A plasma proteomic signature of cancer-related sarcopenia implicates the IGFBP axis in muscle dysfunction

Abstract

Background

Cancer-related sarcopenia is associated with poor clinical outcomes but remains difficult to define and quantify in routine oncology practice. Current assessments rely on imaging and functional scales that are time-consuming and provide limited biological insight. We aimed to identify a plasma proteomic signature of cancer-related sarcopenia and to uncover circulating mediators involved in its pathophysiology.

Methods

Patients were included from two cohorts of the MATCH-R study (NCT02517892): a discovery cohort of advanced cancer patients treated with immunotherapy and an independent validation cohort of metastatic castration-resistant prostate cancer (mCRPC) patients treated with androgen-receptor pathway inhibitors. External validation was performed in the TRACERx cohort of non-small cell lung cancer. Skeletal muscle index at third lumbar vertebra (L3) was quantified using imaging, and ECOG performance status served as a functional proxy. Plasma proteomics was performed using the Olink Explore platform. An extreme gradient boosting (XGBoost) model was trained on a high-contrast subset using a neuromuscular-focused protein panel and validated across cohorts. Functional effects of candidate mediators were assessed in differentiating human myoblasts.

Results

The model generated a continuous sarcopenia probability (SP) score that correlated with muscle mass and functional status and consistently stratified overall survival across cohorts. A reduced four-protein model retained comparable performance, supporting translational applicability. Proteins associated with SP included insulin-like growth factor binding protein 1 and 2 (IGFBP1, IGFBP2), and interleukin-6 (IL6). IGFBP1 and IGFBP2 impaired myoblast differentiation, while IL6 induced IGFBP1 expression in liver cells.

Conclusions

Plasma proteomics enables scalable and biologically informed assessment of cancer-related sarcopenia, identifies tumor–host mediators of muscle dysfunction, and supports objective patient stratification for therapeutic intervention.

Graphical Abstract

We analyzed plasma protein profiles and CT-derived skeletal muscle mass in patients with cancer. Because sarcopenia lacks a single gold-standard definition in oncology, we defined extreme phenotypes based on concordant muscle mass and physical function, using patients with preserved muscle mass and function (ECOG PS 0) and those with low muscle mass and poor function (ECOG PS ≥2). These high-contrast groups were used to identify a muscle-focused plasma protein signature and train a machine-learning model that estimates a continuous probability of sarcopenia. Across multiple independent cancer cohorts, this probability score correlated with muscle mass, functional status, longitudinal functional change, and overall survival, and revealed biologically plausible mediators linking systemic cancer biology to muscle dysfunction. Created in BioRender. (2026) https://BioRender.com/byz3xgi

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12943-026-02716-4.

Background

Sarcopenia is characterized by a progressive decline in skeletal muscle mass, strength, and function, originally described in aging but now recognized across a range of clinical conditions [1]. In cancer patients, sarcopenia represents a major clinical challenge, as it is associated with impaired quality of life, increased treatment toxicity, reduced tolerance to systemic therapies, and poorer survival, thereby amplifying the overall burden of cancer for both patients and caregivers [2]. Sarcopenia frequently coexists with cancer cachexia and other wasting syndromes, further complicating clinical management [3]. However, reported prevalence estimates range widely [4], from 5% to nearly 90%, largely reflecting heterogeneity in assessment methods, cutoff definitions, and patient populations. This lack of consensus has limited the comparability and generalizability of existing studies, ultimately slowing progress in the field.

In geriatric medicine, the definition and assessment of sarcopenia have progressively evolved toward a multidimensional framework incorporating muscle mass, muscle strength, and physical performance [5]. In contrast, oncology research has largely relied on imaging-based quantification of muscle mass, typically derived from routine computed tomography scans. Although this approach has provided important prognostic insights, it captures only a fraction of the biological complexity underlying the functional decline and fails to account for systemic and molecular drivers of muscle dysfunction.

Cancer-derived sarcopenia is increasingly recognized as a multifactorial condition shaped by nutritional status, systemic inflammation, host metabolism, aging, sex, cancer biology, and anticancer treatments. Yet, the contribution of tumor-driven systemic signals and circulating mediators remains incompletely understood.

Emerging evidence, including from the TRAcking Cancer Evolution through therapy (Rx) (TRACERx) study [6], suggests that tumor biology and plasma proteomic profiles are linked to cachexia development, supporting the concept that circulating factors may connect tumor processes to host functional decline. Despite this, scalable approaches to characterize sarcopenia at the systemic molecular level are lacking.

Sarcopenia reflects a disruption of muscle homeostasis, where the balance between regeneration and degradation is shifted toward net muscle loss. Skeletal muscle health depends on coordinated regulation by metabolic, inflammatory, and neural inputs [7, 8]. Skeletal muscle maintenance and regeneration requires effective myoblast differentiation and repair processes [9], which are tightly coupled to endocrine signaling and neuromuscular integrity [10, 11]. Disruption of these integrated systems, rather than isolated muscle atrophy alone, may therefore underlie the development of sarcopenia in cancer.

Here, we leveraged large-scale plasma proteomics to define a molecular signature of muscle mass and physical function across multiple cancer types. We developed a sarcopenia probability (SP) score that enables continuous, biologically informed patient stratification and validated its performance across independent cohorts. In addition, we identified circulating mediators, including IGFBP1, IGFBP2, and IL6, and provide functional evidence linking these factors to impaired myogenesis, thereby offering mechanistic insight into cancer-associated sarcopenia.

Materials and methods

Patients

Patients were selected from two independent cohorts of the MATCH-R (Molecular Analysis for Therapy Choice - Resistance) study [12] (NCT02517892), a prospective, single-center clinical trial conducted at Gustave Roussy between 2015 and 2022. The study was designed to investigate tumor molecular evolution under anticancer therapies and to identify mechanisms of treatment resistance. Patients underwent tumor profiling at baseline and at disease progression, with the aim of characterizing resistance-associated molecular alterations and informing potential treatment strategies. Eligible patients were adults receiving or scheduled to receive systemic anticancer therapy, with at least one tumor lesion accessible for core needle biopsy, who were covered by a social security system and provided written informed consent.

The discovery cohort included patients with advanced solid tumors, predominantly lung and bladder cancers, treated with immune checkpoint inhibitors (MATCH-R immunotherapy, MATCHR-I). The validation cohort consisted of patients with metastatic castration resistant prostate cancer (mCRPC) treated with androgen-receptor pathway inhibitors [13]. The discovery cohort was used for model training and internal testing, whereas the mCRPC cohort served as an independent validation set (Supplementary Fig. 1).

MATCH-R study was approved by the CPP (“Comité de protection des personnes”) and by the ANSM (“Agence nationale de sécurité du médicament et des produits de santé”) and adhered to the principles in the Guideline for Good Clinical Practice and the Declaration of Helsinki. Written informed consent was obtained from all patients included in this study.

External validation was performed using the independent TRACERx cohort, whose data have been previously published. Clinical data, proteomic data, and outcome information were retrieved from the publicly available repository described in the original publication [6]. A schematic overview of the study cohorts and overall design is provided in Supplementary Figure S1.

Muscle mass measurement and sarcopenia definition

Skeletal muscle mass was quantified using the skeletal muscle index (SMI) derived from computed tomography (CT) or positron emission tomography (PET) scans performed within 42 days of plasma collection. Cross-sectional muscle area was measured at the level of the third lumbar vertebra (L3) and normalized to height squared to derive skeletal muscle index (SMI). Low muscle mass was defined using established sex specific cutoffs (52.4 cm²/m² for males and 38.5 cm²/m² for females) [14]. Anthropometer3DNet was also used to extract body composition parameters from CT images. This software, available for research purposes at www.oncometer3d.com, enables automated quantification of anthropometric parameters, including muscle body mass (MBM), from multislice CT scans, based on deep learning-driven segmentation of muscle and adipose tissue (visceral and subcutaneous) [15].

