F). Additionally, 54.13% of the target genes associated with perturbagens in the HT group were identified as genes correlated with breast cancer (Fig. 1G). Taken together, these findings indicate that applying the HT class to screen for perturbagens by calculating the ACPS results in a high concordance rate with diverse biological phenomena and clinical trial statements for breast cancer.

Fig. 1. Overview of the proposed SIMD platform and performance assessment.

A Graphical synopsis of SIMD. B Performance was compared by calculating the AUC for each p-value threshold within each data type. High protein (HP), high RNA (HR), high proteotranscriptome (HT), low protein (LP), low RNA (LR), and low proteotranscriptome (LT). C Proportion of perturbagens exhibiting structural similarities to anti-breast-cancer drugs. D Mean count of BCRG within significantly enriched gene sets from the GSEA-based pathway and GO analysis. Proportion of perturbagens within the HT group in the included clinical trials for cancer (E) or breast cancer (F) at any stage of progression. G Fraction of perturbagen-targeted genes associated with breast cancer.

NE cell isolation and breast cancer trait verification

The scRNA-seq data (PPNE) framework comprises two principal stages. First, using the method and publicly available code proposed by Wu et al.²¹, we distinguished NE cells from normal epithelial NNE cells by analyzing CNV expression patterns derived from single-cell RNA sequencing data of patients with breast cancer. Second, we identified perturbagens whose response transcriptomes exhibited inverse correlations with the gene expression profiles of isolated NE cells. To select perturbagens meticulously, we computed the correlations between NE cells derived from scRNA sequencing data obtained from eight primary breast cancer tumors and the gene expression profiles of each perturbagen classified under the HT category. In total, 32,550 cells that met the quality control criteria were divided into 21 transcriptionally distinct clusters. To establish their cellular identities, the cells were subjected to a rigorous classification process, resulting in the classification of six predominant cell types, B, T, endothelial, ependymal, myeloid, and plasma cells, by leveraging specific cell-type markers, which are elucidated in Data File S7 (Fig. 2A and Figs. S2–3). To assess the unique features exhibited by the major cell types, we evaluated specific canonical cell markers including epithelial (EPCAM), T (CD3D), myeloid (CD68), B (MS4A1), plasma (JCHAIN), and endothelial (ACVRL1) cells (Fig. 2B). We validated the characteristic features of the epithelial cells by meticulously analyzing the expression patterns of crucial breast cancer susceptibility marker genes. Remarkably, a substantial subset of the 11 genes investigated, including ERBB2 and ESR1, exhibited a prevailing expression pattern specific to epithelial cells, surpassing that of their cellular counterparts (Fig. 2C).

Fig. 2. Characteristic validation of NE cells isolated from scRNA-seq analysis of primary breast cancer.

A UMAP visualizations of 32,550 cells analyzed using integrated scRNA-seq data across eight primary breast tumors. Expression patterns of representative marker genes. B Crucial breast cancer susceptibility marker genes. C Marker genes per cell type: epithelial cells (EPCAM), T cells (CD3D), myeloid cells (CD68), B cells (MS4A1), plasma cells (JCHAIN), and endothelial cells (ACVRL1). D UMAP projections of NE and NNE cell types isolated from epithelial cells. E Mean differences in CNV expression between NE and NNE cells. The four asterisks denote p-values < 0.0001. F GSEA pathway enrichment analysis in DEGs within NE cells. G–H Expression of representative marker genes in breast cancer and normal breast epithelial cells of each cell type. I Single-cell GSEA analysis based on DEGs in NE cells. The four asterisks denote p-values < 0.0001.

We distinguished NE cells from their normal counterparts by characterizing the CNV expression patterns derived from scRNA-seq data, adopting the approach proposed by Wu et al. (2021) (Fig. 2D, E and Fig. S3A)²². We identified 5370 cells classified as NE cells and 3318 cells categorized as NE and NNE cells; each cell type contained an average gene count of 2456 and 2647 per cell, respectively (Fig. S4B). Consistent with our expectations, NE cells exhibited a notably higher CNV expression than NNE cells (Fig. 2E). Furthermore, additional similarity analysis of NNE cells using public epithelial datasets—normal epithelium from breast cancer (NOR_EPI, GSE161529) and breast cancer epithelium (BC_EPI, GSE176078) revealed that NNE cells showed the highest transcriptional resemblance to NOR_EPI, supporting their closer association with normal epithelial identity and strengthening the robustness of our findings(Fig. S9B–D). To ascertain biological parallels between NE cells and breast cancer, we examined the expression of representative markers associated with breast cancer progression, including KRT14²³ and SLP1²⁴, and markers characteristic of normal breast epithelial cells, including GATA3²⁵ and TFF3¹⁵, in each cell type. All markers indicative of breast cancer progression were specifically expressed in NE cells (Fig. 2F). Conversely, most markers characteristic of normal breast epithelial cells were predominantly expressed in NNE cells, with the exception of COXSC (Fig. 2G). In addition, we conducted a GSEA using the MsigDB C2 and C5 databases. Several enriched pathways were associated with various breast cancer subtypes (Fig. 2H and Data File S8). Substantial enrichment was identified in the areas of epithelial development, epithelial cell differentiation, biological adhesion, cell migration, regulation of protein kinase activity, and cell-substrate adhesion (Fig. S5A and Data File S9). The GO category of cellular components was significantly enriched in endocytic vesicle lumen, external encapsulating structure, high-density lipoprotein particles, protein-lipid complexes, and vesicle lumen terms (Fig. S5B and Data File S10). Finally, within the realm of biological processes and molecular functions, notable enrichment was detected in the binding functions of calcium ions, calcium-dependent proteins, G-protein-coupled receptors, macrolides, and S100 proteins (Fig. S5C and Data File S11). Most of these biological functions play pivotal roles in cancer cell proliferation, differentiation, and metastasis, thereby influencing breast cancer regulation^16,17,26,27. Single-cell GSEA analysis based on the DEGs of NE cells revealed that the p53 pathway, recognized for its crucial role in the development and progression of breast cancer, exhibited significantly elevated mean enrichment scores in NNE cells compared to those in NE cells²⁸. Furthermore, the Wnt signaling pathway, which is primarily implicated in the intricate mechanisms underlying metastasis and resistance development in TNBC, as revealed by an influential study, exhibited a significantly high mean enrichment score in NE cells²⁹ (Fig. 2I).

