Abstract

Background & Aims

   Currently, preoperative diagnostic methods that can distinguish cancer
   from benign disease of the gallbladder are insufficient, and several
   surgical resections can be avoided if the pathology is known prior to
   surgery. This study aimed to assess whether preoperative plasma
   proteins can distinguish gallbladder cancer (GBC) from cholecystitis,
   with the main goal of identifying proteins for multivariate description
   of the postoperative diagnosis, before surgery.

Methods

   Samples from 82 individuals with suspected GBC who underwent
   bisegmentectomy and lymphadenectomy at Karolinska University Hospital
   between 2009 and 2020 were included in this retrospective,
   observational, single-center study. Preoperative plasma samples were
   analyzed using a 7,500 proteomics panel from SomaScan®.
   High-dimensional statistical methods including machine learning
   regularization, were used to analyze the data.

Results

   In our study, we identified and characterized a panel of 651 proteins
   that exhibited differential expression between GBC and cholecystitis.
   Through multivariate analysis, we demonstrated that circulating
   proteomics data provide valuable insights for diagnosing GBC before
   surgical intervention. Notably, we identified a subset of eight plasma
   proteins (PAHX, CD8A, HRG, CRIS2, Dynactin subunit 2, AT2A3, CSTN2, and
   DEPP) that effectively differentiated GBC from cholecystitis with a
   diagnostic accuracy of 94% when validated on a test set. These findings
   hold potential for clinical validation and could significantly aid in
   preoperative decision-making when GBC is suspected.

Conclusions

   Our findings demonstrate that the preoperative assessment of plasma
   proteins can accurately differentiate cholecystitis from malignancy,
   supporting the potential development of a noninvasive test to assist
   preoperative decision-making when GBC is suspected.

Impact and implications

   This study highlights the potential of plasma proteomic profiling to
   significantly improve the preoperative diagnostic accuracy of
   gallbladder cancer vs. cholecystitis. Using machine learning models, we
   identified biologically relevant plasma proteins associated with the
   diagnosis of gall bladder cancer. A noninvasive preoperative test based
   on selected plasma proteins could potentially enhance clinical
   decision-making, reduce unnecessary surgeries, and mitigate the
   associated risks for patients with suspected GBC, marking a step
   forward in precision medicine.

   Keywords: Cholecystitis, Tumor, Biomarkers, Machine learning

Graphical abstract

   [31]Image 1
   [32]Open in a new tab

Highlights

     * •
       The plasma proteome can differentiate gallbladder cancer from
       cholecystitis with high accuracy.
     * •
       Machine learning models can identify a biologically relevant subset
       of differentiating plasma proteins.
     * •
       Plasma protein tests could have the potential to assist in
       preoperative decision-making when GBC is suspected.

Introduction

   Gallbladder cancer (GBC) and inflammatory disease of the gallbladder
   (cholecystitis) are clinical entities that are difficult to separate
   before histological analysis, especially xanthogranulomatous
   cholecystitis.[33]^1^,[34]^2 Despite advances in modern radiology, to
   date, no diagnostic modality has been sufficiently accurate[35]^3 and
   differential diagnosis is challenging ([36]Fig. 1).

Fig. 1.

   [37]Fig. 1
   [38]Open in a new tab

   Demonstrating the preoperative diagnostic challenge in gallbladder
   cancer (GBC).

   (A) Imaging of a patient with GBC with primary suspicion of
   cholecystitis on preoperative investigation. (B) Imaging of a patient
   with cholecystitis with primary suspicion of malignant disease on
   preoperative investigation.

   Biopsies should be avoided preoperatively because of the risk of tumor
   seeding and spread. Hence, many individuals with radiological suspicion
   of GBC undergo radical cholecystectomy and lymphadenectomy even if the
   diagnosis is often uncertain. Radical cholecystectomy generally
   consists of resection of liver segments 4b and 5 together with
   gallbladder or wedge resection of the adjacent liver parenchyma
   surrounding the gallbladder.[39]^4

   When postoperative histology revealed only inflammation in the
   gallbladder, these patients received an unnecessary invasive treatment,
   with a substantial risk of complications. If accurate methods to
   distinguish between cancer and inflammation of the gallbladder before
   surgery are available, simple cholecystectomy with much lower risks, or
   even no surgery, could have been preferred for these individuals.
   Several attempts have been made to improve preoperative diagnostics in
   this patient group without convincing results.[40][5], [41][6], [42][7]
   To date, clinicians have been faced with the risk of overtreating some
   patients to avoid missing the optimal treatment window for individuals
   with early GBC, where delayed diagnosis leads to poorer prognosis.