Physical function was approximated using Eastern Cooperative Oncology Group performance status (ECOG PS). To improve phenotype labeling for model training, we defined a high-contrast subgroup. Patients with ECOG PS 0 and high SMI were classified as having a low probability of sarcopenia (LS), whereas patients with low SMI and ECOG PS ≥ 2 were classified as having a high probability of sarcopenia (HS). This subgroup was used for supervised model training, maximizing phenotypic contrast and reducing misclassification.

Plasma proteome

Plasma proteomic profiling was performed using the Olink Proximity Extension Assay (PEA) technology (Olink Proteomics AB, Uppsala, Sweden). Baseline plasma samples from the discovery cohort were analyzed using the Olink Explore 1536 platform, comprising 1,472 protein assays and 48 controls distributed across four panels (inflammation, oncology, cardiometabolic, and neurology). Sequencing was performed on a NovaSeq 6000 system using two S1 flow cells with 2 × 50 bp read lengths.

The validation mCRPC cohort and an external validation cohort from TRACERx were analyzed using the Olink Explore 3072 platform. Both platforms are based on the same Proximity Extension Assay (PEA) technology with next-generation sequencing readout, shared quality control procedures, and NPX normalization. For comparability, analyses were restricted to proteins measured across both platforms. Raw sequencing counts were converted into normalized protein expression (NPX) values following Olink’s standard quality control and normalization pipeline, including internal extension and inter plate controls. NPX values are reported on a log2 scale, with higher values corresponding to higher relative protein abundance. Assay validation metrics are available from the manufacturer.

Cell culture

Immortalized human myoblasts (LHCN-M2) were cultured at 37 °C and 5% CO₂ in growth medium consisting of 60% Dulbecco’s Modified Eagle Medium (DMEM) (Sigma Aldrich, M4530) supplemented with 10% (v/v) fetal bovine serum (FBS) (Life technologies, 10270), 30% (v/v) Medium 199 (Sigma Aldrich, M4530), 10 ng/mL basic fibroblast growth factor (Biolegend, 792506), 0.2 µg/mL dexamethasone (Merck, D4902), and 1% (v/v) penicillin streptomycin (Thermo Fisher Scientific, 15140122). HepG2 cells were cultured in DMEM containing 10% (v/v) FBS and 1% (v/v) penicillin-streptomycin and treated with recombinant human IL-6 (0, 0.01, 0.1, 1, 3, 10, 30 ng/mL) for 24–48 h in DMEM. Following the treatment, cells were harvested for protein and RNA extraction. Intracellular IGFBP1 protein levels were analyzed by western blotting from whole cell lysates, while secreted IGFBP1 was measured by western blot analysis of culture supernatants collected at the indicated time points. Absence of mycoplasma contamination was confirmed using MycoAlert assays (Lonza).

Myoblast differentiation

When the density of LHCN-M2 cells reached 90%, the growth medium was replaced with differentiation medium, which was DMEM medium (Sigma Aldrich, M4530) containing 2% (v/v) FBS (Life technologies, 10270), and 1% (v/v) penicillin-streptomycin solution (Thermo Fisher Scientific, 15140122), with or without different concentration of recombinant human IGFBP1(Bio-techne, 871-B1) or IGFBP2 (Bio-techne, 674-B2) or IL6 (Bio-techne, 206-IL) for 4 days. After 4 days of differentiation, the cells were fixed with 4% paraformaldehyde (PFA) for immunostaining or used for protein extraction.

Western blot

Total protein was extracted from cells using nuclear and cytoplasmic extraction buffer (NETN buffer) containing phosphatase inhibitors (Roche, 11836170001) and protease inhibitors (Roche, 04906837001). The extracted proteins were quantified using the BCA protein Assay kit (Thermo Fisher Scientific, 23227) according to the manufacturer’s protocol. Proteins were separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and transferred to polyvinylidene difluoride (PVDF) membranes (Merck Millipore, USA). The membranes were blocked in 5% non-fat milk, incubated with primary antibodies, and then incubated with appropriate secondary antibodies. Primary antibodies used for western blotting were as follows: Myosin heavy chain (MHC; Bio-Techne, MAB4470), IGFBP1 (Abcam, AB181141), IL-6 (Abcam, AB233706), AKT (Cell Signaling Technology, 4691 S), P-AKT (Cell Signaling Technology, 4060 S), ERK1/2 (Abcam, AB184699), P-ERK1/2 (Abcam, AB76299), and β-actin (Santa Cruz Biotechnology, sc-81178). HRP-conjugated goat anti-rabbit IgG (Jackson ImmunoResearch, 111-035-144) and goat anti-mouse IgG (Jackson ImmunoResearch, 315-035-003) were used as secondary antibodies. Signals were detected by an Amersham Imager 800 system (GE Healthcare) using Chemiluminescent HRP Substrate (Merck Millipore, USA). Band intensities were quantified using Fiji software (https://fiji.sc/) and normalized to β-actin expression. Experiments included three biological replicates.

Immunofluorescent staining

Differentiated myotubes were washed with 1x phosphate-buffered saline (PBS), fixed with 4% paraformaldehyde for 10 min, permeabilized with 0.2% Triton X-100, and blocked with 5% normal goat serum. Cells were incubated with anti-myosin heavy chain antibody (Bio-techne, MAB4470), followed by Alexa Fluor 488 conjugated secondary antibody (Thermo Fisher Scientific, A11001). Nuclei were counterstained with Hoechst 33,342 (Thermo Fisher Scientific, H21492). Images were acquired using the Cytation 1 Cell Imaging Reader (BioTek) at ×10 magnification. Image analysis was conducted using the publicly available image processing software FIJI (https://fiji.sc/). Clusters of nuclei (> 2) per myotube were the standard in calculating the fusion index. Fifteen images per condition were randomly selected and analysed from three independent experiments.

RNA sequencing of myoblasts

LHCN-M2 cells were treated with either 1 µg/mL IGFBP1 or IGFBP2 with or without 150 ng/mL IGF1 in the differentiation medium for 4 days. The experiments were conducted using three biological replicates per condition. Total RNA was extracted using the Quick-RNA kit (Zymo Research, USA).

Tumor RNA sequencing

RNA from MATCHR-I and mCRPC tumor samples was extracted using the AllPrep DNA/RNA Mini Kit (Qiagen). Libraries were sequenced on Illumina NextSeq 500 or NovaSeq platforms as paired-end reads following standard quality control procedures.

Gene expression quantification and enrichment analysis

Raw sequencing reads were quality-controlled using Trim Galore. Transcript quantification was performed using Kallisto (v0.44.0) with the GENCODE v27 reference. Transcript-level estimates were aggregated to gene-level counts using TxImport. Genes with low expression (< 2 counts in 10% of samples) were excluded. Differential expression analysis was performed using DESeq2, applying a false discovery rate (FDR) < 0.05 and |log2 fold change| ≥1.

Gene set enrichment analysis was conducted using clusterProfiler with Hallmark, Gene Ontology biological process, and KEGG gene sets from the Molecular Signatures Database (MSigDB).

Single-sample gene set enrichment analysis

To estimate the activity of muscle-related gene pathways at the sample level, we applied single-sample gene set enrichment analysis (ssGSEA). ssGSEA scores were computed from normalized gene expression counts using the GSVA R package (v2.4.1). Predefined gene sets related to skeletal muscle biology were used to derive enrichment scores for each sample, enabling quantitative comparison of pathway activity across samples.