Potent perturbagen efficacy validation in NE-like breast cancer SC-cells

Among the expression data of the 109 perturbagens within the HT class, 49 perturbagens that had a negative correlation with NE cells, as opposed to NNE cells, were the highest among the various data types considered in our analysis (Fig. 3A). We narrowed our focus to five perturbagens with a high potential as breast cancer drug candidates for experimental validation (Table 1). A-443654 acts as a pan-Akt inhibitor³⁰ and inhibits breast cancer cell proliferation³¹; buparlisib is an anticancer agent that targets the PI3K pathway, and vemurafenib is a kinase inhibitor that specifically targets the B-RAF V600 mutation^32,33. Remarkably, most perturbagens acted as inhibitors of primary targets or inhibitors of key signaling pathways, such as PI3K and Akt, which are pivotal in breast cancer progression. In addition, by analyzing the expression of target genes in each cell at the single-cell level, we found that the target genes PI3KCB, AURKB, and AKT3 of buparlisib, tozasertib, and A-443654, respectively, were significantly overexpressed in NE cells (Fig. 3B). Moreover, four genes previously identified to mediate the activation of the respective drugs, as indicated by PubChem bioassay data, were exclusively expressed in NE cells, including LEMD1³⁴, a recently described target in TNBC, and KRT15³⁵, which is implicated in the regulation of breast tissue invasiveness. To identify the cell lines exhibiting the most similar marker patterns to NE cells and to evaluate the efficacy of the selected perturbagens, scRNA-seq data were gathered and analyzed for 32 breast cancer cell lines, comprising 31 cancerous cells and one normal cell line (GSE173634), with 33,123 cells (Fig. S6A–J). We examined the expression patterns of seven genes among the DEGs identified through a comparative analysis of 31 cell lines in relation to MCF12A cells, which serve as a pseudo-epithelial cell line. The expression patterns of KRT17, KRT14, and KRT5 were dissimilar, whereas ERBB2, GABRP, and CDH1, genes recognized as representative markers of breast cancer, exhibited expression levels comparable to those in NE cells (Fig. S6K). In addition, genes, including ERBB2, were expressed at the single-cell level in representative breast cancer cell lines (Fig. 3C).

Fig. 3. Efficacy verification of PPNE-selected perturbagens across scRNA-seq datasets from diverse breast cancer subtype cell lines.

A Correlation patterns among perturbagens within the HT class. The gradient from yellow to green indicates the range of correlation coefficients, with yellow denoting positive and green denoting negative correlations. B Expression patterns of the target and activated genes for five selected therapeutic candidates for breast cancer across various cell types. C Expression profiles of representative marker genes associated with breast cancer across six major breast cancer cell lines. D Cell viability shown by drugs (A-443654, buparlisib, celastrol, vemurafenib, and tozasertib) in breast cancer cells (MCF-7, HCC-1143, HCC-1954, T47D, and BT 474). Results are expressed as mean ± SD. E Flow cytometry (FACS) analysis of annexin expression in the HCC-1143 cell line treated with the pharmacological agent buparlisib. F Bar graph showing the proportion of apoptotic cells. Error bars indicate standard deviation from the mean.

Table 1.

DR candidates for breast cancer treatment and their repurposing strategies

Perturbagen name	Primary use	Proposed anticancer mechanisms	Clinical trials* (Studies)		MoA (ChEMBL)
			Cancer	Breast cancer
Tozasertib	Various cancers (unspecified)	• Blocks cell-cycle progression and induces apoptosis	2	—	Aurora A/B/C kinase inhibitor
A-443654(CID: 10172943)	AML	• Inhibition of breast cancer cell proliferation	—	—	Pan-Akt inhibitor(Akt1, 2, 3)
Buparlisib	Various cancers (unspecified)	• Prevent the PIP3 rebound and disconnection betweenPI3K and Akt activation	87	27	Pan-PI3K inhibitor
Vemurafenib	Erdheim–Chester Disease, melanoma	• Reduction of cytoplasmic phosphorylated Erk andcell proliferation driven by Ki-67 • Inhibition of the MAPK pathway	163	3	B–RAF inhibitor
Celastrol	Various cancer (unspecified)	• Inhibition of the NK-kB signaling pathway	1	—	Proteasome inhibitor

Based on the aforementioned experimental results, we performed 3-(4, 5–dimethylthiazol-2-yl)-2, 5-diphenyltetrazolium bromide (MTT) assays to assess cell viability relative to control conditions using five selected perturbagens (tozasertib, A-443654, buparlisib, vemurafenib, and celastrol) across five cell lines: HCC1954 (HER2-positive), HCC1143 (triple-negative breast cancer; TNBC), BT-474 (Luminal B), T47D, and MCF7 (Luminal A) (Fig. 3D). When cells were exposed to varying concentrations of each perturbagen (5–25 μM) for 16 h, followed by cell viability assessment using an MTT assay, a significant reduction in cell viability was observed across nearly all cell lines with increasing perturbagen concentrations for most of the compounds tested. At a concentration of 10 μM, tozasertib exhibited a significant reduction in cell viability, with rates of 44.90% and 48.87% observed in the MCF-7 and BT474 cell lines, respectively, indicating its potential as a novel therapeutic candidate for Luminal A/B-type breast cancer. The results for most cell types exposed to vemurafenib indicated its antitumor efficacy, but it was less effective than other perturbagens, with <50% cytotoxicity at concentrations below 10 µM. Celastrol is a promising anticancer agent for Luminal B-type breast cancer within the BT474 cell line, with survival rates ranging from 34.69% to 19.07% at treatment concentrations of 5 µM to 25 µM. Of the perturbagens tested, A-443654 exhibited a notable concentration-dependent reduction in cell viability in the MCF-7, T47D, HCC1143, and HCC1954 cell lines. In particular, the HCC1143 cell line displayed remarkable results, with viability registering at 33.44% at a 10-μM concentration and further decreasing to 11.62% at a 20-μM concentration, which was the best cell viability-decreasing perturbagen effect among all cell types. Interestingly, all perturbagens, including buparlisib, showed pronounced anticancer activity in HCC1143, the TNBC cell line. Collectively, these findings indicate that the evaluated perturbagen candidates, including buparlisib, have significant potential as innovative therapeutics for metastatic breast cancer, an intractable form of the disease that is currently devoid of a definitive cure.

