Standardizing the platelet sample preparation procedure
Platelets are easily activated during storage, isolation, and fixation, leading to obvious changes in their subcellular structures37,41. Therefore, when using platelet subcellular structures for clinical trials, how to minimize the interference of these in vitro factors on subcellular structures and reduce the activation of platelets during sample preparation is of great importance for accuracy and reliability. In previous reports, experimental conditions for the preparation of platelet samples (such as storage reagents, storage temperature, storage time, and fixation reagents) in different laboratories varied greatly, and the lack of uniformity makes it challenging to compare results from different studies. Therefore, we tried to establish a standardized protocol to prevent significant heterogeneity arising from the sample preparation.
Microtubules have been shown to undergo significant reorganization in their morphology during platelet activation. Microtubules are commonly thought to form marginal band rings in resting platelets and contract to obviously smaller rings upon platelet activation42-44. Therefore, we used SIM images of microtubules in platelets from healthy volunteers to determine whether sample preparation procedures caused platelet activation. Our results indicated that the predominant structures of microtubules (approximately 60%) in healthy donors were marginal band rings (Fig. 1a), consistent with previous reports44. Interestingly, in addition to the marginal band rings, in healthy donors, we also found wool ball-like structures (Fig. 1a, approximately 25%), which have not been previously reported in healthy subjects.
The sample preparation procedure consists of three steps: blood sample collection, platelet isolation, and immunofluorescence staining (Fig. 1b). In blood sample collection, both anticoagulant EDTA and trisodium citrate have been used in clinical vacutainers to inhibit instant platelet activation and aggregation45. Our results showed that the addition of EDTA can maintain the microtubule morphology longer than that of trisodium citrate (Fig. 1c, d). Even after 24 hours, more than 60% of platelet microtubules in the EDTA groups were marginal band rings (Fig. 1c, d). After blood drawing, in the clinic, whole blood is usually stored at 4 °C. However, our results indicated that after the whole blood was stored at 4 °C for more than 4 hours, microtubules depolymerized significantly, and the marginal band rings could not be maintained (Fig. 1c, d). In contrast, microtubules remained marginal band rings when whole blood was stored at room temperature (RT, 16–25 °C), even for 24 hours (Fig. 1c, d). The results suggest that RT is a better storage temperature than 4 °C for preserving platelet morphology, and the relatively large window of time makes distant collection and centralized processing possible.
Centrifugation is the main method for platelet isolation14. However, a certain degree of platelet activation was observed after centrifugation (Supplementary Fig. 1). Platelet activation and shape changes may be transient or irreversible, depending on the strength of the stimulus41. Therefore, in previous reports, the supernatant platelet-rich plasma (PRP) after centrifugation was usually diluted with a dilution buffer and incubated at 37 °C for a period of time (15 min to 1 day) to recover platelets from activation to a resting state30,37,46. However, these dilution buffers and incubation times varied widely and were not optimized. We incubated PRP in different dilution buffers for different periods of time. The results indicated that incubation in 1:9 ACD and Tyrode's-HEPES buffer for 2 hours recovered the platelets efficiently, resulting in more than 60% of microtubule structures being marginal band ring structures (Fig. 1e-g). These results suggest that the activation of platelets induced by centrifugation was transient and recoverable.
Different fixation reagents and formulations have been used by different laboratories to fix platelets for immunofluorescent staining. Our results indicated that the optimal fixation reagents for microtubules were PHEM buffer containing 3% paraformaldehyde, 0.05% glutaraldehyde, and 0.075% Triton X-100 (Fig. 1h). The morphologies of α-granules, mitochondria, and dense granules were better preserved with no obvious dispersive signals in platelets fixed with PHEM containing 4% paraformaldehyde (Fig. 1i-k). After fixation, platelets could be stored at 4 °C for up to 3 months or directly stained with immunofluorescence according to the manufacturer's instructions and then imaged, with no obvious difference in the quality of images (see Methods sections for details, Supplementary Fig. 2).
Classifying platelet subcellular structures based on SIM imaging
Among the currently available SRM methods, SIM was selected because it is simple to use and compatible with routinely used fluorescent labeling. In addition, it provides not the best but sufficient spatial resolution to discern platelet subcellular structures and, more importantly, is fast enough to be practical in a diagnostic scenario.