Single-cell RNA sequencing

Single-nucleus RNA sequencing (snRNA-seq) was performed by Celsius Therapeutics. Nuclei were isolated from frozen tissue samples following standard protocols. Libraries were generated using the 10x Genomics Chromium Single Cell 3′ v3.1 chemistry and sequenced on an Illumina NovaSeq platform. Raw sequencing reads were processed using STARsolo for alignment, barcode processing, and gene counting. Downstream quality control, normalization, and integration were conducted using the Gustave Roussy single-cell analysis pipeline (https://github.com/gustaveroussy/single-cell/wiki/).

Quality control metrics were computed for each nucleus, including the number of detected genes, total unique molecular identifier (UMI) counts, and the proportion of reads mapping to mitochondrial genes. Nuclei were retained for downstream analysis if they expressed at least 200 genes, contained a minimum of 1,000 total UMIs, and had ≤ 20% mitochondrial gene expression. Nuclei not meeting these criteria were excluded.

Data normalization was performed using SCTransform as implemented in Seurat (v4.0.4), regressing out total UMI counts per nucleus to mitigate technical variability. Principal component analysis (PCA) was applied to the normalized data, and batch effects were corrected using the Harmony algorithm (v0.1.0), with PCA embeddings used as input for integration.

The dimensionality and clustering resolution were selected based on visual inspection of Uniform Manifold Approximation and Projection (UMAP) embeddings to identify biologically coherent structures. Cluster stability across parameter combinations was further assessed using the clustree R package (v0.4.3), ensuring robustness of the chosen clustering configuration.

Quantitative real-time PCR

The samples were prepared in biological duplicates. Total RNA was extracted using the Quick-RNA kit (Zymo Research, USA). cDNA was synthesized using HiScript III RT SuperMix (Vazyme, China), and gene expression was analyzed with SYBR qPCR Master Mix (Vazyme, China) on the FMR-5 S Smart Quant System (Vazyme, China). RT-qPCR primers were synthesized by Eurofins (France). Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) was used as an internal control to normalize the qPCR results. Relative gene expression was calculated using the 2^−ΔΔCt method. All reactions were performed in triplicate. Primer sequences are as follows:

• IGFBP1 F: GAAGGAGCCCTGCCGAATA

• IGFBP1 R: TCCATGGATGTCTCACACTGTC

• GAPDH F: CTGCACCACCAACTGCTTAG

• GAPDH R: AGGTCCACCACTGACACGTT

Statistical analysis

Statistical analyses were performed using R 4.1 and GraphPad Prism 9 (GraphPad Software, San Diego, CA, USA). Data from in vitro experiments are presented as mean ± standard deviation (SD) from at least three independent biological replicates. Comparisons between two groups were conducted using Student’s t-test, and when appropriate the Wilcoxon rank-sum test, while multiple group comparisons were performed using one-way ANOVA. P-values < 0.05 were considered statistically significant.

Kaplan–Meier and Cox proportional hazards models were used for survival analysis. Sarcopenia probability was applied as a grouping variable in Kaplan–Meier curves and as a continuous covariate in the Cox models. Spearman correlation analysis was performed to evaluate the association between predicted sarcopenia probability and SMI. Linear regression models were used to assess the independent association between model-derived sarcopenia probability and clinical variables. Model performance was evaluated using the coefficient of determination (R2), which was also used to assess the relative importance of variables based on their contribution to explained variance.

Machine learning

Plasma proteins were initially screened based on differential expression between high-contrast sarcopenia groups (HS vs. LS) using the Wilcoxon rank-sum test with false discovery rate correction. Candidate proteins were further prioritized based on enrichment in non-sarcopenic patients and biological relevance, supported by annotation from the Human Protein Atlas [16], to enhance interpretability and reduce dimensionality.

A decision tree based extreme gradient boosting (XGBoost) classification model was developed using the tidymodels framework [17]. The model was trained on the high-contrast subgroup of lung cancer patients using fivefold cross-validation with 20 repetitions, stratified by sarcopenia probability. Hyperparameters were optimized based on classification accuracy.

The final model was applied to the full discovery cohort to generate a continuous sarcopenia probability score for each patient. This score was subsequently used for survival analyses and correlation with clinical variables.

The model was independently validated in the mCRPC and in the TRACERx cohort. To derive a reduced plasma protein signature, a second XGBoost regression model was trained using predicted sarcopenia probabilities as the outcome on the MATCHR-I and mCRPC using the tidymodels framework and fivefold cross-validation with 20 repetitions Feature selection was performed iteratively based on variable importance, with root mean square error used for model selection. Performance of the extended and reduced signatures was compared in the external validation cohort.

Results

Patient cohorts and definition of high-contrast sarcopenia groups

To develop and validate a plasma proteomic signature of sarcopenia, we analyzed patients from the MATCH-R study and two independent validation cohorts. Rather than focusing on a single tumor type, we intentionally included heterogeneous populations to assess the robustness and generalizability of the identified proteomic signature across different cancer contexts. We integrated imaging-based muscle mass measurements with functional status to define high-contrast phenotypes for model training and testing. The training cohort included 99 patients with advanced cancers, mainly lung and bladder, treated with immunotherapy from the MATCH-R study. Skeletal muscle mass was assessed using the Skeletal Muscle Index (SMI) at the L3 level on CT or PET scans performed within 42 days of plasma collection. All CT scans and blood samples were collected prior to treatment initiation. The median time between CT scan and blood sampling was − 12.5 days (IQR − 22.3 to − 1.0), with a range from − 42 to + 38 days. Functional status was evaluated using the ECOG performance status (PS) at the time of blood sampling. Among patients with imaging data, 51 had low SMI and 34 had high SMI. To define a high-contrast subgroup, SMI was combined with ECOG PS: patients with high SMI and ECOG 0 were classified as low sarcopenia probability (LS), while those with low SMI and ECOG ≥ 2 were classified as high sarcopenia probability (HS). This yielded 36 patients (21 HS, 15 LS) for model training. Baseline characteristics are summarized in Table 1. For validation, two independent cohorts were used. The first was an internal MATCH-R cohort of 55 patients with metastatic castration-resistant prostate cancer treated with enzalutamide, with 88 plasma samples collected at baseline and progression. The second was an external TRACERx cohort of 151 patients with non-small cell lung cancer, including 141 baseline and 114 relapse samples. These cohorts enabled evaluation of model performance across different cancer types and clinical settings.

Table 1.

Clinical characteristics of the MATCH-R cohorts used for model development and validation. Baseline demographic, clinical, and disease characteristics of patients included in the training cohort (MATCH-R immunotherapy) and in the validation cohort (MATCH-R metastatic castration-resistant prostate cancer). Continuous variables are reported as median (interquartile range), and categorical variables as number (percentage)

Characteristic	Training Cohort		Validation N = 881
	High Contrast N = 361	Low Contrast N = 631
Age	68 (58, 72)	62 (56, 71)	69 (65, 77)
Sex
Female	10 (28%)	26 (41%)	0 (0%)
Male	26 (72%)	37 (59%)	88 (100%)
ECOG
0	15 (39%)	16 (25%)	26 (46%)
1	0 (0%)	38 (60%)	19 (33%)
≥2	21 (58%)	9 (14%)	12 (21%)
Unknown	0	0	31
Site
Bladder cancer	4 (11%)	8 (13%)	0 (0%)
Colon cancer MSI	1 (2.8%)	2 (3.2%)	0 (0%)
NSCLC adenocarcinoma	24 (67%)	42 (67%)	0 (0%)
NSCLC undifferentiated/NOS	2 (5.6%)	2 (3.2%)	0 (0%)
NSCLC squamous	4 (11%)	7 (11%)	0 (0%)
mCRPC	1 (2.8%)	0 (0%)	88 (100%)
Melanome	0 (0%)	1 (1.6%)	0 (0%)
Thyroid	0 (0%)	1 (1.6%)	0 (0%)
Sarcopenia score
Probability ≥ 50%	21 (58%)	40 (63%)	13 (15%)
Probability < 50%	15 (42%)	23 (37%)	75 (85%)
Cardiac morbidity	3 (8.3%)	7 (11%)	4 (4.5%)
High Blood Pressure	12 (33%)	20 (32%)	20 (23%)
Diabetes	4 (11%)	4 (6.3%)	11 (13%)
Metformin use	2 (5.6%)	1 (1.6%)	2 (2.3%)
Corticosteroid use	7 (19%)	10 (16%)	6 (6.8%)

¹ Median (Q1, Q3); n (%)

Discovery of a proteomic signature of sarcopenia

We performed plasma proteomics using the Olink Explore platform in the training cohort. Differential expression analysis between HS and LS patients in the high-contrast subgroup identified 353 proteins enriched in HS, predominantly associated with systemic inflammation, immune dysregulation, matrix remodeling, and metabolic stress, and 64 proteins enriched in LS, many of which were related to muscle or neuronal biology (Supplementary Figure S1, Supplementary Table 1).