   The use of plasma biomarkers for diagnosis is gaining interest.[43][8],
   [44][9], [45][10] In 2014, Wang et al.[46]^8 evaluated the serum levels
   of CA 242, CA 125, CEA, and CA 19-9, concluding that CA 242 and CA 19-9
   might assist in the diagnosis in the early stage and as prognostic
   markers for GBC compared with benign gallbladder disease.[47]^8
   Recently, in 2023, Baichan et al.[48]^9 analyzed tissue and blood using
   liquid chromatography-mass spectrometry and identified nine
   dysregulated proteins (APOA1, APOA2, RET4, TTR, HEMO, HBB, HBA, PIGR,
   and APOE) in both tissue and plasma compared with benign biliary
   pathology. Several studies have evaluated GBC protein biomarkers at
   different tumor stages,[49]^9^,[50]^11 including extracellular
   vesicles[51]^12 and plasma proteins. In 2021, a study identified 29
   proteins from plasma-derived extracellular vesicles that were uniquely
   dysregulated in early stage GBC and clinically verified three proteins
   (NT5E, ANPEP, and MME).[52]^12 Tissue[53]^13^,[54]^14 and
   plasma-derived extracellular vesicles[55]^12 have been identified as
   potential biomarkers for GBC stages. It is noteworthy that, to our
   knowledge, no GBC biomarker discovery study has screened more than
   2,600 proteins.[56]^10 These previous studies relied on statistical
   tests to compare GBC and control groups. The use of such
   frequency-based analyses can lead to high false-positive rates owing to
   multiple testing.[57]^16 They only show associations between the
   disease and the biomarker, and cannot necessarily be translated to
   predictive potential.[58]^17 Bayesian model-based machine learning (ML)
   methods can improve biomarker discovery in large datasets as follows:
   (1) they are fitted to high-dimensional data, particularly when the
   number of variables outweighs observations and (2) they optimize
   biomarker selection based on outcome predictability (i.e. GBC vs.
   cholecystitis).[59][18], [60][19], [61][20]

   In this study, we aimed to investigate whether preoperative assessment
   of plasma proteins in samples can distinguish GBC from cholecystitis
   and to explore whether proteomics could be used in the future to
   predict cancer preoperatively.

Materials and methods

   This retrospective single-center study was approved by the Swedish
   Ethical Review Authority (Dnr 2013/188-31, 2014/2118-32, 2020-02702)
   and was performed in accordance with the Strengthening the Reporting of
   Observational Studies in Epidemiology guidelines.[62]^21

Study population and design

   Adults (≥18 years of age) who underwent bisegmentectomy or wedge
   resection (anatomical resection of liver segments 4b and 5 or part of
   these segments) with lymphadenectomy and frozen sections of the cystic
   duct for suspected GBC at Karolinska University Hospital (2009–2020)
   were screened for inclusion in the study.

   Based on the pathology report after resection, patients were divided
   into two groups: inflammation (cholecystitis) and GBC. Individuals with
   advanced tumors requiring more extensive resection, immunosuppressive
   treatment, surgery for incidental GBC from a previous cholecystectomy,
   or diagnoses other than GBC or cholecystitis were excluded. After
   screening, two groups of similar sizes were obtained.

   Clinical data of patients from the cholecystitis (n = 38) and GBC (n =
   44) groups and their correlation with proteomics data were explored
   using a computational technique called singular value decomposition
   (SVD) analysis. In SVD, principal component analysis (PCA) was first
   performed to reduce data dimensionality, and then the correlation of
   clinical parameters with the principal components was statistically
   tested (see [63]Fig. S1A).

Surgical management

   During the study period, all patients with suspected GBC were discussed
   at a multidisciplinary team conference where the decision to recommend
   surgery and the operating strategy were taken. Individuals with a
   preoperative suspicion of GBC were operated (2009–2020) and divided
   into two groups, cholecystectomy and GBC, based on the histology of the
   operative specimen. All patients had a frozen section from the cystic
   duct, and if positive for cancer or high-grade dysplasia, resection of
   the extrahepatic bile duct was performed with hepaticojejunostomy. All
   patients had at least one regional lymphadenectomy extending to
   stations 8a, 12a, b, c, and p.

Proteomics screening

   All patients who underwent surgery at Karolinska University Hospital
   since 2019 were asked to provide written consent for inclusion in the
   prospective biobank consisting of all hepatobiliary surgeries.
   Preoperative EDTA plasma samples were collected on the day of surgery,
   centrifuged, aliquoted, frozen, and stored at -80 °C. Plasma samples
   acquired from patients preoperatively were analyzed using an
   aptamer-based proteomics platform, SomaScan®[64]^22 (SomaLogic,
   Boulder, CO, USA), using a 7 K assay. This novel technology ensures
   specific binding to different proteins and enables their accurate
   quantification. All raw data are publicly available at
   [65]https://doi.org/10.6084/m9.figshare.26388166. Additional details
   are provided in the [66]Supplementary CTAT Table.