Utilizing SIM, we obtained superresolution images of four subcellular structures (i.e., mitochondria, dense granules, α-granules, and microtubules) of approximately 60,000 platelets from a variety of tumor patients (n = 20, including patients with glioma, cervical carcinoma, endometrial carcinoma, ovarian carcinoma, and myeloproliferative neoplasms) and healthy volunteers (n = 10). Our results showed that mitochondria and dense granules appeared as small dots distributed throughout the cells, which was consistent with previous reports (Fig. 2a, b)30,37. No statistically significant differences between the tumor patients and healthy controls were observed in the morphology (i.e., number and radial distribution) of mitochondria and dense granules (Fig. 2a, b, and Supplementary Fig. 3), implying that these two subcellular structures are relatively stable in response to tumor-related stimuli. Given the limited number of tumor types detected and the diversity of tumor types, changes in these structures cannot be ruled out in patients with undetected tumor types.
For microtubules, our quantification indicated that the predominant structures in healthy donors were the marginal band rings (approximately 60%) and wool ball-like structures (approximately 25%) (Fig. 1a). Remarkably, other unreported significantly different distribution patterns were found in the tumor patients, as shown in Fig. 2c. Based on these nanoscale distribution patterns, we classified microtubules into four categories: (i) a circular coil pattern, including marginal band rings and wool ball-like structures. It is worth noting that although wool ball-like structures are distinguished from marginal band rings, we put them into one category because in healthy donors, microtubules were mainly organized in these two patterns (approximately 85%), and their respective proportions varied greatly, but the total proportion did not change much between healthy individuals (Supplementary Fig. 4); (ii) a dispersion pattern characterized by a diffuse distribution throughout the cytoplasm; (iii) an irregular pattern characterized by twisted lines, such as semicircles and saddle shapes; and (iv) an aggregation pattern characterized by clustering in a small region (Fig. 2c). It is worth noting that although conventional confocal microscopy can discern the marginal band rings of microtubules, it is difficult to distinguish the wool ball-like structures from the dispersion pattern, whereas SIM can clearly distinguish them (Fig. 2d).
α-Granules are the most abundant secretory granules in platelets and are commonly thought to be distributed inside cells as small dots (200-500 nm in diameter). Their number is usually used as a parameter to characterize their performance39,47,48. Remarkably, our results showed that in addition to the number, there was also a significant difference in the subcellular distribution of α-granules between healthy controls and tumor patients. Through statistical analysis of their numbers and distribution patterns, we divided α-granules into four categories (Fig. 2e): (i) a scattered dot pattern with fewer than 30 dots in each platelet (i.e., n < 30 α-granule pattern); (ii) a scattered dot pattern with 30 or more dots in each platelet (i.e., n ≥ 30 α-granule pattern: the number 30 was used as the classification boundary because we found that the numbers of α-granules in healthy donors (n = 7404 platelets from 10 individuals) were mainly below 30 (Fig. 2f-h)); (iii) a circle pattern characterized by a circle of dots distributed around the cell membrane; and (iv) an aggregation pattern characterized by aggregation in a small region. It is worth noting that, to our knowledge, this is the first time the dispersion, irregularity, and aggregation patterns of microtubules and the circle and aggregation patterns of α-granules have been described.
Developing an unbiased high-throughput automated image analysis workflow
SIM can deliver high-throughput superresolution images, and it is impossible to manually analyze the resulting large datasets, especially in diagnostic scenarios. Therefore, we tried to develop objective and automated image analysis for platelet microtubules and α-granules based on their distribution patterns. An overall workflow is schematically shown in Fig. 3a, including three convolutional neural network (CNN) models for cell segmentation, microtubule classification, and α-granule classification.
First, we trained a CNN model based on the ResUNet framework with both fluorescence and bright field images of platelets to segment and obtain fluorescence images of individual platelets49,50. Care was taken to minimize the potential influence of platelet adhesion and stacking on the analysis of the distribution patterns (Fig. 3a i-iii, see Methods sections for details). The accuracy of this network in the test set was approximately 98.0% (Fig. 3b, see Methods sections for details).
Then, two CNN models based on the ResNet-50 framework were trained using approximately 33,000 images of individual platelets to classify microtubules and α-granules based on their distribution patterns (Fig. 3a iv, see Methods sections for details)51. Microtubules were divided into four categories: circular coil, dispersion, irregularity, and aggregation, and α-granules were divided into three categories: scattered dot, circle, and aggregation. The classification activation maps (CAM) show that these models could effectively capture the morphological characteristics of the fluorescence images (Fig. 3a v, vi). Most of the images were classified into the correct groups in the diagonal line of the confusion matrix, with an average diagonal element value of approximately 91% (Fig. 3c, d). Finally, after classification into a scattered dot distribution pattern, α-granules were further divided into two categories (n < 30 or n ≥ 30) according to the number of α-granules per platelet (Fig. 3a vi). The sparse deconvolution algorithm52 and Otsu method53 were used to automatically count the number of α-granules in each platelet, with an accuracy of approximately 95% (Fig. 3e, see Methods sections for details).