To focus on proteins reflecting muscle function rather than tumor biology, we prioritized proteins enriched in LS and underrepresented in HS. A biology-driven selection further narrowed candidates to those implicated in skeletal muscle homeostasis (e.g., MSTN, WFIKKN1, CD34, PAMR1, DPP4), neuromuscular connectivity (e.g., CNTN family, adhesion/guidance molecules), extracellular matrix remodeling (e.g., integrins), and autophagy/lysosomal pathways (e.g., LAMP2, IDS). Proteins primarily involved in systemic metabolism, immunity, endocrine regulation, or oncogenic signaling were excluded.

This process resulted in a 23-protein panel that served as the basis for the XGBoost classification model (model p23; Supplementary Table S2). The model was trained to predict sarcopenia probability (SP) as a continuous variable and classify patients as sarcopenic (SP > 50%) or non-sarcopenic (SP ≤ 50%) within the high-contrast training set. Given the limited size of the high-contrast training set, model development relied on repeated cross-validation with stratification by sarcopenia status to ensure balanced representation across folds.

Model performance and association with SMI and ECOG

The XGBoost model (p23) was first evaluated within the high-contrast training set. The model achieved an accuracy of 0.893 in classifying patients as sarcopenic or non-sarcopenic, demonstrating robust discrimination between HS and LS cases (Supplementary Figure S2). When applied to the full training cohort, the predicted probability of sarcopenia (SP) correlated inversely with SMI in both male and female patients (Spearman ρ = -0.39, p = 0.0045 and ρ = -0.42, p = 0.016 respectively; Fig. 1A), confirming that the model captured meaningful biological variation in muscle mass.

Fig. 1.

Association between proteomic-based probability of sarcopenia, muscle mass, and clinical performance. A-D Correlation between the proteomic-based probability of sarcopenia and skeletal muscle index in the training cohort A, the mCRPC validation cohort B, and the TRACERx cohort at baseline C and at relapse D. E,F Box plots showing the distribution of the proteomic-based probability of sarcopenia across ECOG performance status categories in the training cohort E and in the mCRPC validation cohort F

SP also reflected functional status: increasing SP was associated with higher ECOG PS scores, indicating concordance between the proteomic signature and clinically assessed physical function (Fig. 1E).

In the internal validation cohort of 55 mCRPC patients, the model maintained a similar correlation with SMI (ρ = -0.41, p = 0.008) and ECOG PS (Fig. 1B F respectively), and in the subset of 18 high-contrast cases, classification accuracy reached 0.889, comparable to the training set (Fig. 3A).

Fig. 3.

Validation and longitudinal changes in proteomic-based sarcopenia probability. A Distribution of proteomic-based sarcopenia probability in the high-contrast validation cohort (LS: low probability of sarcopenia ; HS: high probability of sarcopenia). B Change in proteomic-based sarcopenia probability between paired baseline and progressive disease samples in the MATCH-R prostate cohort, stratified according to changes in ECOG performance status. C Correlation between proportional changes in skeletal muscle index and changes in proteomic-based sarcopenia probability in paired baseline and relapse samples from the TRACERx cohort. Dot color indicates the disease-free interval between the two time points. D Heatmap showing the correlation between candidate circulating mediators and the proteomic-based probability of sarcopenia in the 4 cohorts, namely MATCH-R immuno (MR immuno), MATCH-R metastatic castration resistant prostate cancer (mCRPC), TRACERx recurrence (Rrx_rec) and TRACERx baseline (Rrx_B)

Importantly, in the external TRACERx cohort, which included both baseline and relapse samples, SP also correlated with cross-sectional skeletal muscle area at L3 (for male: baseline ρ = -0.29, p 0.012; relapse ρ = -0.42, p = 0.0018; for female: baseline ρ = -0.24, p 0.073; relapse ρ = -0.35, p = 0.047), confirming the generalizability of the model across independent patient populations (Fig. 1C-D). To further dissect the respective contributions of body composition and global clinical status, we performed multivariable analyses including both SMI and ECOG PS. Both variables were independently associated with SP, although ECOG explained a larger proportion of variance (R2 0.1 vs. 0.4), consistent with its role as a global indicator of patient condition. Importantly, SMI retained an independent contribution, supporting the ability of the proteomic signature to capture muscle-specific biological information beyond general performance status (Supplementary Table S3). Moreover, to further investigate the potential confounders, we explored the impact of patients’ therapies and comorbidities No significant association was observed between SP and hypertension (p = 0.26), diabetes (p = 0.53), metformin use, or cardiovascular disease (both p > 0.8). Similarly, no correlation was found with age (Spearman’s ρ, p = 0.67). These findings are consistent with previous evidence suggesting that cancer-related factors may play a predominant role in the development of sarcopenia, potentially outweighing the contribution of age and comorbidities.

Despite the differences in terms of tumor types, the proteomic score showed consistent associations with muscle mass across cohorts, supporting the existence of shared biological mechanisms underlying cancer-related sarcopenia.

Together, these findings demonstrate that the 23-protein signature provides a robust, scalable measure of sarcopenia probability, consistently reflecting both muscle mass and functional status across multiple cancer types and cohorts.

Prognostic impact of the proteomic sarcopenia signature

We next investigated the prognostic relevance of the proteomic-based probability of sarcopenia (SP) across cohorts. As expected, in the high-contrast training set, SP strongly stratified overall survival (OS), with sarcopenic patients (SP > 50%) exhibiting a markedly shorter median OS compared with non-sarcopenic patients (3.8 vs. 34.8 months; log-rank p < 0.0001; Fig. 2A). Importantly, this prognostic separation was preserved when extending the analysis to the remaining, lower-contrast patients in the training cohort, in whom sarcopenic patients also experienced significantly worse outcomes (median OS 7.8 vs. 22.1 months; p = 0.012; Fig. 2B).

Fig. 2.

Association between proteomic-based probability of sarcopenia and overall survival. A,B Overall survival according to the proteomic-based probability of sarcopenia in the high-contrast training cohort A and in the remaining patients from the training cohort with lower contrast B. C Overall survival stratified by proteomic-based probability of sarcopenia in the mCRPC validation cohort at baseline. D,E Overall survival according to the proteomic-based probability of sarcopenia in the TRACERx external validation cohort at baseline D and at relapse E. Time is expressed in months

In the internal validation cohort of patients with metastatic castration-resistant prostate cancer (mCRPC), baseline SP similarly identified a subgroup with significantly inferior survival. Sarcopenic patients had a median OS of 8.9 months compared with 21.9 months in non-sarcopenic patients (p < 0.0001; Fig. 2C), confirming the prognostic robustness of the signature in a cancer type with distinct biology and treatment context.