Statistical and computational analyses

   A total of 80 patients was sufficient to reach an acceptable power of
   80% (medium effect size and a significance level of 0.05), with a
   significance level of 0.05. All statistical analyses were performed
   using R version 4.3.1 (2023-06-16, R Foundation for Statistical
   Computing, Vienna, Austria). The data, in the ‘.adat’ format were
   loaded and manipulated using the SomaDataIO package.[67]^7 The quality
   of the data was controlled and then centered and scaled for downstream
   analysis. Proteins from organisms other than humans and duplicate
   isotopes were excluded from analysis.

ML methods

   The elastic net (EN) regularization method, based on generalized linear
   models, was used as a feature (variable) selection method using the
   glmnet v4.1.8 package in R and the caret package v6.0-94. Data were
   split into training and test sets (80% to 20%) with respect to the GBC
   and cholecystitis groups, and we used the leave-one-out
   cross-validation technique. The models were run iteratively with
   different hyperparameters (alpha and lambda) to determine the lambda
   parameter with the highest model performance, best fit, and lowest
   complexity for a given alpha. Setting alpha defines the regularization
   method that is used. The least absolute shrinkage and selection
   operator (LASSO) regression (alpha = 1) defines a stringent cut-off of
   less informative variables, whereas a strictly positive alpha that is
   <1 implements an EN regression where collinear variables are maintained
   in the model ([68]Fig. 2). Model performance was evaluated using the
   root mean square deviation (RMSE), mean square deviation (MSD),
   correlation (R^2), receiver-operating characteristic curve (ROC), and
   area under the ROC curve (AUC) metrics.

Fig. 2.

   [69]Fig. 2
   [70]Open in a new tab

   Flow chart of the study design and methods used.

   GBC, gallbladder cancer; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of
   Genes and Genomes.

Association analysis and clustering

   We calculated the standardized mean differences and the 95% CIs for
   each protein using an independent t test to compare the mean levels
   between the cholecystitis and GBC groups. Dimensionality reduction
   (PCA) was used to visualize the data. The clustering potential of
   plasma proteins identified by ML was tested using unsupervised
   clustering methods (k-means clustering and hierarchical clustering).

Enrichment analysis

   Using all the proteins found by the EN methods and the statistical
   significance test, we performed enrichment analysis, using Kyoto
   Encyclopedia of Genes and Genomes (KEGG) pathway
   ([71]https://www.genome.jp/kegg/pathway.html) and Gene Ontology (GO,
   [72]https://geneontology.org/) databases, clusterProfiler v4.8.3 R
   package, Enrichr,[73]^23 and ShinyGO web-tool.[74]^24 All scripts used
   in computational analysis are available at GitHub
   ([75]https://github.com/MedH-AB-group/Cancer_plasma_proteomics).

Results

   Overall, 82 patients with clinical and radiological suspicion of GBC
   were included in this study. Of these, 44 patients had GBC in the
   pathology report and 38 had cholecystitis. The clinicopathological
   characteristics are shown in [76]Table 1.

Table 1.

   General description of clinical characteristics of the cohort.
   Cholecystitis n = 38 GBC n = 44 p value (Χ^2)
   General description
   Age, median (IQR) 63.5 (18.2) 69 (12.5) 0.036
   Sex, n (%)
    Female 16 (42) 26 (59)
    Male 22 (58) 18 (41)
   BMI, median (IQR) 27 (8) 25.9 (5.7) 0.241
   Diabetes, n (%) 7 (18) 4 (9)
   Immunosuppressive medication, n (%) 0 0
   Primary sclerosing cholangitis (PSC), n (%) 0 7 (16)
   Cirrhosis, n (%) 0 1 (2)
     __________________________________________________________________

   Baseline tests
   CRP, median (IQR) 2 (4.5) 2 (5.8) 0.227
   Albumin, median (IQR) 36 (4) 33 (6) 0.009
   Bilirubin, median (IQR) 6 (4) 6 (6) 0.616
   CA199, median (IQR) 15 (12) 110 (110.5) 0.076
     __________________________________________________________________

   Imaging
   CT, n (%) 34 (89) 39 (88)
   MRI, n (%) 17 (45) 17 (39)
   Both CT and MRI, n (%) 13 (34) 13 (29)
   Preoperative biliary stenting, n (%) 1 (2) 4 (9)
     __________________________________________________________________