Cell lines HCC1143 and MCF-7 were analyzed after treatment with the compound A-443654 and buparlisib at concentrations of 10 μM and 20 μM, respectively (Fig. 3E, F and Fig. S7). In the MCF-7 cell line, A-443654 treatment with concentrations of 10 μM and 20 μM led to early apoptosis rates of 15.26% and 21.96%, respectively, and late apoptosis rates of 23.19% and 32.08%, respectively. The late apoptosis rate showed a significant increase of 29.41% relative to that of the control group. In MCF-7 cells, a clear positive correlation was observed between the drug concentration and the magnitude of apoptosis, suggesting that an increase in drug concentration corresponds to enhanced apoptotic activity. In contrast, within the HCC1143 cell line, the cell death rate increased significantly to 23.36% at 10 μM and further increased to 51.66% at 20 μM, predominantly due to necrosis. Intriguingly, cells exposed to each compound appeared to circumvent the intermediate phases of early and late apoptosis, advancing directly toward necrosis. In MCF7 cells exposed to buparlisib at concentrations of 10 µM and 20 µM, the early apoptosis rates were 16.7% and 16.59%, respectively, whereas the late apoptosis rates were 18.23% and 25.26%, respectively. These results underscore the distinct association between increased drug concentration and augmentation of late apoptosis. Consistent with this trend, HCC-1143 cells treated with buparlisib showed increased apoptosis, which correlated with increasing drug concentrations. At 10 µM, early apoptosis was observed at 14.12%, with late apoptosis measuring at 17%. At an increased drug concentration of 20 µM, early apoptosis increased to 19.52%, whereas late apoptosis remained consistent at 18.52%. Significantly, amplification during early apoptosis was concurrent with an enhanced drug concentration.

Deciphering perturbagen activity in breast cancer phosphoproteome

A total of 1246 differentially expressed phosphosites (DEHs) that met a significance threshold of p < 0.05 were identified from a pool of 15,714 phosphosites that passed quality control. Among these DEHs, 233 were downregulated and 895 were upregulated (Fig. 4A, Data File S12, and Fig. S8A). To prioritize the key kinases associated with DEHs in breast cancer, we computed a kinase substrate score based on a composite of kinase recognition motifs. Prominent kinases such as MAPK, CDK, and Akt, which are known to play significant roles in breast cancer progression and survival, have previously been shown to exhibit regulated expression patterns influenced by DEHs (Fig. S8B)^36,37. The Gene lists with associated Drug Actions (GDA) of CID-10172943, buparlisib, vemurafenib, and tozasertib included 2, 64, 12, and 14 breast cancer-specific DEHs, respectively, with p-values < 0.05. However, no DEHs were identified in the remaining perturbagens (Data File S13). Among the top three perturbagens, buparlisib, vemurafenib, and tozasertib, which showed high drug efficacy in TNBC cell lines, the GDAs in buparlisib contained five DEHs (S1487-PI3KC2B, S338-PI3KC2A, T2474-MTOR, Y453-PI3KR2, and Y524-PI3KCD) within the protein domain of the PI3K/AKT/mTOR pathway, which is widely recognized as a critical regulator of breast cancer growth and survival (Fig. 4B)³⁷. Through the analysis of breast cancer-derived DEHs belonging to GDAs related to buparlisib, the top 30 major kinases affected by DEHs were investigated. Remarkably, among the identified kinases, CDK1, a widely recognized kinase associated with enhanced survival in breast cancer upon upregulation³⁷, and AKT1, a critical player implicated in the initiation, progression, and development of breast cancer³⁸ were identified (Fig. 4C). DEHs within the GDAs of buparlisib and AKT1 exhibited various elevated kinase substrate scores, reflecting the extent of the regulation of phosphorylation levels per site (Fig. 4D). Analysis of the interconnections between the selected kinases revealed a noteworthy correlation between PIM1, a highly promising therapeutic target for breast cancer, and AKT1 (Fig. 4E). Furthermore, various members of the MAPK family of genes, known for their pivotal involvement in the advancement and progression of breast cancer, exhibited interconnected relationships with one another³⁹ (Fig. 4E). The PI3K/Akt/mTOR pathway is involved in regulating cell proliferation, apoptosis, metabolism, and angiogenesis and is activated in breast cancer. The activation of this pathway increases tumor progression and decreases patient survival⁴⁰. In addition, mTOR activation is correlated with poor prognosis in breast cancer²¹. Here, buparlisib inactivates Akt by inhibiting the PI3 subunits p100α, β, and γ. Next, TSC1/2, which is inhibited by Akt, is activated, and Rheb is converted to a GDP-bound inactive state. Finally, inactivated RHEB was unable to bind to mTOR, resulting in the inhibition of mTOR, which contributed to the suppression of breast cancer (Fig. 4F). These findings led us to speculate that the modulation of buparlisib efficacy might be governed by the expression patterns of GDAs influenced by DEHs.

Fig. 4. Buparlisib suppresses breast cancer progression by targeting the PI3K pathway.

A Volcano plot depicting 1246 DEHs between normal and breast cancer tissues. Pink, dark green, and gray symbols denote the overexpressed, underexpressed, and non-significantly expressed entities, respectively. The threshold is set at a p-value below 0.05 and an absolute log₂ fold change >1. B Significant DEHs of breast cancer associated with GDAs of buparlisib. Each score represents the log₂ fold change between normal and breast cancer samples in the phosphoproteome. C Combined kinase-substrate score for the top 30 phosphosites, where a higher combined score indicates a more favorable alignment with a kinase motif and the kinase-substrate phosphorylation profile of a phosphosite. D Plots illustrating profile, motif, and combined scores, along with the prediction score, for PI3K phosphosites associated with AKT1 kinase. The motif and profile scores represent the scoring of the kinase-recognition motif and experimental perturbation, respectively, while the combined score amalgamates the motif-matching and profile scores. Prediction score denotes the likelihood of predicting a specific kinase for a given site. E Plots depicting an interaction network of kinases associated with DEHs of GDAs of buparlisib. F Illustration of the action mechanism of buparlisib using the SIMD platform.