Consistent results were observed in the external TRACERx cohort. Across resected and relapsed NSCLC samples, sarcopenic patients had significantly shorter OS compared with non-sarcopenic patients (median OS 25 vs. 46 months p < 0.001 for baseline; median OS 30 vs. 44 months, p = 0.043 at relapse, Fig. 2D–E). Notably, in this cohort, the proteomic sarcopenia classification only partially overlapped with cachexia status, with most cachectic patients being sarcopenic but a substantial proportion of sarcopenic patients not fulfilling cachexia criteria, reinforcing the biological and clinical distinction between these two syndromes.

Multivariable Cox proportional hazards models were used to assess the independent prognostic value of the proteomic sarcopenia probability across cohorts, adjusting for available clinical covariates.

In the MATCHR-I cohort, sarcopenia probability remained independently associated with overall survival after adjustment for skeletal muscle index at L3, ECOG performance status, age, and sex (HR 4.13, 95% CI 1.38–12.3, p = 0.011), while SMI showed only a trend towards significance.

In the mCRPC validation cohort, sarcopenia probability and ECOG performance status were both independently associated with survival, whereas age was not. SMI could not be included due to a high rate of missing imaging data (> 50%), reflecting routine clinical follow-up in mCRPC patients, which is primarily based on PSA monitoring rather than systematic radiological assessments.

In the external TRACERx cohort, sarcopenia probability was consistently associated with overall survival both at baseline and at relapse, independently of age, sex, and cross-sectional muscle area, which was not significantly associated with outcome in multivariable analyses. (supplementary table S4)

Together, these results demonstrate that the plasma proteomic sarcopenia signature captures clinically meaningful systemic vulnerability that translates into a strong and consistent prognostic impact across tumor types, disease stages, and independent patient cohorts.

SP reflects longitudinal changes in sarcopenia

To determine whether the proteomic-based probability of sarcopenia (SP) captures dynamic changes in muscle health over time, we investigated its longitudinal behavior in patients with paired plasma samples collected at baseline and at disease progression or relapse.

In the mCRPC validation cohort, paired proteomic samples were available for 33 patients treated with enzalutamide, of whom 20 had contemporaneous clinical documentation of ECOG performance status at both time points. Changes in SP closely mirrored clinical trajectories. Patients whose performance status improved between baseline and progression consistently showed a marked reduction in SP, with both individuals transitioning from sarcopenic to non-sarcopenic classification and exhibiting a mean SP reduction of 58%. In contrast, patients with stable ECOG PS showed minimal variation in SP, whereas those experiencing clinical deterioration displayed a corresponding increase in sarcopenia probability (Fig. 3B). These findings support the ability of the proteomic signature to track clinically meaningful changes in functional status over time.

We next extended this analysis to the external TRACERx cohort, where paired plasma proteomics and quantitative muscle mass assessments were available for 73 patients at baseline and relapse. In this independent dataset, longitudinal changes in SP were significantly correlated with proportional changes in skeletal muscle mass between the two time points (Spearman ρ = −0.32, p = 0.005; Fig. 3C). Notably, the strength of this association appeared modulated by the disease-free interval separating baseline and relapse, suggesting that longer inter-sample intervals may allow for more pronounced and biologically detectable muscle remodeling.

Collectively, these longitudinal analyses demonstrate that the proteomic sarcopenia signature is not merely a static classifier but reflects dynamic changes in muscle health across disease evolution, treatment exposure, and relapse, reinforcing its potential utility for real-time patient monitoring.

Reduced four-protein model correctly predicts sarcopenia

As a final feature selection step, we sought to determine whether a more streamlined proteomic signature, consisting of a reduced number of proteins, could still reproduce the classification performance of our original model. We then trained an Xgboost regression model on MATCH-R cohorts and found that a model comprising a progressively reduced number of proteins maintained a good correlation with initial signature, with a model based on 4 proteins (model P4, based on CNTN3, CBLN4, MSTN, and ITGA11) resulting in an RMSE of 0.053 on the validation fold (supplementary figure S2 and supplementary table S5). Model p4 showed similar result in the high contrast groups, and AUC of 1, in the discovery cohort, and an accuracy of 0.83 and AUC of 1 on the prostate cohort. On the overall cohorts, both prediction from the 2 models showed a high correlation of 0.92, 0.98 and up to 0.99. Thus, the results from the classification model were closely approximated by the regression models with lesser plasma proteins.

p23 signature reveals potential mediators of sarcopenia

To identify potential circulating mediators underlying the proteomic signature of sarcopenia, we treated the signature-derived probability of sarcopenia as a continuous variable and investigated its association with individual plasma proteins across cohorts. This strategy allowed us to move beyond classification performance and interrogate biological processes consistently linked to sarcopenia severity.

Across the discovery cohort, several plasma proteins showed significant positive correlations with the probability of sarcopenia after false discovery rate correction. When extending this analysis to the internal validation cohort of patients with metastatic castration-resistant prostate cancer and to the external TRACERx cohort, IGFBP1, IGFBP2 and IL6 emerged as the most consistently associated proteins, showing significant positive correlations with sarcopenia probability in all evaluated datasets (Supplementary Table S6 and Fig. 3D).

We also observed several proteins with known or suspected roles in muscle pathophysiology. These include proteins implicated in muscle disease (e.g., SLC39A14 [18]), markers of muscle damage (CKB, SERPINA3 [19]) and ITIH3, previously associated with disease activity in myastenia gravis [20] and recently associated with muscle wasting [21].

Single-cell transcriptomics reveals the origin of potential mediators of sarcopenia

We next investigated the transcriptomic correlates of our sarcopenia signature. Bulk RNA-seq analyses performed in our two internal cohorts revealed no meaningful correlation between IGFBP1/2 plasma levels and their corresponding mRNA expression in tumor tissue. At the pathway level, however, increasing sarcopenia probability was consistently associated with suppression of the Hallmark myogenesis program as well as multiple GO Biological Process pathways related to muscle development and function, supporting a muscle-specific biological signal captured by our signature (Fig. 4A-D).

Fig. 4.

Transcriptomic correlates of the plasma proteomic signature of sarcopenia. A, B Gene set enrichment analysis (GSEA) of Hallmark pathways associated with the plasma proteomic signature of sarcopenia in the MATCH-R immunotherapy cohort A and the MATCH-R metastatic castration-resistant prostate cancer cohort B. C, D Gene Ontology Biological Process (GOBP) pathway enrichment analysis showing activation and suppression patterns associated with the proteomic signature of sarcopenia in the MATCH-R immunotherapy C and MATCH-R prostate D cohorts. E-G Single-cell RNA sequencing analysis showing cell type–specific expression of IGFBP1 in hepatic cells F and IGFBP2 in tumor cells G

To explore the potential cellular origin of putative mediators, we analysed single-cell RNA sequencing data generated from a subset of 16 patients from the MATCHR-I study (clinical characteristics reported in Supplementary Table S6). IGFBP1 expression was predominantly enriched in hepatocytes, together with ITIH3, thus validating in patients what was recently described in murine models [21]. In contrast, IGFBP2 expression was mainly restricted to a subset of tumor cells. IL6 did not map clearly to a specific cellular compartment, a finding that may reflect its multifocal origin, including potential production by skeletal muscle, a compartment not represented in tumor biopsies (Fig. 4 and supplementary Fig. 3).

IGFBP1 and IGFBP2 impair myogenic differentiation

To provide functional support for candidate circulating mediators identified by the proteomic signature, we selected a subset of proteins showing strong association with sarcopenia probability (SP) and consistent behavior across cohorts, and investigated their effects on muscle cells in vitro. Given that skeletal muscle homeostasis relies on efficient myoblast differentiation and regeneration, processes known to be impaired in sarcopenia [22], we examined whether IGFBP1, IGFBP2, and IL-6 directly interfere with myogenic differentiation.

We first determined the concentrations of IGFBP1 and IGFBP2 in a subset of non-sarcopenic vs. sarcopenic patients in the training cohort using ELISA. Concentrations of IGFBP1 were relatively low (3.82 vs. 23.80 ng/ml), while IGFBP2 concentrations were substantially higher (0.56 vs. 1.14 µg/ml, Supplementary Figure S4 and supplementary table S7).