   Pathological report
   Inflammation All –
   T stage, n (%) –
    In situ 1 (3)
    1 3 (9)
    2 17 (53)
    3 11 (34)
    4 0
   Lymph node status, n (%) –
    N0 (0 positive nodes) 19 (43)
    N1 (1–3 positive nodes) 11 (25)
    N2 (≥4 positive nodes) 2 (4)
    NA 12 (27)
   M status, n (%) –
    M0 42 (95)
    M1 (aortocaval nodes) 2 (5)
   [77]Open in a new tab

   p values are calculated based on the Χ^2 test. CRP, C-reactive protein;
   GBC, gallbladder cancer; n, number of occurrences; n, number of
   patients.

   Before surgery, radiological investigation with either computed
   tomography (CT) scan (57%), magnetic resonance imaging (MRI) (10%), or
   both (32%) was performed. The use of preoperative radiology was similar
   in patients with confirmed GBC and those with a postoperative
   cholecystitis diagnosis.

   We assessed the relationships between clinical factors and proteomics
   profile using SVD analysis ([78]Fig. S1A). Liver cirrhosis and primary
   sclerosing cholangitis did not show any significant associations,
   whereas diabetes displayed correlation (significant at p <0.05) with
   ∼5% of the proteomic data ([79]Table 1 and [80]Fig. S1A).

Large-scale plasma proteomic signature distinguishes cholecystitis from GBC

   The primary objective was to determine whether circulating proteins
   detected preoperatively in the peripheral blood can accurately
   distinguishing between patient groups (malignancy or cholecystitis). We
   identified 267 proteins in the plasma that were differentially
   expressed between the groups ([81]Fig. 3A), using statistical tests.
   Notably, three of these proteins exhibited the highest standard mean
   difference (SMD >0.73) in this study, indicating prominent group
   differences: the cysteine-rich secretory protein 2 (CRIS2) with a raw
   mean difference (MD) of 381 (95% CI 142–619), the homeobox protein HMX2
   (MD 70.7, 95% CI 26–115), and the sarcoplasmic endoplasmic reticulum
   calcium ATPase 3 (AT2A3) (MD 143, 95% CI 51–235). Given the large
   number of proteins tested (7,500) relative to the modest sample size (n
   = 82), we recognize the importance of addressing the multiple testing
   issue to avoid false positives.

Fig. 3.

   [82]Fig. 3
   [83]Open in a new tab

   A set of 651 proteins were associated with diagnosis.

   (A). Volcano plot of all the proteins (in blue) and those significantly
   (p <0.05) linked to the diagnosis (in pink), in t test. (B) Venn
   diagram of the overlap between results from elastic net (n = 577), and
   t test at p <0.05 (n = 267). (C) Principal component analysis (PCA) of
   the whole dataset (7,500 proteins) with no apparent data structure. (D)
   PCA using the 651 diagnosis-associated proteins showed separation of
   GBC and cholecystitis. (E) Unsupervised hierarchical clustering of
   diagnosis-associated proteins largely distinguished GBC and
   cholecystitis, yet in the central region these groups overlapped (11%
   of the patients). Dim, dimension; GBC, gallbladder cancer; PC,
   principal component.

   Given the high dimensionality of the data, we opted for the
   regularization method EN, an ML method that (1) can handle
   high-dimensional datasets, (2) prevents overfitting by utilizing
   Bayesian statistics, and (3) balances between retaining correlated
   variables and performing variable selection.[84]^25 Unlike traditional
   methods, which may exhibit multicollinearity, EN can identify proteins
   that contribute jointly to the model even if they are highly
   correlated. To explore the multivariate association within the
   proteomic data, we applied an EN-regularized regression. This ML model
   incorporated a subset of 577 proteins associated with postoperative
   diagnosis. The EN model achieved a high performance on the test set,
   with an AUC of 94%. The model’s mean squared error and the root mean
   squared error were low, at 0.01 and 0.12, respectively, while the R^2
   metric was 93%, indicating a strong fit to the data. Collectively,
   these measurements underscored the capacity of this subset of proteins
   to differentiate between the GBC and cholecystitis profiles, supporting
   the robustness of our multivariate approach.

   Notably, 193 proteins (29.6%) identified through EN also appeared in
   the statistical t test results ([85]Fig. 3B), exhibiting a medium to
   large effect SMD ranging between 0.4 and 0.75. This overlap underscores
   the significance of these proteins, even when considering the
   limitations of a small sample size and the burden of multiple testing.

