RNA profiling of blood platelets noninvasively differentiates colorectal cancer from healthy donors and noncancerous intestinal diseases: a retrospective cohort study

The RNA profiles of tumor-educated platelets (TEPs) possess pathological features that could be used for early cancer detection. However, the utility of TEP RNA profiling in detecting early colorectal cancer (CRC) versus noncancerous colorectal diseases has not yet been investigated. This study assesses the diagnostic capacity of TEP RNA profiles in a cohort of patients with CRC and noncancerous diseases. Transcriptome sequencing for platelets isolated from 132 patients with CRC at early and late stages and 190 controls consisting of healthy donors and patients with ulcerative disease, Crohn’s disease, polyps, and adenomas was performed and analyzed using binary particle swarm optimization coupled with support vector machine to identify genes that contributed to the classification of CRC patients versus controls. The area under the receiver operating curves (AUROCs) and the accuracy of TEP RNA profiles in CRC diagnosis were assessed. TEP RNA profiling achieved high performance in distinguishing and staging CRC patients from the controls. Using the swarm intelligence algorithm, the 921 most contributive genes that classified CRC patients from the controls were identified. AUROCs of 0.928 for the training set via leave-one-out cross-validation and 0.92 for the validation set were achieved, both of which were significantly higher than the clinically utilized serum biomarkers: carcinoembryonic antigen and cancer antigen 19-9. Notably, an AUROC of 0.915 in an external validation set was achieved. For predicting different CRC stages, an AUROC of 0.984 was achieved in the training set and 1.000 in the internal validation set. RNA profiles of TEPs are of potential diagnostic value for identifying early CRC from noncancerous diseases. Prospective studies are needed to validate its clinical relevance.

surgically resectable stages without distant metastasis can indeed improve both survival outcomes and life quality of patients. Given that the development of CRC malignancy often takes two to three decades before clinical manifestation [4], a wide time window for detecting CRC before metastasis is provided. Therefore, it is critical to develop new methods and technology for the diagnosis of CRC at early stages, and particularly for patients who suffer from inflammatory bowel diseases (IBD, such as Crohn's disease and ulcerative colitis), polyps, or adenomas, because these patients have a high risk of developing CRC [5][6][7]. Conventional serum protein biomarkers, such as carcinoembryonic antigen (CEA) and cancer antigen 19-9 (CA199), have low sensitivity and specificity. The sensitivity of CEA and CA199 to detect CRC was reportedly 46.6% and 14.4%, respectively, while the specificity was 80% and 89%, respectively [8]. Precancerous lesions, such as advanced adenoma and highrisk adenoma, also induce aberrant expression of CEA and CA199, which complicates early CRC detection [9]. Besides, some molecular biological markers of CRC, such as serum microRNAs and methylated septin-9 DNA, were also reported, but they did not distinguish well CRC from polyps or adenomas [10,11]. Thus, a new accurate and efficient method is needed for screening CRC from patients with IBD, polyps, and adenomas.
Blood platelets, small anucleate cells originated from bone marrow megakaryocytes, play important roles in response to tumor progression. Apart from genetic materials, proteins, and RNAs derived from megakaryocytes, platelets also actively sequester proteins and spliced/ unspliced RNAs, as well as oncogenic and angiogenic factors from cancer cells, such as VEGF, PDGF, and TGF-β [12]. Blood platelets reportedly contribute to creating tumor microenvironment supporting cancer development and progression [13]. The interplays between tumor cells and blood platelets impact tumor growth and dissemination [14][15][16]. Tumor cells can regulate blood platelets by transferring tumor-associated factors into platelets, consequently altering the expression profiles of blood platelets [17]. Meanwhile, platelet activation can release growth factors to facilitate tumor cell survival [18]. Platelets can encompass circulating tumor cells (CTCs) in the bloodstream, helping CTCs escape from immune cell-induced apoptosis and fluidic shearing force. Moreover, in addition to stimulating cancer cell adhesion and extravasation, platelets also contribute to CTC transmigration through the blood vessel walls and distant metastatic lesion formation [19].
RNA profiling of TEPs has emerged as a new liquid biopsy-based cancer detection method, allowing for noninvasive cancer detection [12]. The advantage of using RNA profiles of TEPs as a new strategy for early cancer detection is that platelet is the second most abundant cell types in the blood and can be stored at room temperature up to 48 h [20]. RNA profiles of TEPs were reported to achieve an AUROC of 0.99 and an accuracy of 96% in differentiating healthy donors and multiple cancer types at advanced or metastatic stages [20]. Of note, primary tumor sites and oncogenic alterations at the DNA level could be pinpointed using TEP RNA profiles [20]. The performance of using RNA profiles of TEPs to differentiate nonsmall cell lung cancer (NSCLC) from other noncancerous diseases was robust, but the control diseases were not closely related to NSCLC, and the classification accuracy was not as good as its performance in distinguishing healthy donors from cancer patients at advanced stages [21]. The accuracy of TEP-based detection of nonsmall cell lung cancer was reportedly 81% for early stages and 88% for late stages [21]. This performance is independent of age, smoking habits, and inflammatory states [21]. However, the analysis on RNA profiles of TEPs towards detecting early CRC, especially in the context of noncancerous intestinal diseases that were associated with the development of CRC, has not yet been explored. Tumorigenesis of sporadic CRC follows canonical multistep development, starting from polyps and adenomas to carcinoma and involving diverse genomic and epigenomic alterations [4], which collectively complicate early CRC detection. This study was set to assess the diagnostic performance of TEP RNA profiles in detecting CRC at an early stage from a cohort of 322 donors by expanding controls to cover a wide range of noncancerous diseases, such as IBDs, polyps, and adenomas.