Comparing subcellular structures of platelets from tumor patients, healthy controls and benign mass patients
Using platelets to support cancer diagnosis and monitoring is not a novel concept. Previous studies on platelet-based cancer diagnosis have mainly focused on platelet clinical characteristics (e.g., count and volume)21-24 or molecular contents (e.g., RNA and protein)25-28. The diagnosis of cancer by SRM based on platelet subcellular structures has not been reported. Which features, if any, of platelet subcellular structures obtained from SRM are potential biomarkers and their relevance to cancer diagnosis remain unknown.
After approval by the institutional review committee for human research, we prospectively collected peripheral blood samples (4 mL) from patients with newly diagnosed invasive cancers or benign masses prior to any chemotherapy or surgical treatment. The patient cohort (Fig. 4a) included five tumor types, i.e., myeloproliferative neoplasm (n = 12), glioma (n = 16), cervical carcinoma (n = 31), endometrial carcinoma (n = 22), and ovarian carcinoma (n = 27) and two benign diseases, i.e., endometrial hyperplasia (n = 8) and ovarian cyst (n = 14). None of the patients had coexisting inflammation or were taking medications known to affect platelet function. Healthy volunteers (n = 33) aged 21-54 with no history or signs of cancer or other serious disease were recruited from the general population as controls. Details of the participants are presented in Supplementary Table 1-7. Following our standardized sample preparation procedure and automated image analysis workflow, a total of approximately 200,000 superresolution images of individual platelets were obtained and automatically analyzed. For each participant, at least 500 superresolution images of individual platelets were collected for each subcellular structure.
As shown in Fig. 4b, α-granules in platelets from healthy donors were mainly scattered dots (approximately 85%) with fewer than 30 in each platelet (i.e., n < 30 α-granule pattern). No significant difference was observed between patients with myeloproliferative neoplasm and healthy donors (Fig. 4c-f and Supplementary Fig. 5). In contrast, patients with glioma and the gynecological cancers (i.e., cervical, endometrial and ovarian cancers) had significantly lower percentages of platelets with the n < 30 α-granule pattern than the healthy controls (Fig. 4c). Statistical analysis of each distribution pattern of α-granules indicated that the glioma patients showed a marked increase in the percentage of platelets with n ≥ 30 α-granule pattern, which was 7 times higher than that in healthy donors (20.42±19.82% vs. 2.89±5.13%, Fig. 4d, e). The frequency distribution curve of glioma patients also showed a significantly wider range than that of healthy controls (Fig. 4f and Supplementary Fig. 5). The receiver-operating characteristic (ROC) analysis demonstrated that the parameter (n ≥ 30) discriminated well between the glioma cases and healthy controls with an area under the curve (AUC) of 84.8% (Fig. 4g). Strikingly, in addition to the increase in the number of α-granules (Fig. 4f and Supplementary Fig. 5), patients with the gynecological cancers had a significant increase in platelets with a circular α-granule distribution pattern (i.e., circle, Fig. 4d, e). Moreover, the increase occurred only in the cancer patients, and there was no significant difference between patients with the gynecological benign diseases (i.e., ovarian cysts and endometrial hyperplasia) and healthy subjects (Fig. 4h). The AUC of the parameter (circle) was 85.2%, implying that the parameter has the potential to distinguish benign and malignant gynecological diseases well (Fig. 4i).
We also compared the distribution pattern of α-granules in platelets from patients with cervical carcinoma (n = 3) and ovarian carcinoma (n = 3) before and after clinical treatment. Remarkably, after clinical treatment, the distribution pattern of α-granules of the tumor patients changed from 69% circle pattern to 72% n < 30 pattern with an AUC of 91.7% and were no longer significantly different from those of healthy donors (Fig. 4j, k), consistent with the significant reduction in carbohydrate antigen 125 (CA125) and squamous cell carcinoma antigen (SCC-Ag) levels (Supplementary Table 8). The results suggest that the distribution patterns of α-granules in platelets responded sensitively to clinical treatment and have potential for use in monitoring the treatment process of cervical cancer and ovarian carcinoma.
To our surprise, although we found that the distribution pattern (dispersion, irregularity, or aggregation) of some microtubules in platelets of the tumor patients was different from that of healthy controls (mainly circular coil) (Fig. 2c), statistical analysis showed no significant difference between the tumor and healthy groups (Fig. 4l). The results suggest that the microtubules in platelets were morphologically heterogeneous, and a sufficient number of samples rather than individual samples with a distinct distribution pattern and statistical analysis were necessary to ensure the reliability of the results.