Discussion

We developed an innovative SIMD strategy for drug repositioning based on the synergistic effect of bidirectional data from proteotranscriptomes and single cells. We successfully verified the performance of the ACPS through a comprehensive assessment conducted by comparing it with individual single-layer omics data in diverse aspects, including drug prescription records of TCGA patients and clinical trials, structural similarity to approved breast cancer medications, and alignment with biological functionality.

Numerous studies have hypothesized that compounds that exhibit inverse transcriptional correlations with disease states may offer therapeutic benefits by applying LINCS L1000 perturbation profiles using a range of computational approaches. For example, anti-correlation analysis identified CGP60474 as a candidate for sepsis, demonstrating cytokine suppression and survival benefit in endotoxemia models⁴¹, while characteristic direction–based similarity metrics revealed kenpaullone as an inhibitor of Ebola-associated expression profiles, validated in vitro⁴². Dutasteride was prioritized for LPS-induced neuroinflammation using QSAR-guided BBB filtering in combination with transcriptomic anti-correlation⁴³, and BRD-K60230970 was identified via deep embedding of perturbation signatures as a MAPK pathway modulator⁴⁴. Furthermore, large-scale dose–response modeling across over 140 million profiles enabled the quantitative assessment of drug–gene associations, exemplified by tamoxifen’s transcriptional efficacy in MCF7 cells⁴⁵.

A separate study developed a LINCS L1000-based repositioning pipeline that identified six candidate compounds, including suramin, by selecting those that significantly downregulated genes associated with elevated HS and CS expression, thereby suggesting their potential to enhance anticancer peptide efficacy⁴.

The primary distinctions of SMID relative to prior LINCS L1000-based drug repositioning studies reside in two key aspects. The primary distinction between SMID and previous LINCS L1000-based drug repositioning studies lies in its multi-layered validation approach, which comprehensively captures gene expression dynamics across the central dogma, integrating transcriptomic and proteomic changes. While most previous studies relied solely on disease transcriptomic profiling for drug repositioning, SMID integrates bulk RNA-seq and proteomic data, followed by additional validation using neoplastic epithelial cells derived from breast cancer single-cell datasets, thereby enabling precise in silico validation of drug efficacy. Second, unlike many studies that validated the predicted drug efficacy solely via in vitro assays or literature evidence, our approach enhances predictive accuracy by additionally incorporating in vivo validation and phosphosite expression data related to drug targets, thus providing a more comprehensive evaluation of the drug response.

Our ACPS framework is distinguished from previous approaches by its use of diverse data types and evaluation of scoring strategies from multiple perspectives. In pursuit of a more intricate drug selection methodology, we aimed to select novel pharmaceutical compounds with high ACPS, indicative of an anti-correlation with the gene expression profile of a subset of 5370 NE cells expressing markers, such as SLP1, which is recognized for its association with aggressive and metastatic breast cancer. The experimental validation of these five pharmaceutical agents demonstrated their potent inhibitory efficacy in a breast cancer cell model. Notably, the mechanism of action of the five selected pharmaceutical agents was related to the inhibition of kinases that play pivotal roles in phosphorylation. Buparlisib exhibited the greatest potential to inhibit breast cancer progression. Buparlisib, acting as a comprehensive pan-PI3 kinase inhibitor, is well documented for its ability to specifically target PI3 kinase signaling pathways, ultimately leading to the inhibition of tumor growth through the meticulous regulation of mTOR and ERK pathways⁴⁶. Originally developed to address a diverse range of malignancies, buparlisib is currently the subject of investigation in 75 ongoing clinical trials for breast cancer, encompassing both monotherapy and combination therapies. This extensive clinical scrutiny bolsters our confidence in its predictive value. Furthermore, the therapeutic potential of buparlisib has been assessed in the context of TNBC, a subtype that accounts for 15–20% of breast cancer cases⁴⁷ and is characterized by substantial heterogeneity and limited treatment options⁴⁸. Vemurafenib, known for its ability to inhibit tumor growth by targeting the B-RAF-MEK-ERK signaling axis, demonstrated a significant dose-dependent decrease in cell survival. Among the strongest responses observed in HCC-1143, a TNBC cell line tested, vemurafenib elicited the most pronounced effect. Furthermore, a clinical case report highlighted the successful local treatment of TNBC with vemurafenib, which was accompanied by minimal side effects. Vemurafenib is associated with the B-Raf V600E mutation in TNBC⁴⁹. Further predicted pharmaceutical candidates include the aurora kinase inhibitor (tozasertib), pan-Akt inhibitor (A-443654), and IKK-α, β inhibitor (celastrol), all of which have demonstrated therapeutic efficacy in treating TNBC^49–51. In the experimental validation of breast cancer using a cell viability assay and FACS analysis for SIMD-predicted perturbagen evaluation, A-443654 circumvented the intermediate phases of early and late apoptosis, directly inducing necrosis, whereas buparlisib predominantly induced apoptosis.

To enhance the predictive capabilities of SIMD and overcome its limitations, we address several pivotal considerations that require meticulous consideration. First, the L1000 dataset utilized in our study relied on a restricted set of landmark genes. Hence, the performance enhancement of SIMD can be anticipated by utilizing drug response data encompassing a wider gene spectrum or refining improved gene expression imputation methodologies. Second, multi-omics profiling and public proteomic datasets, such as CPTAC, are increasing across diverse disease contexts. SIMD is poised to establish itself as an exceptionally precise and efficient tool for compound screening, particularly when capitalizing on the substantial magnitude of these proteomic datasets. As a representative example, the combination experiments across data types in Fig. 1b demonstrated that the proteotranscriptomic-based ACPS consistently yielded the highest AUC values compared with single-omics approaches across all thresholds, exceeding 0.7 at p-values below 0.04. However, owing to the limitations in data quantity, the performance appeared to be sensitive to threshold variation. In future studies, optimizing the methodology by incorporating large-scale proteomic resources, such as CPTAC data from other cancer types, will be critical for enhancing the robustness of the approach. Third, SIMD must encompass diverse characteristic datasets such as metabolomics⁵², which elucidate the metabolic responses of drugs, and epigenetics⁵³, which are critical regulators of gene expression. In subsequent research endeavors, the integration of diverse multimodal omics data into SMID has the potential to predict not only the efficacy of individual drugs with greater accuracy, but also precise dosage and metabolic pathways. Lastly, in this study, we isolated the NNE population from breast cancer tissues and, through multiple lines of evidence—including the observation of partial expression of normal epithelium–associated genes—predicted drug responses targeting bona fide cancer cells. Despite these efforts, it remains challenging to regard NNE cells adjacent to NE cells as completely normal epithelial cells, representing an intrinsic limitation of this study.Therefore, future investigations will aim to overcome this limitation by incorporating experimental comparisons with normal breast epithelium to better delineate the molecular boundary between normal and malignant epithelial states. SIMD is a synergistic DR platform that harmoniously reflects the strengths of single-cell data with distinct characteristics of heterogeneous breast cancer cells and proteotranscriptome data that encapsulate the molecular dynamics of individual genes. It has been validated across multiple dimensions, including clinical, pharmacological, and biological perspectives. Our platform incorporates stage-specific molecular dynamics in patients with breast cancer and has the potential to offer greater clinical relevance than previous screening methods that solely assess drug activity using single-omics data from specific cell lines. Furthermore, it is anticipated that this platform will offer drug response predictions or personalized therapeutic strategies for individual patients because it encompasses the molecular signature of each patient. Finally, SIMD, which prognosticates novel drug efficacies rooted in distinct molecular signatures inherent to particular diseases, provides new therapeutic avenues for incurable diseases that currently have no effective therapies, such as Alzheimer’s disease and rare disorders.