   Combining the results from the frequentist statistical analysis and the
   ML results, we identified diagnosis-associated proteins (651) affected
   by both upregulation (316 proteins) and downregulation (341 proteins).
   In the PCA performed using the whole proteomic dataset (7,500
   proteins), no data pattern was observed ([86]Fig. 3C). However, using
   the proteins identified by the EN and t test, amounting to 651
   proteins, revealed a strong differentiation of GBC and cholecystitis
   ([87]Fig. 3D and E).

ML methods identify a minimal plasma protein signature for preoperative
diagnosis of GBC

   By adjusting the alpha parameter in our ML model and performing LASSO
   regression, we identified a more parsimonious set of 13 proteins
   ([88]Fig. 4) associated with the postoperative diagnosis. This refined
   model achieved an AUC of 98% in the test set, suggesting a strong
   discriminatory capacity for these proteins within the cohort. Although
   not aimed at predictive applications, the high AUC underscores the
   potential of these selected proteins to capture key multivariate
   relationships distinguishing GBC from cholecystitis profiles.
   Interestingly, 12 of these proteins had the highest SMD in this study,
   ranging from 0.64 to 0.74. The exception was dickkopf-related protein 2
   (DKK2), which had a medium SMD of 0.3. This streamlined protein set
   further reinforces the robustness of our multivariate approach and
   offers insight into potentially influential biomarkers.

Fig. 4.

   [89]Fig. 4
   [90]Open in a new tab

   Characteristics of the most descriptive proteins of postoperative
   diagnosis.

   (A) The variance of the protein levels in patients with cholecystitis
   and GBC. (B) Direction and contribution of each protein to the data
   variance, assessed by principal component analysis (PCA). (C) Heatmap
   showing the difference of expression of the proteins between the two
   groups. (D) Data variance and group differences are well described by
   the proteins, visualized by PCA. (E) Low correlation between the
   proteins confirming non-redundant information in the ML model. (F)
   Thirteen proteins could separate the patients into two clusters (k-mean
   unsupervised clustering) that largely correspond to the cholecystitis
   and GBC groups (89% of the patients were rightfully assigned). All nine
   wrongfully assigned patients were in the middle area. PCA, principal
   component analysis.

   Next, we further explored the proteins selected by the ML model in
   relation to the dataset. Individually, no protein showed a distinct
   frequency difference between GBC and cholecystitis ([91]Fig. 4A and C).
   However, PCA demonstrated that the proteins largely accounted for data
   variability and spanned their directions ([92]Fig. 4B and D). K-means
   clustering, using the 13 same proteins, resulted in an unsupervised
   bifurcation of the patients into two clusters (i.e. groups), largely
   aligning with the cholecystitis and GBC groups ([93]Fig. 4F). Eleven
   percent of the patients were misclassified and all fell within the
   intermediate area. K-means clustering optimizes grouping by minimizing
   intragroup distances and maximizing intergroup distances. However, for
   data points in the intermediate area, patients were not accurately
   assigned to their respective groups. In fact, upon iteration with
   different cluster quantities (changing the K value), the optimal
   cluster count was three, where patients in the middle constituted a
   distinct group.

Pathway enrichment analysis reveals biological mechanisms underlining
cholecystitis and GBC

   To gain insight into the molecular functions and biological pathways
   overrepresented in these 651 diagnosis-associated proteins, we
   performed an enrichment analysis. Cytokine receptor binding as well as
   activity proteins and the pro-inflammatory cytokine pathway (TNF
   signaling pathway) were significantly (adjusted p <0.01) enriched both
   in the proteins’ molecular functions and the biological pathways in
   which they were involved ([94]Fig. 5A and B), as expected in the
   presence of inflammation. The set was also enriched (adjusted p <0.01)
   in proteins related to growth development (serine activity and growth
   factor binding), membrane receptors (transmembrane receptors and G
   protein binding), and glycosaminoglycan-binding functions. Based on
   KEGG brites ([95]Fig. 5C), several diagnosis-associated proteins
   belonged to membrane trafficking, chromosome-associated proteins, and
   transcription factors. Notably, four transcription factor families were
   significantly enriched in diagnosis-associated proteins (p <0.01)
   ([96]Fig. 5D and [97]Table S2). Finally, we connected HGF, CDK2, FADD,
   and other proteins to the cancer pathways in the KEGG database
   ([98]Fig. 5E).

Fig. 5.

   [99]Fig. 5
   [100]Open in a new tab

   Enrichment analysis of the proteins linked to diagnosis.

