Positive surgical margins (PSMs) are a serious concern in kidney cancer surgery. When cancer cells remain at the cut edge of tissue after a partial or radical nephrectomy, it signals incomplete tumor removal, which is linked to higher rates of local relapse and reduced overall survival.
The standard tool for checking surgical margins intraoperatively is frozen section analysis (IFS), but it has notable drawbacks. It can take 20 to 60 minutes to complete, forcing surgeries to pause while patients remain under anesthesia, and its accuracy is limited by artifacts that arise when tissue is rapidly frozen for examination.
There is a pressing clinical need for a faster and more accurate method to assess whether cancer cells are present at the surgical margin in real time, without the delays and error rates associated with current frozen section techniques.
Desorption electrospray ionization mass spectrometry imaging (DESI-MSI) is an ambient ionization technique that maps the chemical composition of a tissue surface without requiring complex sample preparation. It works at room temperature and in open air, making it well-suited for clinical environments.
The technique generates spatial maps of hundreds to thousands of small metabolites and lipids across a tissue slice, capturing molecules ranging from simple metabolites like lactate and succinate to complex glycerophospholipids. This chemical fingerprint can reflect the biological state of the underlying tissue.
Because DESI-MSI is nondestructive and compatible with subsequent standard pathology staining, the same tissue section used for mass spectrometry can also undergo hematoxylin and eosin (H&E) staining, preserving clinical workflow without sacrificing tissue.
The researchers collected 112 banked frozen tissue samples from 44 normal kidney regions and 68 renal cell carcinoma specimens spanning three subtypes: clear cell RCC (ccRCC), papillary RCC (pRCC), and chromophobe RCC (chRCC). Each tissue was confirmed by an expert genitourinary pathologist.
DESI-MSI data was processed using MassExplorer, a computational pipeline that converts raw mass spectra from individual pixels into a training matrix. The training matrix had over 52,000 rows (pixels) and 27,523 columns (detected molecular peaks), providing an enormously rich dataset for model building.
A multinomial lasso classifier was trained on this pixel-level data. Lasso regression selects a sparse subset of informative molecular features, which it uses to assign each pixel a probability of belonging to one of four categories: normal, ccRCC, pRCC, or chRCC. Tissue-level predictions were made using a majority rule across all pixels in a sample.
The multinomial lasso model selected 281 molecular features from over 27,000 detected species. In cross-validation on the Stanford training set, it achieved 84.3% accuracy with 86.8% sensitivity and 82.6% specificity for distinguishing all RCC subtypes from normal tissue.
On an independent Stanford test set of 48 tissues (20 normal, 28 RCC), the model achieved 85.4% accuracy with 100% specificity (no normal tissues misclassified as cancer). On an external Baylor-UT Austin test set of 57 tissues from a completely different institution, accuracy rose to 91.2% with 95.1% sensitivity.
The area under the ROC curve (AUC) was 0.807 on the training cross-validation set, 0.865 on the Stanford test set, and 0.854 on the Baylor-UT Austin test set, confirming that model performance was stable across diverse patient populations and institutions.
A key concern for any machine learning model is whether it learns stable, biologically meaningful patterns rather than dataset-specific noise. To test this, the researchers examined whether features elevated in cancer or normal tissue in the training set maintained that same trend in the independent test sets.
They found that 74.6% of peaks upregulated in cancer training tissues were also upregulated in cancer tissues in the Stanford test set, and 54.3% held the same pattern in the Baylor-UT Austin test set. This level of consistency across institutions suggests the model captures real biological signals rather than technical artifacts.
For example, arachidonic acid (m/z 303.23) was consistently elevated in normal kidney tissue compared to ccRCC and pRCC tissues across all three datasets, with fold differences of 1.6, 3.3, and 2.3, respectively, confirming reproducibility of a key biomarker signal.
Suppression of arachidonic acid metabolism emerged as a conserved molecular hallmark shared between ccRCC and pRCC tissues across all datasets. Arachidonic acid and its downstream eicosanoid pathway products, including thromboxane derivatives, were consistently lower in these cancer subtypes compared to normal kidney tissue.
This finding has broader biological significance: arachidonic acid is a key precursor to inflammatory signaling molecules. Its suppression in RCC suggests a consistent disruption of lipid-based inflammatory signaling pathways during tumor progression, which may affect how these tumors interact with the immune system.
Importantly, these arachidonic acid patterns were detected across different RCC subtypes, even though the subtypes otherwise differ substantially in their molecular profiles and clinical behavior, suggesting this metabolic change represents a fundamental aspect of RCC biology across histological categories.
The clinical promise of DESI-MSI lies in its potential to replace or supplement frozen section analysis for real-time surgical margin evaluation. The technique requires minimal sample preparation, works at room temperature, and generates results rapidly, making it operationally compatible with intraoperative use.
Prior work by the same group in pancreatic cancer demonstrated that DESI-MSI detected positive margins in 8 of 32 patients whose IFS results were negative. Those 8 patients had a median survival of only 10 months, similar to known PSM outcomes, while patients confirmed negative by both methods had 26-month median survival. This suggests DESI-MSI may be more sensitive than IFS in detecting microscopic residual disease.
The next step for kidney cancer would be prospective testing of DESI-MSI directly against IFS on fresh nephrectomy specimens. If the performance holds, DESI-MSI could potentially reduce reoperation rates, shorten surgery time, and improve patient outcomes by giving surgeons more accurate real-time margin information.
Beyond surgical applications, DESI-MSI provides a window into the metabolic landscape of RCC. The technique identified consistent lipid metabolism alterations across histologically distinct subtypes, offering a complementary perspective to genomic studies that have traditionally dominated RCC biology research.
The observation that different RCC subtypes share certain metabolic signatures while differing in others could help define clinically relevant subgroups with distinct metabolic phenotypes. Such subgroups might respond differently to immunotherapy, and the metabolic signatures might serve as novel predictive biomarkers for treatment selection.
Future integration of DESI-MSI metabolic data with genomic and proteomic datasets could deepen understanding of how the metabolic rewiring observed in RCC drives tumor progression and shapes the immune microenvironment, potentially revealing new therapeutic targets.