Methods

Datasets

We compiled the proteotranscriptomic and clinical data of patients with breast cancer according to the following conditions: (1) The normalized RNA expression profiles consisted of 113 and 1098 normal and breast cancer specimens from TCGA (https://portal.gdc.cancer. gov, May 2018). (2) The LC-MS/MS datasets were composed of 105 subset specimens from patients identical to the RNA-seq data from the CPATC data portal (https://proteomics.cancer.gov/data-portal, version 20200511). (3) The clinical data of patients with breast cancer included 72 types of prescribed drugs. Normalized L1000 datasets were downloaded from GEO (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE92742; LINCS level 3), and 15,207 transcriptome expression profiles were perturbed by 313 perturbagens tested on breast cancer cell lines (BT20, HME1, HS578T, MCF10A, MCF7, MDAMB231, and SKBR3). Owing to nomenclature inconsistencies and errors in perturbagen names, we standardized the compound names using PubChem CID with semi-automatic methods, as detailed in our previous paper²⁶. Among the perturbagens in L1000, 301 were mapped to the CID in PubChem. To validate the therapeutic effect of the deduced potential therapeutic candidates as aspects of 2D structure chemical similarity and action mechanism of perturbagens between approved breast cancer drugs and to deduce potential therapeutic candidates, we downloaded 298 Structure-Data Files (SDF) and 253 bioassay data from PubChem using the REST protocol (https://pubchemdocs.ncbi.nlm.nih.gov/pug-rest). Breast cancer susceptibility genes were obtained from the COSMIC Cancer Gene Census. We downloaded the processed single-cell RNA-seq (scRNA-seq) data of patients with breast cancer from ArrayExpress accession number E-MTAB-8107 and selected nine samples (sc5rJUQ026, sc5rJUQ033, sc5rJUQ039, sc5rJUQ042, sc5rJUQ050, sc5rJUQ053, sc5rJUQ060, sc5rJUQ064, and scrJUQ068)⁶⁵.

Between the mean CNV profile and GIS of each cell (Fig. S4C), we partitioned the medoid clusters using the PAM function of the embedded PAM algorithm in the R cluster package (ver. 2.1.4), based on the inferred optimal K (2–4) using the silhouette scores. Based on the parameters of the code (https://github.com/Swarbricklab-code/BrCa_cell_atlas/blob/main/inferCNV) provided by Neftel et al., we applied two-cluster s.d. to the left as the threshold to distinguish NE and NNE cells and 1.5 s.d. under the first cancer cluster. For cancers in which partitioning around the medoids could not be set over one cluster, the thresholds were defined as 1 s.d. to the left and 1.25 s.d. under the cluster means. We obtained 5370 NE and 3701 NNE cells from 9,106 epithelial cells, of which 35 remained unassigned (Fig. S4D).

Anti-correlated proteotranscriptome perturbation score (ACPS)

Each of the 15,207 perturbed gene expression profiles of PSP with 22,268 genes was calculated using the log₂ fold change between each vehicle control and PSP experimental conditions within the same cell line in the L1000 dataset. For perturbagen–cell line pairs with multiple doses or time points, gene-wise expression values were averaged prior to calculating log₂ fold changes between the PSP-treated and matched vehicle control conditions. To standardize the expression signature gene identifiers, all gene designations were mapped to Ensembl gene IDs using the BioMart R package (version 3.7).

The ACPS is a merged score that considers gene expression profile trends that change through central dogma processes using proteotranscriptomics to measure the repurposing ability of each perturbagen. First, we computed the log₂-based fold change based on each transcriptome or proteome expression abundance derived from identical gene i for comparison between the two groups, including normal versus breast cancer tissue and non-treatment versus treatment with perturbagen k. To reflect the overall proteotranscriptome perturbation change trend in the breast cancer condition of each gene in a summarized score, we obtained the corresponding proteotranscriptome expression value, which is the mean value between the log₂ fold value of the transcript and protein in relation to gene i. To quantify the degree of potency of each PSP using a contrastive gene perturbation pattern for the breast cancer gene, we computed the Pearson correlation coefficient ($r_{{\mathit{psp}}_k}$) according to Eq. 1, where GEP denotes the gene expression profile perturbed by PSP_k.

$$\begin{array}{c}{r}_{{\mathit{PSP}}_k}=-\left|\frac{{\sum \left({\mathit{GEP}}_{\mathit{ik}}-\overline{{\mathit{GEP}}_k}\right)}^2}{\sqrt{\sum {\left({\mathit{GEP}}_{\mathit{ik}}-\overline{{\mathit{GEP}}_k}\right)}^2\sum {\left({\mathit{PTE}}_k-\overline{{\mathit{PTE}}_k}\right)}^2}}\right|\hfill \\ {\mathit{GEP}}_k=\left\{{\mathit{GEP}}_k^i\in |{\mathit{GEP}}_{\mathit{ik}}\dots {\mathit{GEP}}_{\mathit{nk}}\right\}\hfill \end{array}$$ 1