   (A) Enriched KEGG pathways in the diagnosis-associated proteins (651)
   where cytokine receptors are the most significant. (B) Gene Ontology
   analysis of the molecular functions, the cellular components, and the
   biological processes of the diagnosis-associated proteins. (C) Enriched
   KEGG brites in diagnosis-associated proteins. (D) Transcription factor
   families enriched in the protein list. (E) Cancer pathways in KEGG
   database with dysregulated proteins (from this study) highlighted in
   red. FDR, false discovery rate; KEGG, Kyoto Encyclopedia of Genes and
   Genomes. N. of genes, number of genes. p values are calculated based on
   Fisher's exact test.

Discussion

   In this retrospective exploratory study, we investigated the potential
   of the plasma proteome to preoperatively differentiate between GBC and
   cholecystitis. We aimed to identify a minimum set of proteins that
   could reflect the postoperative diagnosis before surgery. Plasma
   samples from 82 individuals who underwent liver resection were
   analyzed. Using ML techniques, protein subsets could differentiate
   patients with a postoperative histological diagnosis of GBC from
   patients with cholecystitis with 98% accuracy. Indeed, reliable
   clustering is not feasible for a minor subset of patients. This study,
   although exploratory and lacking external validation, used a reasonable
   patient pool, screened a large number of proteins (7,500 proteins), and
   used cutting-edge and suitable computational techniques. Our findings
   suggest that the preoperative measurement of plasma proteins could be
   informative regarding the presence of GBC. Preoperative blood tests
   could potentially be developed to enhance diagnostic accuracy,
   particularly in individuals with pre-existing inflammatory conditions.

   The SomaScan® technology from SomaLogic is a pioneering high-throughput
   proteomics technology that allows data-driven biomarker discovery.
   Modified aptamers are called SOMAmers and are small DNA sequences with
   high-affinity target-specific tertiary structures. Their slow off-rate
   binding kinetics were measured using microarray-based fluorescence
   intensity readout.[101]^22 This affinity-based approach overcomes the
   challenge of detecting proteins at vastly different concentrations,
   known as the dynamic range issue, and outperforms mass
   spectrometry.[102]^26 Our choice of ML regularization method was
   motivated by the high dimensionality of the data. EN and LASSO are
   robust variable (proteins in this case) selection methods that
   integrate ML techniques, such as cross-validation and training, to
   optimize the model for outcome predictability (diagnosis in this case).
   In addition, approximately one-third of the proteins identified by EN
   overlapped with the statistical t test results, reinforcing the
   relevance of these proteins despite the challenges posed by the limited
   sample size and multiple testing. This overlap suggests that although
   individual statistical significance may not have survived strict
   multiple testing correction, the proteins identified by both methods
   likely reflect biologically meaningful associations with the disease.
   Together, these analyses highlight the importance of considering
   complementary approaches, such as EN in large proteomic studies where
   traditional statistical methods may be overly conservative.

   Among the identified diagnosis-associated proteins, several hold
   promise for clinical application. For example, CD8 chain (T-cell
   surface glycoprotein CD8 alpha chain [CD8A]), HMX2, histidine-rich
   glycoprotein (HRG), ATPase Sarcoplasmic/Endoplasmic Reticulum Ca2+
   Transporting 3 (AT2A3), and the decidual protein induced by
   progesterone (DEPP)[103]^27 can be tested using commercially available
   ELISA or Western blot kits. In addition, some of these proteins are
   associated with oncogenesis and progression of GBC. Arylamine
   N-acetyltransferase 1 (ARY1) is a key enzyme in various cancers,
   including GBC.[104]^28^,[105]^29 It is upregulated in several cancer
   types and can be a target for cancer therapy and patient
   stratification.[106]^30 Vascular endothelial growth factor receptor 2
   (VEGFR2) is a well-known target in cancer therapy because of its role
   in tumor induced angiogenesis.[107]^31^,[108]^32 Histidine-rich
   glycoprotein (HRG) has been detected in plasma and has been suggested
   as a potential prognostic and diagnostic biomarker in
   cancer.[109]^33^,[110]^34 Finally, CD8A, a component of the CD8
   co-receptor complex, has been linked to GBC subtypes and
   prognosis.[111]^35 In contrast, diagnosis-associated proteins include
   several transcription factors, including HNF4A and PRDM1, which were
   previously associated with cholangiocarcinoma.[112]^36^,[113]^37 The
   NFAT2 family, the most significantly enriched (p = 0.0001) family in
   our results, has been linked to the progression of bladder urothelial
   carcinoma[114]^38 and various human solid tumors and hematologic
   malignancies.[115]^39 Our results highlight the potential relevance of
   these proteins to GBC pathobiology.