Blood processing and platelet isolation
This study was conducted according to the Helsinki human subject doctrine and was approved by the Huazhong University of Science and Technology Review Board and Ethics Committee. Written consent to participate was acquired from all patients. A total of 322 blood samples were obtained from healthy donors (n = 21), patients with Crohn's disease (CD, n = 40), ulcerative colitis (UC, n = 22), polyps (n = 48), and adenoma (Ad, n = 59) or CRC (n = 132) in Wuhan Union Hospital. The number of CRC patients at stages I, II, III, and IV was 25, 48, 58, and 1, respectively. Blood samples were stored in 5-mL purple-capped vacutainers equipped with the anticoagulant EDTAK 2 (purchased from Zhiyuan Medical Technology Co., Ltd.). Platelets were isolated using gradient centrifugation according to the standard experimental method described previously [22]. To evaluate platelet purity, morphological analysis was implemented to check freshly isolated and randomly selected platelet samples.
Samples with 0-5 nucleated cells per 10 million platelets were included in the follow-up processes. Isolated platelets were lysed with 1 mL RNAiso (takara NO9109), followed by pipetting RNAiso solution to complete lysis. Platelet total RNA was purified with Direct-zol RNA Miniprep (ZYMO RESEARCH R2052), then 500 pg of total RNA was subjected to SMARTer mRNA amplification and sequencing. During blood acquisition for library preparation, 134 (29.4%) samples were excluded due to low blood volume (< 2 mL, six samples), nucleated cell contamination (60 samples), or poor RNA quality (total RNA < 5 ng and/or RIN value < 6, 68 samples).

Transcriptome sequencing
The quality of total RNA was examined using an Agilent 2100 bioanalyzer and then was subjected to cDNA synthesis and amplification using SMARTer kit (Clontech Laboratories, Inc.) according to the manufacturers' protocol. By using Agilent 2100 bioanalyzer with DNA high-sensitivity chip, we performed quality control to amplified sequencing. Samples were prepared using the Ovation ® SoLo RNA-seq Systems (HUMAN PART NO0500) according to the manufacturer's protocol. Finally, we pooled high-quality samples with product sizes ranging 300-500 bp in equimolar concentrations, then submitted for 100 bp paired-read sequencing on the Illumina HiSeq X-ten.

Sequencing data analysis
We used STAR to perform spliced alignment of clean reads to human reference genome hg19 guided by annotated transcripts of Ensembl version 75 [23]. The intron-spanning read count table of each transcript was collected by STAR during sequence alignment. The read count table obtained from the STAR aligner was equal to those obtained from the HTSeq tool [24]. Both coding and noncoding RNAs were included in abundance estimation and downstream analyses.