Fisher’s z-transformation of sample correlation coefficients is known to substantially improve normality. Therefore, we transformed the correlation coefficient, r, for each PSP_k (Eq. 2).

$$Z_{{\mathit{PSP}}_k}=\frac{\mathrm{In}\left(\left(1+r_{{\mathit{PSP}}_k}\right)÷\left(1-r_{{\mathit{PSP}}_k}\right)\right)}{2}$$ 2

The left-tailed p-values were calculated based on the z-values of each PSP to establish a statistically significant cutoff for selecting anti-breast cancer drugs that have a confident anti-correlation with cancer. Finally, we deduced the ACPS of PSP_k by normalizing it to the range [0,1] using the Z_k values (Eq. 3).

$${\mathit{ACCP}}_{{\mathit{PDT}}_k}=\frac{Z_{{\mathit{PDT}}_k}-\min \left(Z_{{\mathit{PDT}}_k}\right)}{\max \left(Z_{{\mathit{PDT}}_k}\right)}-\min \left(Z_{{\mathit{PDT}}_k}\right)$$ 3

In cases where some perturbagens had several PSPs, we calculated the mean ACPS value with a passed p-value cutoff as the representative value. To evaluate the performance of ACPS by comparing multiple conditions, we categorized the data types into six groups (high-protein (HP), high-RNA (HR), high-proteotranscriptome (HT), low-protein (LP), low-RNA (LR), and low-proteotranscriptome (LT)) according to two criteria: 1) division into three data types and 2) grouping perturbagens that passed the baseline p-value threshold (0.04) into high and low categories. As the gold standard for approximating the area under the curve (AUC) for validation, 72 prescribed anti-breast cancer drugs from TCGA with standardized names and compound identifiers (CIDs) from our previous research¹² were extracted.

Meta-analysis of target genes of perturbagens

To ascertain the current clinical trial status of the selected perturbagens, a comprehensive retrieval was conducted using queries with specific terms, including, “Cancer,” “Tumor,” “Tumour,” and “Neoplasm,” as well as refined variants like “breast cancer,” “Breast Carcinoma,” “Breast Neoplasm,” and “Breast Tumor” on clinicaltrials.gov. Based on the aforementioned investigation, we categorized perturbagens into distinct risk classes, namely HT-low, HT-moderate, and HT-high, according to the following criteria (Data File S6).

• Low-risk class: This class encompasses chemicals that not only demonstrate ongoing clinical trials but also exhibit continued validation of efficacy across a diverse range of cancer types, including breast cancer.

• Moderate-risk class: Within this category, chemicals undergo ongoing validation of their efficacy in various cancer types, excluding breast cancer.

• High-risk class: This classification pertains to challenging chemicals that are at the forefront of novelty and lack prior studies in the context of cancer research.

Chemical structure similarity

Among the 367 perturbagens, consisting of 67 anti-breast cancer drugs and 302 perturbagens from L1000, we collected 337 SDF formats of the PubChem database (https://pubchem.ncbi.nlm.nih.gov/) using the PubChemPy python library. The smile structures of the chemicals from SDF were processed using the extractDrugExtended function in the RCDK R package (ver. 3.6). To compare the chemical structures, we calculated the fingerprint similarity between all anti-breast cancer and small-molecule pairs using the calcDrugFPSim function and Tanimoto metric. We calculated the proportion of perturbagens with a Tanimoto similarity higher than the cut-off for each class.

PPI network construction and enrichment analysis

To benchmark the relationship between the COSMIC breast cancer susceptibility genes and drug target genes from the PubChem bioassay data, we constructed PPI networks as follows: 1) All collected gene symbols were mapped to the STRING protein ID via ensemble protein IDs using BioMart (ver 3.16) and STRINGdb (ver 3.16; 9606.protein.aliases.v11.5.txt.gz file). 2) PPIs were obtained from the STRING database with a confidence score (9606.protein.links.v11.5.txt.gz). 3) We counted the number of PPIs with breast cancer-related gene sets (BCRG) per perturbagen. 4) We computed p-values from t-tests to examine the significance of the mean difference in PPIs between the omic-class pairs (HP-DP, DR-HR, and DT-HT). To define BCRG, only genes with PTEi ≥ 1 and p-value < 0.05 were considered for enrichment analysis. We performed KEGG and GO enrichment analyses using ClusterProfiler (ver 4.2) with sets for Homo sapiens (HS) from msigdbr (ver 3.16), based on the following two input data types: Initially, the difference in log2FC between each PTEi and GEPi was used to infer the functional and biological characteristics of the perturbagen. Second, only PTEs were used to obtain the BCRG, which reflects the environment of breast cancer progression. For KEGG analysis, we used predefined breast cancer-related pathways of KEGG as a BCRG, including KEGG_HOMOLOGOUS_RECOMBINATION, KEGG_MAPK_SIGNALING_PATHWAY, KEGG_HOMOLOGOUS_RECOMBINATION, KEGG_CELL_CYCLE, KEGG_P53_SIGNALING_PATHWAY, KEGG_WNT_SIGNALING_PATHWAY, and KEGG_NOTCH_SIGNALING_PATHWAY. The BCRG proportions for each perturbagen were calculated using significant gene sets (NES ≥ 1 and <0.05) and predefined BCRG and were summarized according to the average rate in each class.

Search strategy and selection criteria for clinical trials

We conducted a systematic review of clinical trials to assess the potential of perturbagens predicted using the SIMD platform as novel interventions for breast cancer. Our search utilized ClinicalTrials.gov (https://clinicaltrials.gov/), which encompasses clinical trial studies from over 200 countries, including the United States. We discerned pertinent records through a targeted search employing the following query within an exhaustive dataset encompassing all study statuses and disease stages: “(Drug name) OR (Drug MoA) AND (Cancer) OR (Tumor) OR (Neoplasm) OR (breast cancer) OR (Cancer, breast).” To ensure a precise count of clinical trials, we diligently excluded duplicates of search terms, both in the overarching search terms (Cancer; Tumor; Neoplasm) and subordinate search terms (Breast cancer; Cancer, breast), as well as within the subordinate search terms themselves ((Breast cancer) and (Cancer, breast)).