   However, knowledge regarding the etiology of GBC is limited. A history
   of gallstones and cholecystitis is the strongest known risk
   factor.[116]^40 Other factors include advanced age, female sex,[117]^41
   chronic bacterial infections (Salmonella and Helicobacter),[118]^40 and
   primary sclerosing cholangitis.[119]^42 The etiology and related
   biochemical alterations in blood during GBC development remain largely
   unknown. Here, we describe a set of proteins that may shed light on the
   pathogenesis of GBC. Several proteins identified in our analysis are
   implicated in the pathophysiology of GBC and other biliary tract
   cancers, suggesting few pathways that are critical to tumor progression
   and immune evasion. For example, phytanoyl-CoA dioxygenase (PAHX) is
   involved in peroxisomal metabolism,[120]^43 and its alterations may
   disrupt lipid metabolic pathways, which have been increasingly linked
   to tumorigenesis in biliary tract cancers.[121]^44 The DEPP protein,
   known for its role in oxidative stress responses,[122]^45 may
   contribute to the heightened oxidative environment that facilitates
   cancer progression.[123]^46 CRISP2 is a member of the cysteine-rich
   secretory protein family and has been linked to immunomodulation and
   pathophysiology of chronic pancreatitis.[124]^47 This protein was
   likely linked to the diagnosis of cholecystitis in our study. VEGF sR2,
   a soluble form of VEGF receptor, is crucial for vascular endothelial
   growth in biliary pathophysiology and angiogenesis regulation. Its
   dysregulation is a known factor in biliary tract cancers, where
   angiogenesis supports tumor growth and metastasis.[125]^48^,[126]^49
   HRG and DKK2 are involved in immune modulation and Wnt
   signaling,[127]^50 respectively. Both pathways play essential roles in
   cancer progression and anticancer immune response.[128]^51^,[129]^52
   Additionally, CD8A, a protective gene with the highest correlation with
   T cells, is a promising prognostic biomarker in bladder cancer,[130]^53
   calsyntenin-2 (CSTN2) has been identified as a potential biomarker for
   colorectal cancer,[131]^54 and dynactin subunit 2 has been identified
   as an oncogene in hepatocellular carcinoma.[132]^55

   Notably, we found a significant association between serpin family A
   (SERPINA1), Annexin A protein 3 (ANXA3), collagen type VI alpha 1 chain
   (COL6A1), proteinase 3 (PRTN3), and keratin (KRT18) with GBC diagnosis,
   through statistical tests and ML techniques. These proteins have been
   previously associated with early stage GBC.[133]^13^,[134]^56
   Additionally, KRT1, ANXA2, peptidyl-prolyl cis-trans isomerase B
   (PPIB), colony stimulating factor 1 (CSF1), and receptor
   tyrosine-protein kinase (ERBB3) were identified using ML methods in
   this study. The first three proteins were recently detected in tissue
   samples and were associated with lymph node metastasis in GBC,[135]^14
   whereas ERBB3 has been described as a targeted therapy for GBC.[136]^59
   Recently, CSF1 was described as a potential preoperative plasma marker
   for biliary tract cancer.[137]^60

   Collectively, these proteins provide insights into the multifaceted
   biological mechanisms driving the development of GBC and other bile
   duct cancers and may serve as promising future plasma biomarkers for
   preoperative GBC diagnosis. Our findings complement those of previous
   studies, in which plasma biomarkers have been reported for the
   diagnosis and prognosis of GBC.[138]^8 In a study by Wang et
   al.[139]^8, a combination of plasma markers (CA125, CA242, and CA199)
   enhanced the diagnostic sensitivity of GBC compared with individual
   markers. Notably, our extensive protein panel did not include CA19-9 or
   CA242, and we found a low SMD for CA125 in our cohort. Furthermore,
   carcinoembryonic antigen (CEA) exhibited an SMD of 0.2, suggesting a
   limited role as a diagnostic marker in our analysis. This highlights
   the need for further investigation of alternative biomarkers that may
   provide improved diagnostic capabilities for GBC.