Differential gene expression analysis and data normalization
Differential gene expression analysis was performed using the R-package DESeq2 [25]. Transcripts with less than five read counts in all samples were excluded. We constructed DESeq object from the read count table obtained above and performed variance stabilizing transformation with the vst function [26]. Construction of the DESeq object included estimation of size factors, estimation of dispersion, and fitting negative binomial general linear model to the data. The differential expression analysis statistics including log2 fold-change, test statistics, and p-value were extracted from the DESeq object. We employed the R routine results to extract the result table from the DESeq analysis. We used the test statistics to perform gene set enrichment analysis for cancer hallmark pathways and platelet signatures downloaded from MsigDB [27]. We obtained normalized gene expression data after variance stabilizing transformation. We calculated the correlation of the expression level of each gene with the age of individuals. Genes that exhibited significant correlation with age were excluded from the downstream analysis to avoid the impact induced by age. We then performed surrogate variable analysis to remove unwanted variations within the data via the R-package sva [28].

Feature selection via binary particle swarm optimization
Particle swarm optimization (PSO) mimics natural phenomena such as movements of bird flocks. The optimization procedure was first initiated with multiple particles. In the feature selection setting, particles are subsets of different predictors. Each particle has its position, velocity, and fitness value in the searching space. In our study, we used support vector machine (SVM) as the classifier and area under the receiver operating characteristic curve (AUROC) as the fitness value of the classification model. The fitness of the model was iteratively evaluated on the last position and current velocity, and the best position was determined. The PSO algorithm was firstly proposed for real-value optimization and later adapted to discrete optimization [29]. Let x id and v id denote the coordinates and velocity of the i th particle in D-dimensional space and g as the index of the best particle in the neighborhoods identified so far [28]. The movement of the i th particle is as follows: where p id and x id are binary values, i.e., 0 or 1, and is the random positive number generated for particle i th .
The velocity vid was transformed by logistic function S(v id ) as follows: The function rand is a uniform random number generator in [0.0, 1.0], and the range of S(vid) is [0.0, 1.0]. We ran the algorithm for 100 iterations. The feature sets with the highest AUROC were used to build the final SVM classifier and subsequently evaluated its performance on internal and external validation sets. The R package caret (version 6.0-78) was used to build and optimize the parameters of the SVM classifier. The calculation of AUROC and the visualization of ROC were performed with the R package pROC (version 1.10.0). We used the if (rand() < S(v id )) then x id = 1, else x id = 0 R function multiclass.roc in the pROC package to calculate the multiclass AUROC, which implemented the multiclass AUROC calculation proposed by Hand and Till [30]. Tenfold cross-validation was used as an optimization cohort.
The median sequencing reads of 58.8 million per sample were obtained; RNA sequence alignment was performed by STAR [21], and the median percentage of uniquely mapped reads was 83.2% per sample (for detailed information regarding sequencing data and mapping results, see Additional file 1: Table S2). Intron-spanning reads of each gene were collected during sequence alignment (see the "Methods" section). After exclusion of genes with low coverage, 16,300 genes were finally obtained for the following analyses. We performed differential gene expression analysis and subsequently gene set enrichment analysis (GSEA). In total, 863 genes exhibited significant differences in CRC patients versus all controls (adjusted p < 0.1): 161 upregulated and 702 downregulated genes; 1095 genes exhibited differential expression across CRC patients, healthy donors, and patients with noncancerous diseases. The GSEA analysis showed that immune-related signaling circuits were significantly downregulated in CRC patients, whereas the circuits of platelet signatures and platelet activation were significantly upregulated in CRC patients. This finding was consistent with a previous study by Best and colleagues [20]. The downregulated circuits included TNF-α signaling via