Data processing of scRNA-seq

Beginning with merging nine scRNA-seq data of patients with breast cancer (E-MTAB-8107) using the “Merge” function, all statistical analyses were performed using the Seurat (ver 4.2) package in R. QC was performed using the RunMiQC function in MiQC with default parameters (posterior.cutoff = 0.75, model.slot = “flexmix_model,” and percentage of mitochondrial) to remove low-quality genes. The data were normalized using the LogNormalize method in the “NormalizeData” function, and highly variable genes were identified using the “FindVariableFeatures” function with the vst method. The RunHarmony function of the SeuratWrapper R package was applied to perform batch effect correction after determining principal components 1–20. FindNeighbors and FindClusters were used to build KNN graphs based on Principal Components (PCs) and to identify clusters with a resolution of 0.6. We applied the RunUMAP function to visualize cell manifolds in 2D space. Cell clusters were annotated using Garnett in monocle version 3, with the marker genes of breast cancer by Wu et al., downloaded from https://github.com/Swarbricklab-code/BrCa_cell_atlas²¹.

Single-cell RNA-seq datasets were obtained from GSE161529 (normal epithelium; NOR_EPI) and GSE176078 (breast cancer epithelium; BC_EPI), comprising 3 normal samples (GSM4909255, GSM4909256, GSM4909258) and 25 breast cancer samples (“CID3838”, “CID3921”, “CID3941”, “CID3946”, “CID3948”, “CID3963”, “CID4040”, “CID4066”, “CID4067”, “CID4290A”, “CID4398”, “CID44041”, “CID4461”, “CID4463”, “CID4465”, “CID4471”, “CID4495”, “CID44971”, “CID44991”, “CID4513”, “CID4515”, “CID45171”, “CID4523”, “CID4530N”, “CID4535”), respectively. Cells were filtered using the MiQC algorithm with the parameters percent.mt = “percent.mt”, nFeature_RNA = “nFeature_RNA”, posterior.cutoff = 0.75, and model.slot = “flexmix_model”. Batch effects were subsequently corrected using harmony integration. Following quality control and integration, the dataset comprised 16,806 normal epithelial cells and 45,881 cancerous epithelial cells. Among the cancer cells, 7443 were classified as estrogen receptor-positive (ER), 1578 as HER2-enriched, and 7785 as triple-negative breast cancer (TNBC). Differentially expressed genes (DEGs) were identified using the FindMarkers function with parameters logfc.threshold = 0, min.pct = 0.1, and test.use = “wilcox”. Based on the criteria of adjusted p-value (p_val_adj) <0.05 and absolute log2 fold change ( | avg_log2FC | ) ≥ 1, a total of 3106 DEGs were identified (NOR_DEG.csv). To evaluate the transcriptomic similarity between epithelial cell groups, Jaccard similarity was calculated using gene expression profiles. For each group, the gene expression matrix was set to include only commonly expressed genes across groups. Let G_i denote the binary expression vector (gene presence/absence) for group i, the Jaccard similarity between group i and j is defined as (Eq. 4).

$$J\left(G_i,G_j\right)=\frac{\left|G_i\cap G_j\right|}{\left|G_i\cup G_j\right|}$$ 4

The Jaccard distance was computed using the vegdist function from the vegan R package, and the similarity was derived by subtracting the distance from 1.

Characterization of NE cells

The FindAllMarkers Seurat function was used to identify markers of NE cells, using the likelihood-ratio test by comparing different cell types (thresh.use = 0.25, min.cells.feature = 3). The selected top seven DEGs among the genes with adjusted p-value < 0.05 and log2FC ≥ 1 were visualized using the dittoDotPlot function in the dittoSeq R package (ver 1.10). To identify the enriched gene sets in NE cells, enrichment analysis was performed using the DEGs of NE cells and the HS-C2 set in the Molecular Signatures Database (MSigDB) by applying the fast GSEA (fgsea) of the fgsea R package (ver1.24). We selected gene sets with p-value < 0.05 and Normalized Enrichment Score (NES) ≥ 2.5 and visualized them using the ggplot2 package (ver 3.4). We also performed GSEA at the single-cell level using scRNA-seq counts of NE/NNE cells and the HS-Kyoto Encyclopedia of Genes and Genomes (KEGG)/HS-BP database of msigdb by applying the enrichment function of the escape R package (ver. 0,1).

Extracting NE cells of breast cancer

To perform analysis to classify breast cancer NNE cells, we adopted the method established by Neftel et al.⁵⁴. The gene expression matrix and metadata of epithelial (9071), T (9403), and B cells (830) were extracted from the cell population to generate the input data for the InferCNV objects. The CNVs for each cell were inferred using the inferCNV R package (ver1.4) with the following inferCNVrun() parameters: window length = 101 and maximum centered threshold = 3. After filtering out low-quality cells with a mean count of <0.1, the score was denoised based on the sd_amplifier 1.3 threshold. We used T and B cell types as references to infer CNV expression intensity. After scaling the inferred CNV profiles between −1 and 1, we defined the mean of the squares of the score as the genomic instability score (GIS) of each cell. For each sample, we deduced the mean CNV profile based on the highest GIS of the selected top 5% of cells. For each sample, we deduced the mean CNV profile based on the highest GIS of the selected top 5% of cells. We then calculated and plotted the correlation coefficients between the mean CNV profile and GIS of each cell (Fig. S4C). We partitioned the medoid clusters using the PAM function of the embedded PAM algorithm in the R cluster package (ver. 2.1.4), based on the inferred optimal K (2–4) using the silhouette scores. Based on the parameters of the code (https://github.com/Swarbricklab-code/BrCa_cell_atlas/blob/main/inferCNV) provided by Neftel et al., we applied a two-cluster s.d. threshold to the left to distinguish NE and NNE cells. Specifically, we used a cutoff of 1.5 s.d. below the mean of the first cancer cluster. In the case of cancers in which partitioning around the medoids could not be set over one cluster, thresholds were defined at 1 s.d. to the left and 1.25 s.d. under the cluster means. We obtained 5370 NE and 3701 NNE cells from 9106 epithelial cells, of which remained unassigned⁴⁶ (Fig. S4D).