Immortalized human myoblasts were induced to differentiate and exposed to increasing concentrations of IGFBP1, IGFBP2, and IL-6 as detailed in the Methods. After four days of differentiation, myogenic progression was assessed by immunostaining and western blotting.

Both IGFBP1 and IGFBP2 impaired myogenic differentiation in a dose-dependent manner, with a significant reduction in fusion index (1.5-fold and 2-fold, respectively; Fig. 5A–C) and decreased MHC protein expression (2-fold and 4-fold, respectively; Fig. 5D–F). Notably, the effects of IGFBP2 were observed at concentrations comparable to those measured in patient samples. For IGFBP1, in vitro effects were detected at higher concentrations than those observed clinically, consistent with prior studies reporting similar requirements to elicit functional effects in muscle cells [21].

Fig. 5.

IGFBP1 and IGFBP2 inhibit myogenic differentiation via an IGF-1–dependent mechanism. A–F LHCN-M2 human myoblasts were grown to 100% confluence and induced to differentiate for 96 h in DMEM supplemented with 2% FBS in the presence or absence of increasing concentrations of IGFBP1 or IGFBP2 (0.1, 1, 2.5 and 5 µg/mL). A Immunostaining analysis of LHCN-M2 cells induced to differentiate in the presence of either IGFBP1 or IGFBP2. Cells were stained with Hoechst 33342 (blue) and myosin heavy chain (MHC, green) antibody to visualize nuclei and myosin, respectively. Scale bar, 200 µm. B, C Fusion index of  LHCN-M2 cells treated with either IGFBP1 B or IGFBP2 and induced to differentiate. D Western blot analysis of MHC protein expression LHCN-M2 cells treated with IGFBP1 or IGFBP2 and induced to differentiate. E, F Densitometric quantification of MHC protein levels in IGFBP1-treated E and IGFBP2-treated F cells, normalized to β-actin **P < 0.01, ***P < 0.001, ****P < 0.0001, vs. untreated control. G–L LHCN-M2 cells were grown to confluence and differentiated for 96 h in the presence of IGFBP1 (1 µg/mL) or IGFBP2 (1 µg/mL), with or without increasing concentrations of IGF-1 (0, 50, 150 or 300 ng/mL). Experimental conditions included the control (2% FBS only), IGFBP1 or IGFBP2 alone, IGF-1 alone, and combined IGFBP1+IGF-1 or IGFBP2+IGF-1 treatments. G–J Immunostaining analysis of the effect of IGF-1 on IGFBP1- or IGFBP2-mediated inhibition of myogenic differentiation. H, J Fusion index of LHCN-M2 cells treated with IGFBP1+IGF-1 H or IGFBP2+IGF-1 J and induced to differentiate. K, L Western blot analysis K and quantification L of MHC protein expression demonstrating partial or complete rescue of myogenic differentiation upon IGF-1 supplementation in the presence of IGFBP1 or IGFBP2. #p<0.05, ##p<0.01, ####p<0.0001, vs. control; *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, vs. IGFBP1-only or IGFBP2-only conditions

Given that IGFBPs can exert both IGF-1–dependent and IGF-1–independent biological effects²¹, we next evaluated whether increasing IGF-1 availability could rescue this phenotype. Indeed, supplementation with increasing concentrations of IGF-1 fully restored myoblast differentiation in the presence of both IGFBP1 and IGFBP2, supporting a predominantly IGF-1–dependent mechanism whereby IGFBPs sequester IGF-1 and suppress its pro-myogenic signaling (Fig. 5G–J).

In contrast, exposure to IL-6 did not impair myoblast differentiation or MHC expression, indicating that IL-6 does not act as a direct inhibitor of myogenesis in this experimental context (Supplementary Figure S5A-C). Based on prior evidence indicating that IL-6 can induce hepatic IGFBP1 expression [23], we next explored this axis in liver-derived cells. IL-6 exposure led to increased IGFBP1 expression in HepG2 cells, accompanied by elevated levels of secreted IGFBP1 in the culture supernatant, supporting an indirect mechanism whereby systemic inflammation may promote muscle dysfunction through liver-mediated modulation of the IGF axis rather than through direct effects on muscle cells (Supplementary Figure S5D-F).

IGFBP1 and IGFBP2 perturb myogenesis and IGF1/KRAS signaling.

To gain insight into the molecular mechanisms of myogenesis inhibition by IGFBP1 and IGFBP2, we treated LHCN-M2 human myoblasts with IGFBP1, IGFBP2 or the IGFBP1 + IGF1 or IGFBP2 + IGF1 and induced them to differentiate as described in Materials and Methods. RNA-seq analysis identified 158 and 286 misregulated genes in LHCN-M2 cells treated with IGFBP1 and IGFBP2, respectively (Fig. 6A, B; Supplementary Tables S8-S14), with largely overlapping dysregulated pathways induced by both peptides. The most affected pathways were MYOGENESIS and KRAS_SIGNALING (Fig. 6C), the former being largely suppressed and the latter being overactive in the IGFBP-treated cells (Fig. 6E). IGF1 and KRAS signaling are functionally interconnected through shared downstream pathways, and misregulation of KRAS signaling has been linked to impaired muscle differentiation [24]. Genes of the KRAS_SIGNALING cascade upregulated in IGFBP-treated cells include SERPINA, frequently upregulated in inflammatory and degenerative muscle diseases [25] and cell cycle activator CCND2, a potent inhibitor of muscle differentiation [26].

Fig. 6.

Effect of IGFBP and IGFBP2 on gene expression in differentiating human myoblasts. LHCN-M2 human myoblasts were treated with IGFBP1, IGFBP2 or the IGFBP1+IGF1 or IGFBP2+IGF1 and induced to differentiate. A PCA analysis of the samples; B Volcano plots showing the genes deregulated in the LHCN-M2 cells treated with either IGFBP or  IGFBP2 vs. untreated control. Top 10 genes by padj deregulated in each direction are labeled. C, Overrepresentation analysis against the HALLMARK genesets of the MSigDB for the genes significantly deregulated in the LHCN-M2 cells treated with either IGFBP or  IGFBP2 vs. untreated control. D, Scatter plots showing the relationship between gene expression changes in LHCN-M2 cells treated with either IGFBP1 (left)/ IGFBP2 (right) and cells  treated by a combination of either IGFBP1+IGF1 or  IGFBP2+IGF1. The genes significantly deregulated in opposite directions in these comparisons are highlighted. Top 10 genes by padj deregulated in each direction are labeled. Pearson R and p-value are reported. E, Pathway activity scores for the MYOGENESIS, KRAS_SIGNALLING_UP and KRAS_SIGNALLING_DN pathways in  the LHCN-M2  cells treated with either IGFBP or  IGFBP2 vs. untreated control. t-test p-values are reported

IGFBP1- and IGFBP2-treated cells also exhibited downregulation of key myogenic genes, including TNNT1, TNNT3 (encoding the fast skeletal muscle isoform of troponin T, that is strongly upregulated at late stages of differentiation and whose disregulation has been found in impaired muscle differentiation and congenital myopathies [27]) and multiple MYH isoformes (MYH3, MYH7, MYH8), whose inactivation leads to impaired muscle function and myopathies [28]. Importantly, these transcriptional alterations were largely attenuated by concomitant IGF1 supplementation that substantially restored the activity of KRAS and myogenesis pathways (Fig. 6E), as well as many other genes deregulated by IGFBP1/2 (Fig. 6D). This further supports a central role for impaired IGF1 signaling in mediating the observed differentiation defects.