   The molecular functions and biological pathways identified in our study
   are related to cancer factors and infections. Signal transduction and
   signaling molecules and interactions, such as cytokine receptor-related
   pathways and the TNF signaling pathway, were significantly enriched.
   These pathways are characteristic of inflammation as seen in
   cholecystitis.[140]^61 The MAPK, JAK-STAT, and PI3K/AKR pathways, which
   were significantly dysregulated in our study, are associated with
   inflammation, cholecystitis, and tumor progression.[141]^62^,[142]^63
   ErbB and the coagulation and complement cascade pathways were
   associated with thromboinflammatory process regulation in tumor
   progression[143]^2 and were enriched in downregulated proteins in our
   analysis. The axon guidance pathway, which is altered in pancreatic
   ductal adenocarcinoma,[144]^1 was enriched in this study. Lastly, the
   Wnt signaling pathway was significantly enriched with upregulated
   proteins, including DDK2, which was highly associated with GBC
   diagnosis in our study. This pathway promotes colon
   carcinogenesis.[145]^64

   This study had several limitations. Because of the rarity of GBC, a
   relatively small number of samples were studied. However, this was
   considered when choosing analysis methods that were adapted to such
   conditions. Nevertheless, the results are exploratory and further
   validation of the present results is required. A validation step is
   feasible and cost-effective because of the low number of proteins,
   their availability for clinical testing, and their detection at the
   plasma level. Importantly, validation efforts would require a larger
   collection of samples and, potentially, multicenter and international
   collaboration.

   In conclusion, to our knowledge, our study provides the largest
   proteomic characterization of GBC that can serve as a resource for
   further biomarker discovery. We show that preoperative assessment of
   plasma proteins can accurately differentiate cholecystitis from
   malignancy, supporting the development of a noninvasive test to assist
   preoperative decision-making when GBC is suspected.

Abbreviations

   ANXA, annexin A; ARY1, arylamine N-acetyltransferase 1; AT2A3,
   sarcoplasmic/endoplasmic reticulum calcium ATPase 3; AUC, area under
   the ROC curve; CD8A, T-cell surface glycoprotein CD8 alpha chain; CEA,
   carcinoembryonic antigen; COL6A1, collagen type VI alpha 1 chain;
   CRIS2, cysteine-rich secretory protein 2; CSF1, colony stimulating
   factor 1; CSTN2, calsyntenin-2; CT, computed tomography; DEPP, decidual
   protein induced by progesterone; DKK2, dickkopf-related protein 2;
   dynactin sub2, dynactin subunit 2; EGRF, epidermal growth factor
   receptor; EN, elastic net; ERBrbB3, receptor tyrosine-protein kinase
   erbB-3; FDR, false discovery rate; GBC, gallbladder cancer; GO, Gene
   Ontology; HMX2, homeobox protein HMX2; HRG, histidine-rich
   glycoprotein; Is, in situ; ITI H1, inter-alpha-trypsin inhibitor heavy
   chain H1; KEGG, Kyoto Encyclopedia of Genes and Genomes; KRT, keratin
   protein; LASSO, least absolute shrinkage and selection operator; MD,
   mean difference; ML, machine learning; MRI, magnetic resonance imaging;
   MSD, mean square deviation; PAHX, phytanoyl-CoA dioxygenase
   peroxisomal; PC, principal component; PCA, principal component
   analysis; PPIB, peptidyl-prolyl cis-trans isomerase B; PRTN,
   proteinase; RMSE, root mean square deviation; ROC, receiver operating
   characteristic curve; SERPINA, serpin family A; SMD, standard mean
   difference; SVD, singular value decomposition; TNF, tumor necrosis
   factor; VEGF sR2, soluble vascular endothelial growth factor receptor
   2.

Financial support

   This study was supported by grants from the Swedish Society for Medical
   Research (SSMF), the Center for Innovative Medicine at Karolinska
   Institutet and RegionStockholm. The funding sources were not involved
   in the design or conduct of the research, nor the analysis and
   interpretation of the data, or the writing of this manuscript and
   decision to submit the article for publication to this journal.

Authors’ contributions

   Contributed to the study concept and design: MC, AB, ES. Contributed to
   data acquisition: MC, AA, HJ. Performed the bioinformatics and
   computational analyses: GN. Interpreted data and drafted the
   manuscript: GN, MC, AB, ES. Critically revised and approved the final
   version of the manuscript: all authors.

Data availability statement

   Raw data is available on
   [146]https://doi.org/10.6084/m9.figshare.26388166. Public access to the
   data is restricted by the Swedish laws and regulations that prohibit
   the release of individual-level datasets that could potentially allow a
   personal identification. In this study, this particularly concerns some
   of the clinical metadata and patients’ s characteristics. However, data
   access can be granted in the framework of a defined academic research
   collaboration with requirement of data transfers agreement. Anyone
   wishing to gain access to the data should contact Martin Cornillet and
   Ernesto Sparrelid ([147]martin.cornillet.jeannin@ki.se,
   [148]ernesto.sparrelid@ki.se).

Conflicts of interest

   The authors declare no conflicts of interest that pertain to this work.

   Please refer to the accompanying ICMJE disclosure forms for further
   details.

Acknowledgements