NF-κB, interferon α/γ responses, and IL2 signaling, while the upregulated circuits included myogenesis, heme metabolism, platelet signature, responses to elevated platelet cytosolic CA2, and platelet activation and aggregation (Fig. 3). On the contrary to CRC, those immune-related signaling pathways (TNF-α signaling via NF-κB, interferon α/γ responses, and allograft rejection) were significantly upregulated in both polyp or adenoma patients (Additional file 1: Fig. S3). The heatmap representation of differentially expressed genes (Fig. 4) suggests that RNA profiles of TEPs in healthy donors are clearly separable from patients diagnosed with CRC, polyps, adenoma, Crohn's disease, and ulcerative colitis in our cohort (Fisher's exact test, all p < 0.001) and in an external cohort from Best and colleagues (Fisher's exact test, p < 0.001) [20].
Binary particle swarm optimization (PSO) coupled with support vector machine (SVM) was used to identify a panel of genes that contributed the most to the classification of CRC patients versus controls. We excluded the genes significantly associated with age of individuals to avoid unwanted impact on classification. Variance stabilizing transformation of gene expression matrix (after exclusion of age-associated genes) was used as the inputs for binary PSO-based feature selection. Eventually, 921 genes were identified as the most contributive genes and used to build the cancer-versus-control classifier (Additional file 1: Table S3). The area under the receiver operating curves (AUROCs) for the repeated cross-validation of the training set and the internal validation set were evaluated iteratively (Additional file 1: Fig. S4), which indicates that SVM fits well to both the training and the validation sets. We achieved an AUROC of 0.928 (95% CI 0.891-0.965) on the training set (Fig. 5A) as measured by leave-one-out cross-validation and 0.92 (95% CI 0.869-0.971) on the internal validation set (Fig. 5B). Additionally, an AUROC of 0.915 (95% CI 0.859-0.970) was obtained for the external dataset reported by Best and colleagues (Fig. 5C) [20]. The AUROC values in the training set and the internal validation set were 0.785 (95% CI 0.708-0.863) and 0.679 (95% CI 0.562-0.796) for CEA, respectively, and 0.676 (95% CI 0.581-0.771) and 0.546 (95% CI 0.416-0.482) for CA199, respectively. The identified panel markers achieved significantly higher AUROC in both the training set and the internal validation set than did CEA (p = 0.001 and p = 0.0003, Fig. 3 Gene set enrichment analyses of differentially expressed genes in the pathways of cancer hallmarks and platelet signatures respectively) and CA199 (all p < 0.0001). For the training set, the classification accuracy, sensitivity, and specificity were 87.6%, 97.5%, and 81.1%, respectively; for the internal validation set, the classification accuracy, sensitivity, and specificity were 87.5%, 88.5%, and 86.8%, respectively; for the external dataset from Best and colleagues, the classification accuracy, the sensitivity, and the specificity were 86.1%, 76.1%, and 94.5%, respectively. Other classification metrics including positive predictive value, negative predictive value, kappa coefficient, and F1 score were shown in Table 1. Moreover, we observed that the classification accuracy was comparable when stratified by different disease types and TNM stages. The sensitivities of identifying CRC patients in the training set were 93.3% (14/15) at stage I, 96.7% (29/30) at stage II, and 100% (34/34) at stage III. For the internal validation set, the sensitivities were 80% (8/10) for stage I, 88.9% (16/18) for stage II, and 91.7% (22/24) for stage III. We achieved comparable classification performance when incorporating the serum levels of CEA and CA199 into the identified gene panels in both the training set (0.926, 95% CI 0.877-0.976) and the validation set (0.933, 95% CI 0.883-0.984) (Fig. 5D). The sensitivities were 73% in the training set at 98% specificity and 76% in the internal validation set at 99% specificity (Fig. 5D).
TEP RNA profiling also showed high sensitivity and specificity in predicting the stages of CRC. An AUROC of 0.984 (95% CI 0.966-1.000) was achieved on the training set and 1.000 (95% CI 1.000-1.000) for the internal validation set (Fig. 5E). Apart from CRC identification and staging, TEP RNA profiles also allowed the classification of healthy donors, patients with noncancerous diseases, and CRC patients. We obtained a gene panel of 929 genes from the binary PSO algorithm for the classification of healthy donors, patients with noncancerous diseases, and CRC (Additional file 1: Table S4). The AUROCs were   Table S5).