To validate misregulation of the IGF1/KRAS pathway, we treated LHCN cells with either IGFBP1, IGFBP2 or IGFBP1 + IGF1, IGFBP2 + IGF1 combination in differentiation medium for 48 h. Both IGFBP1 and IGFBP2 could inhibit the phosphorylation of AKT, a key downstream effector of IGF1 signaling, an effect that was reversed by IGF1 supplementation (Supplementary Figure S6).

Discussion

In this study, we demonstrate that plasma proteomics can capture a biologically meaningful representation of physical function and muscle mass in oncology patients, offering a potentially scalable and objective alternative to conventional clinical assessments. We used a two step approach. First, by focusing on proteins enriched in non-sarcopenic individuals, we identified a proteomic signature reflective of preserved systemic and neuromuscular integrity, rather than advanced cancer-related wasting. This enabled the delineation of patient subgroups with distinct muscle mass and functional phenotypes based on molecular features, providing biological resolution that is not achievable using performance status alone. Importantly, our approach estimates a continuous probability of sarcopenia-related muscle dysfunction, rather than assigning a binary diagnosis, reflecting the biological heterogeneity and absence of a single gold-standard definition in oncology. Secondly, after SP computation, we identified candidate mediators enriched in sarcopenic patients.

Our findings suggest that, beyond tumor-specific effects, common host-related pathways may contribute to muscle dysfunction across cancer types, supporting the hypothesis that systemic factors, including components of the IGFBP axis, may play a central role in mediating these effects.

Furthermore, although the prostate cancer cohort included only male patients, the training and external validation cohorts comprised both sexes. The consistent relationship between the proteomic signature and muscle mass observed in both male and female patients suggests that the identified signature captures core features of muscle biology that are largely independent of sex. Nevertheless, sex-specific differences in body composition and hormonal regulation may modulate the development of sarcopenia and deserve further investigation in dedicated studies.

Despite major advances in precision oncology, patient stratification in clinical practice remains largely tumor-centric, with limited incorporation of the host systemic state and its biological determinants. Skeletal muscle mass is increasingly assessed using imaging-based approaches, yet these measures are inconsistently implemented, and provide limited insight into the underlying biology of muscle dysfunction. Functional status is routinely assessed using coarse clinical scales such as the Eastern Cooperative Oncology Group performance status (ECOG PS), which suffer from moderate interobserver agreement [29], limited resolution, particularly within intermediate categories [30], and lack direct biological interpretability. Consequently, the molecular processes underlying muscle loss, functional decline and sarcopenia in patients with cancer remain poorly characterized and difficult to quantify at scale. This limitation has important implications for interventional studies in cancer sarcopenia, including trials of exercise, nutritional, or pharmacological interventions, which critically rely on robust, objective, and scalable endpoints. The lack of biologically grounded surrogate markers of muscle function and repair has contributed to heterogeneous trial designs and inconclusive results across the field.

A notable aspect of the identified signature is the enrichment of proteins linked to neuronal and neuromuscular biology. In particular, CBLN4 and CNTN3 emerged as major contributors to the model. Both proteins are involved in synaptic organization and plasticity, suggesting that preserved physical function in patients with cancer may depend not only on muscle-intrinsic properties but also on intact neuromuscular and neuronal signaling. This observation aligns with growing evidence that muscle aging and sarcopenia are influenced by alterations at the neuromuscular junction [31] and by central neural regulation [32, 33]. However, given the absence of direct neuromuscular or functional measurements in our study, these interpretations should be considered hypothesis-generating.

Beyond descriptive stratification, our analyses identified candidate mediators of cancer-related sarcopenia. Among the most consistently correlated proteins across cohorts were IGFBP1 and IGFBP2, both key regulators of the insulin-like growth factor (IGF) axis. While liver-derived IGFBP1 has recently been implicated in muscle wasting [21], our data extend this paradigm by identifying tumor-derived IGFBP2 as an active contributor to muscle dysfunction. Functional experiments suggests that IGFBP2 impairs myoblast differentiation, thereby limiting muscle repair and regeneration, an effect that could be reversed by increasing IGF1 availability. At the molecular level, transcriptomic analyses revealed coordinated suppression of myogenesis-related pathways and activation of KRAS signaling in IGFBP-treated cells, both of which were largely rescued by IGF1 supplementation. These findings provide mechanistic support for a model in which disruption of IGF1 signaling, coupled with aberrant activation of KRAS-related pathways, contributes to impaired muscle regeneration in cancer-associated sarcopenia. These findings support a model in which tumors may contribute to sarcopenia through endocrine or paracrine modulation of IGF signaling.

The broader proteomic landscape associated with the signature further highlights the systemic nature of cancer-related sarcopenia. Several correlated proteins are involved in inflammatory signaling and acute-phase responses, including IL-6, MDK [34], and C9, consistent with the established role of chronic inflammation in muscle wasting. Others have been linked to insulin resistance at the muscular level, a metabolic alteration increasingly recognized as a contributor to sarcopenia development. Additional proteins associated with extracellular matrix remodeling, tumor proliferation, and invasion likely reflect aggressive tumor biology, which may indirectly exacerbate functional decline through sustained systemic stress rather than direct effects on muscle tissue. Together, these findings suggest that sarcopenia in cancer emerges from the convergence of inflammatory, metabolic, and tumor-derived signals, resembling an accelerated state of inflammaging.

Notably, the incomplete overlap between sarcopenia and cachexia observed in the external validation cohort reinforces the concept that they represent overlapping but distinct biological conditions. This distinction highlights the limitation of weight-based definitions and supports the need for muscle-specific, biology-driven stratification approaches.

In clinical practice, such a blood-based signature could complement imaging and functional assessments by enabling scalable screening, longitudinal monitoring, and early identification of patients who may benefit from supportive interventions.

This potential is consistent with recent longitudinal proteomic studies in non-cancer populations, which have demonstrated that circulating protein signatures can capture sarcopenia progression and enable early detection of functional decline [35].

Beyond its biological insights, our findings illustrate how molecular phenotyping may contribute to more individualized approaches to patient management in oncology. By combining plasma proteomics with machine learning, our approach enables scalable, biologically informed stratification of muscle health and physical function.

This study has several strengths, including the use of a large, multi-cancer cohort and a broad plasma proteomic panel, supporting the generalizability of the identified biology across tumor types. At the same time, the relatively small size of the training set represents a limitation; however, the use of a high-contrast design allowed us to maximize biological signal and reduce misclassification. The validation across cancer types suggest that cancer-related sarcopenia is accompanied by shared systemic mechanisms rather than being entirely tumor-specific. Limitations include the absence of direct measurements of muscle strength; however, it is noteworthy that most sarcopenia studies rely primarily on radiological assessments of muscle mass. We mitigated this limitation by incorporating performance status as a functional proxy and by training the model on patients with highly contrasted clinical and radiological phenotypes.

Future studies integrating matched tumor, muscle, and plasma samples will be essential to further dissect the causal pathways linking tumor biology to systemic functional decline. In particular, combining peripheral proteomics with direct assessments of muscle biology may help refine understanding of the mechanisms involved. Ultimately, such integrative approaches could enable biologically informed patient stratification and support the development of targeted interventions aimed at preserving physical function in oncology patients. More broadly, the identification of plasma-based, mechanism-linked signatures of sarcopenia may help establish objective and scalable surrogate endpoints for future clinical trials in cancer cachexia and sarcopenia, including exercise-based interventions, thereby accelerating therapeutic development.