Cell culture and MTT assay

A-443654 (CAS: 552325-16-3), Buparlisib (CAS: 944396-07-0), Celastrol (CAS: 34157-83-0), Tozasertib (CAS: 639089-54-6), and Vemurafenib (CAS: 918504-65-1) were purchased from MedChemExpress. These compounds were solubilized in dimethyl sulfoxide (DMSO) to create stock solutions of 10 mM and preserved at −4 °C for future use in cell-based experiments. MCF-7, HCC-1143, HCC-1954, T47D, and BT-474 cell lines were acquired from the Korea Cell Line Bank (KCLB, Seoul, Korea). The cells were cultured at 37 °C with 5% CO₂ using 10% fetal Dulbecco’s Modified Eagle Medium (DMEM) (Welgene) supplemented with bovine serum (HyClone, Logan, UT, USA), along with penicillin (100 U/mL) and streptomycin (100 mg/mL) (Welgene, Gyeongsan, Korea). For the experiments, cells were seeded at a density of 5 × 104 cells/well in a 96-well plate. Following cell seeding, the drugs were added to each well and the cells were incubated for 16 h. Subsequently, 3-(4,5-dimethylthiazol-2-yl)-2,5 diphenyltetrazolium bromide (MTT) solution (5 mg/mL) was added to each well, and the plates were incubated for 4 h. The culture medium was aspirated, and the formazan crystals formed within the cells were dissolved in DMSO. The optical absorbance was measured at 540 nm using an ELISA reader (Biochrome Ltd., Holliston, MA, USA). The cell viability of perturbagen k (CBpk) was determined per cell type using the mean value obtained from the j-th concentration of the i-th perturbagen and dividing it by the average control measurement value before multiplication (Equation 5).

FACS assay

For the experiments, the cells were seeded at a density of 1.5 × 10⁵ cells in a cell culture dish (100/20 MM). MCF-7 and HCC 1,143 cell lines were treated with A-443654 and Buparlisib at concentrations of 10 μM and 20 μM for 16 h, respectively. Apoptosis was assessed using Muse Annexin V and Dead Cell Kit Reagent (Luminex, USA). The cells were then stained with the reagent for 20 min at room temperature in the dark. The stained cells were analyzed using FACS Guava software version (Goma Biotech, Seoul, Republic of Korea).

Statistical analyses were conducted using SPSS version 18.0 (IBM Corp., Armonk, NY, USA). Duncan’s post hoc tests were used to discern variations in mean values among the experimental cohorts. The results are presented as the mean ± standard deviation of triplicate experiments. Moreover, each experimental dataset is represented as the mean ± standard deviation. Statistical significance was set at p < 0.05.

Proteotranscriptomic and phosphoprotein data analysis

TCGA RNA-seq data with 10,967 genes for each sample were analyzed using generalized linear models in the EdgeR (version 3.7) package to deduce the representative gene expression signature by comparing breast cancer and unpaired normal samples (adjusted p < 0.05). The proteomics data contained 74 datasets (37 global proteomes and phosphoproteomes) from 108 patients with breast cancer and healthy samples (105 and 3, respectively). Each dataset comprised three breast cancer samples and one pooled internal reference sample. All raw files of the 74 datasets were downloaded from the CPTAC data portal (global proteomes: PDC000173; phosphoproteomes: PDC000174) and studied as the “Proteotranscriptomics” or “Phospho-proteotranscriptomics” dataset in the following analysis. All raw files were reprocessed using MaxQuant software (Ver. 1.6.17.0) using the UniProt protein sequence database of humans (Proteome ID: UP000005640; reviewed June 2020). Searches were performed using cysteine carbamidomethylation as a fixed modification, and protein N-terminal acetylation and oxidation of methionine as variable modifications. Phospho (STY) was incorporated as an additional variable for the phosphoproteomic analysis. Trypsin was chosen as the protease, permitting up to two missed cleavages, and the peptide mass was constrained to a maximum of 4600 Da with a minimum length of seven amino acids. The precursor mass and fragment tolerance were set at 20 ppm. Both the protein and peptide false discovery rates were set at 1%, as ascertained using a reverse database search. Unless specified explicitly, the search parameters in MaxQuant remained at their default values. Finally, 9950 proteins were identified in each sample.

Active Kinases inference and signaling pathways through phosphoproteomics data analysis

Gene lists of associated drug actions (GDA) were obtained from PubChem using the REST protocol. Tozasertib, CID: 10172943, Buparlisib, and Vemurafenib contained 85, 21, 772, and 112 GDAs, respectively. After applying the selectGrps function in PhosR (version 3.17) to filter out identified 58,148 phosphosites with >70% quantified values, the remaining 15,714 phosphosites were obtained for subsequent analysis within the software package⁴⁷. Missing values were imputed for every phosphosite, with >50% quantified values among replicates within each condition, using the scimpute and timpute functions. In subsequent analyses, we employed median centering and scaling for normalization, utilizing the imputed values of the phosphosites. Using the limma package, we applied a fit linear model (lmfit) to obtain a log2 fold change of 1246 differentially expressed phosphosites (DEH), which was statistically significant with a p-value threshold of <0.05. To identify prominent kinases and groups of DEHs, the expression values of DEHs were subjected to z-transformation using a standardized function within GDAs, and the kinase-substrate score was computed using the kinaseSubstrateScore function, with the specific condition parameters set at numMotif = 5 and numSub = 1. The results were effectively visualized as a signalome network using the plotKinaseNetwork function with a score threshold of 0.5.

Author contributions

SY. L and JH. A conceived the project and led the study. SY. L conducted the algorithm development and computational studies. MY. S performed proteomic and phosphoproteomic data analyses, along with a meta-analysis of the biological mechanisms of the drugs. YR. L designed and conducted molecular experiments. All authors drafted and reviewed the manuscript.

Data availability

All data used in this study were acquired from publicly available databases. RNA-seq data were acquired using the ICGC portal (https://portal.gdc.cancer.gov). Gene expression profiles induced by perturbagens in the L1000 dataset were obtained from the GEO database (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE92742; LINCS Level 3). Phosphoproteomic and proteomic data were sourced from CPTAC (https://proteomic.datacommons.cancer.gov/pdc/). scRNA-seq data from breast cancer patients were retrieved from ArrayExpress (https://www.ebi.ac.uk/biostudies/arrayexpress/studies/E-MTAB-8107). Publicly available normal and breast cancer epithelial datasets were obtained from GEO (GSE161529 and GSE176078, respectively). All datasets generated and utilized in this study are comprehensively documented in the manuscript and Supplementary Files. All codes used for data processing and analysis are available at https://github.com/SMID-SYL/SMID.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41523-025-00869-x.