Discussion
Blood-based liquid biopsy provides a potential noninvasive alternative for early CRC detection. Among various liquid biopsy procedures (such as circulating cell-free DNA sequencing and gut microbiome profiling) [31][32][33], TEP RNA profiles emerged as a promising marker of molecular diagnostics to detect CRC at early stages. TEP RNA profiling has been employed for multiclass cancer diagnosis and reportedly achieved high accuracy in detecting several cancer types, including CRC [20]. However, previous studies did not include IBDs, polyps, and adenomas, which are common among human populations and often confound CRC early detection. Here, we showed that the TEP RNA profile can effectively detect CRC patients at an early stage from the population with the inclusion of noncancerous diseases, as well as predicting the stages of CRC. Our study revealed that RNA profiles of blood platelets from healthy donors are distinct from those of patients with CRC and other noncancerous diseases, whereas RNA profiles of blood platelets from patients with CRC and noncancerous diseases were admixed in the linear space of heatmap representation even though they were separable (Additional file 1: Fig. S5, Fisher's exact test, p = 0.002). The SVM algorithm has high classification power as it projects data into higher dimensional space with kernel methods, which can model the nonlinear features embedded in the TEP RNA expression data that were not captured in heatmap representation. This underscored the importance of including noncancerous diseases in the control group when developing TEP-based CRC early detection methods. In the internal validation set, five patients with adenoma, two with polyps, and two with Crohn's disease were misclassified as CRC (Fig. 5). Follow-up of these misclassified noncancerous patients in this study should be conducted to determine their risks of CRC development in the future.
Pathologically, there is a cross-talk between blood platelets and cancer cells. Cancer cells participate in platelet activation and reshape platelet RNA profiles by their oncogenic transformation mechanisms. Meanwhile, platelets contain growth and angiogenic factors facilitating cancer progression, and also interact with immune cells, such as natural killer cells and neutrophils, to promote cancer cell evasion from immune surveillance. Functional analysis from our study indicates that gene sets related to platelet activation and platelet signature are upregulated in CRC patients, while immune-related pathways, such as TNF-α signaling via NF-κB and interferon responses, are downregulated in CRC patients. This observation is consistent with a previous study [18]. Notably, the platelet RNA profiles of patients with polyps or adenomas are clearly distinguishable from healthy donors, which has not been reported before, suggesting that the development of noncancerous diseases involves pathological interactions with platelets.
The identified gene panels achieved significantly higher performance than did serum protein biomarkers, such as CEA and CA199 (p < 0.001). The performance was validated by an internal validation set and an external validation set whose control group however only contained healthy donors. The accuracy of detecting CRC patients at different stages was comparable. When incorporating CEA and CA199, the sensitivity of the performance of the classifier was marginally improved, suggesting that the selected gene panel possesses the CEA/CA199 comparable or even Table 1 Classification metrics of SVM a Kappa measured the agreement between the predicted classification with true labels. F 1 was the harmonic average of precision (positive predicted value) and recall rates (sensitivity)

Performance metrics
The classification metrics of SVM across three data sets Although some other biomolecules, such as cell-free DNA and intestinal microbiome, employed by several liquid biopsies were also interrogated in the early CRC detection, they have their own limitations. The recall rates based on mutations of cell-free DNA depend on the sequencing depth and vary across different stages [31]. The profiles of microbiome collected from fecal samples and intestinal microenvironment during colonoscopy examination were reportedly valuable for early CRC detection [32], but microbiome data are readily affected by antibiotics and sample collection procedures [32,34,35]. The classification power of these early CRC detection approaches, including our TEP RNA profiling, is expected to increase in the future by including more samples, incorporating multiple types of datasets, and using deep learning algorithms that have higher feature representation learning capability.
Although we demonstrated the applicability of TEP RNA profiling for CRC screening, some limitations should be considered. In this study, the sample exclusion rate reached 29.4%, mainly due to the insufficient blood volume, nucleated cell contamination, or poor sample quality, thus burdening the sample collection process. These underscore the importance of optimizing platelet and platelet RNA isolation procedure for TEP RNA profiling. Furthermore, the CRC diagnostic efficiency of TEP RNA profiling still needs to be evaluated in prospective studies.

Conclusions
In summary, we showed that RNA profiles of blood platelet are potentially applicable for early CRC detection from noncancerous diseases. However, further validation, especially prospective validation, is required for further demonstrating the diagnostic significance of TEP RNA profiling.