Abbreviations

MATCH

R-Molecular Analysis for Therapy Choice-Resistance study

mCRPC

Metastatic Castration-Resistant Prostate Cancer

TRACERx

Tracking Cancer Evolution through therapy (Rx) study

L3

Third Lumbar vertebra

ECOG

Eastern Cooperative Oncology Group

XGBoost

Extreme Gradient Boosting model

SP

Sarcopenia Probability

IGFBP1

Insulin-like Growth Factor-Binding Protein 1

IGFBP2

Insulin-like Growth Factor-Binding Protein 2

IL6

Interleukin-6

CPP

Comité de Protection des Personnes

ANSM

Agence Nationale de Sécurité du Médicament et des produits de santé

SMI

Skeletal Muscle Index

CT

Computed Tomography

PET

Positron Emission Tomography

IQR

Interquartile Range

MBM

Muscle Body Mass

ECOG PS

Eastern Cooperative Oncology Group Performance Status

LS

Low probability of Sarcopenia

HS

High probability of Sarcopenia

PEA

Proximity Extension Assay

NPX

Normalized Protein Expression

DMEM

Dulbecco’s Modified Eagle Medium

PBS

Phosphate-Buffered Saline

FDR

False Discovery Rate

MSigDB

Molecular Signatures Database

KEGG

Kyoto Encyclopedia of Genes and Genomes

ssGSEA

Single-sample Gene Set Enrichment Analysis

GSVA

Gene Set Variation Analysis

snRNA

Seq-single-nucleus RNA sequencing

UMI

Unique Molecular Identifier

PCA

Principal Component Analysis

UMAP

Uniform Manifold Approximation and Projection

NSCLC

Non-Small Cell Lung Cancer

MSTN

Myostatin

PAMR1

Peptidase domain-containing Associated with Muscle Regeneration 1

DPP4

Dipeptidyl Peptidase 4

CNTN

Contactin

LAMP2

Lysosome-Associated Membrane Protein 2

IDS

Iduronate 2-Sulfatase

OS

Overall Survival

CI

Confidence Interval

HR

Hazard Ratio

PSA

Prostate-Specific Antigen

CBLN4

Cerebellin-4

ITGA11

Integrin alpha-11

RMSE

Root Mean Square Error

AUC

Area Under the Curve

CKB

Creatine Kinase B

SERPINA3

Serpin Family A Member 3

ITIH3

Inter-alpha-Trypsin Inhibitor Heavy chain 3

PI3K

Phosphoinositide 3-kinase

mTOR

Mechanistic Target Of Rapamycin

AKT

Protein kinase B

ERK

Extracellular signal-Regulated Kinase

MAPK

Mitogen-Activated Protein Kinase

CCND2

Cyclin D2

MYH3

Myosin Heavy Chain 3

MYH7

Myosin Heavy Chain 7

MYH8

Myosin Heavy Chain 8

TNNT3

Troponin T type 3

TNNT1

Troponin T type 1

GO

Gene Ontology

MHC

Myosin Heavy Chain

FI

Fusion Index

IGF

1-Insulin-like Growth Factor-1

MDK

Midkine

Rrx_rec

TRACERx recurrence

Rrx_B

TRACERx baseline

Authors' contributions

FGDO, YV, and BB conceived and designed the study. WSZ , AS and MA performed the statistical analyses. XS, ES and NM conducted the in vitro experiments and functional assays. LL, DB, and NL contributed to imaging analyses and radiological data interpretation., CB, YL and KB contributed to proteomic data analysis and biological interpretation. FC, PB, RI, MG, DC, CN, MNC contributed to data collection, clinical annotation, and sample processing. CP and MJH provided critical expertise and contributed to data interpretation. FB, CE, AI, YV, and BB supervised the study. FGDO drafted the manuscript. All authors critically revised the manuscript and approved the final version.

Funding

This work was partially supported by Canceropole Île-de-France (grant number 2024-1-EMERG-06) and by the Fondation Gustave Roussy.

Data availability

The plasma proteomic datasets used in this study are not publicly available. The training cohort (MATCH-R immunotherapy) has been included in a previous publication reporting plasma proteomic analyses (PMID 34416362); however, the underlying dataset is not publicly accessible. The independent validation cohort of metastatic castration resistant prostate cancer (mCRPC) comprises unpublished plasma proteomic data. Due to the presence of potentially identifiable personal information, the data are subject to data protection regulations, including the General Data Protection Regulation (GDPR), and are therefore not publicly available. Access to de-identified data may be granted upon reasonable request to the corresponding authors, subject to institutional approval, data transfer agreements, and compliance with applicable regulatory requirements. The code used for model development and analysis is available from the corresponding authors upon reasonable request under the same conditions. The code used for model development and analysis is available from the corresponding authors upon reasonable request, subject to institutional and data protection regulations.

Declarations

Ethics approval and consent to participate

The MATCH-R study was approved by the Comité de Protection des Personnes (CPP) and by the Agence nationale de sécurité du médicament et des produits de santé (ANSM), and was conducted in accordance with the principles of the Declaration of Helsinki and the Guideline for Good Clinical Practice. Written informed consent was obtained from all patients prior to inclusion in the study.

EudraCT number: 2014-A01147-40; CPP dossier reference: Am7501-4-3183.

Consent for publication

Non applicable as no individual data are included in the publication.

Competing interests

Fabrice Barlesi reports institutional relationships (no personal financial interests) with AbbVie, ACEA, Amgen, AstraZeneca, Bayer, Bristol Myers Squibb, Boehringer Ingelheim, Eisai, Eli Lilly Oncology, F. Hoffmann–La Roche Ltd, Genentech, Ipsen, Ignyta, Innate Pharma, Loxo, Novartis, MedImmune, Merck, MSD, Pierre Fabre, Pfizer, Sanofi-Aventis, Summit Therapeutics and Takeda. Caroline Even reports consulting or advisory roles with Innate Pharma, Bristol Myers Squibb, MSD Oncology, Merck Serono, Novartis, F-star Therapeutics, Merus and GlaxoSmithKline (institutional), and travel, accommodation or expenses from MSD Oncology and Merck Serono. Nathalie Lassau reports participation on an advisory board for Jazz Pharmaceuticals. Yohann Loriot reports honoraria from Janssen, Bristol Myers Squibb, Roche, Gilead, MSD and Pfizer; institutional research funding from Amgen, Janssen Oncology, MSD Oncology, Lilly, AstraZeneca, Orion, Exelixis, Incyte, Pfizer, Sanofi, Astellas Pharma, Gilead Sciences, Merck KGaA, Taiho Pharmaceutical, Bristol Myers Squibb, Roche and Tyra Biosciences; and travel, accommodation or expenses from Astellas Pharma, Pfizer, MSD Oncology and AstraZeneca. Antoine Italiano reports research grants from AstraZeneca, Bayer, Bristol Myers Squibb, Merck, MSD and Pharmamar. Carla M. Prado has previously received honoraria and/or paid consultancy from Abbott Nutrition, Nutricia, Nestlé Health Science and Novo Nordisk. Mariam Jamal-Hanjani reports consulting for Astex Pharmaceuticals, Pfizer and Achilles Therapeutics; membership on the Scientific Advisory Board and Steering Committee of Achilles Therapeutics; and speaker honoraria from Pfizer, Astex Pharmaceuticals, Oslo Cancer Cluster, Bristol Myers Squibb and Genentech. Benjamin Besse reports institutional honoraria and speaker’s bureau participation for AbbVie, AstraZeneca, Chugai Pharmaceutical, Daiichi Sankyo, Hedera Dx, Janssen, Merck Sharp & Dohme, Roche, Sanofi Aventis and Springer Healthcare Ltd; consulting or advisory roles (institutional) for AbbVie, BioNTech SE, Bristol Myers Squibb, Chugai Pharmaceutical, CureVac AG, Daiichi Sankyo, F. Hoffmann–La Roche Ltd, Pharmamar, Regeneron, Sanofi Aventis and Turning Point Therapeutics; and institutional research funding from AstraZeneca, BeiGene, Genmab A/S, GlaxoSmithKline, Janssen, Merck Sharp & Dohme, Ose Immunotherapeutics, Pharmamar, Roche-Genentech, Sanofi and Takeda.All other authors declare that they have no competing interests.