Optimization of WALDI-MS for analyzing unprocessed FFPE tissue sections
Since WALDI-MS is a convenient ambient MS method for sample preparation-free tissue profiling, we have evaluated the possibility to analyze FFPE histological sections without any of the traditional sample preparation steps (i.e. the deparaffinization and the rehydration). The first assessment was performed on a FFPE rat brain tissue section which was analyzed in both positive and negative ion mode (Fig. 1). Surprisingly, a population of ions associated with the presence of lipids were detected albeit at lower intensity (10e3-10e4 TIC range) than in case of fresh-frozen tissue samples analyzed with the WALDI-MS (10e5-10e6 TIC range)(16). The detected species correspond to specific fatty acids (m/z 200-400) and phospholipids species (m/z 700-900) in positive (Fig. 1a) as well as negative (Fig. 1b) ion mode (top panel). The signals are tissue specific since they are not detected when analyzing paraffin alone (bottom panel). Lower ion intensities can be explained by the extremely low water content of the sample which reduces the desorption/ionization efficiency of the WALDI process. Thus, in order to increase the signal intensity of the detected species, we studied the effect of adding a matrix on top of the FFPE tissues. Since a sprayed layer of water tends to shrink into a single large droplet due to low viscosity and high surface tension, we have chosen to test the addition of a thin glycerol matrix layer. Glycerol has two main advantages. First, the glycerol O-H bond has a large absorption band that is remarkably close to the one of water (2882 vs. 2940 nm max absorption wavelength, respectively), and which provides important resonant excitation to generate a similar desorption/ionization process. Second, glycerol has already been successfully used as an infrared MALDI matrix in the past (28). Because of the high viscosity of glycerol, it is not possible to spray it pure and we have chosen to mix it with isopropyl alcohol (IPA). The tissues were, thus, uniformly sprayed with a 50% glycerol /50% IPA solution. After depositing glycerol on the FFPE tissue, the lipid intensities increased by 100-folds reaching absolute TIC intensities up to 10e6 (Fig. 1 middle panel) in both positive (Fig. 1a middle) and negative ion mode (Fig. 1b middle). In the negative ion mode, there is a clear increase of the fatty acid species, such as m/z 255.2 Palmitic acid, m/z = 281.2 Oleic acid, m/z 283.2 Stearic acid and m/z 303.2 Arachidonic acid; lipid species at m/z 747.5 [PA (22:6_18:0)-H]-, m/z 885.5 [PI (18:0_20:4)-H]- and m/z 907.5 as well as a ganglioside GM 1 (18:0) at m/z 1544.7. In the positive ion mode the major increase is observed for lipid species m/z 734.5 [PC (32:0)+H]+, 760.5 [PC(34:1)+H]+, m/z 782.5 [PC(34:1)+Na]+ and m/z 810.6 [PC(34:1)+Na]+. The extensive list of MS/MS detected fragments and identified lipid species is shown in Table S1. Furthermore, we evaluated the evolution of the MS molecular profiles and the wealth of information they contain according to the number of laser shots and time of spray. All spectra are shown in the Supplementary Information. Fig. S1 shows the mass spectra recorded in both polarities from a single laser shot. As revealed by the optical image (Fig. S1a) a partially ablated area of 0.16 mm2 is observed in case of the 8 µm thick tissue section. Importantly, there are no significant changes in the content of the molecular profiles in either (Fig. S1b) positive or (Fig. S1c) negative ion mode for a single laser shot but a slight decrease in overall TIC intensity compared to 10 laser shots. This indicates that only a few laser shots are necessary to achieve the requested sensitivity. Next, variation of the time (1-5 min) of the glycerol deposition was tested with a single laser shot (Fig. S2 and Fig. S3). In negative mode, the overall intensity of the mass spectrum is inversely proportional to the time of spray with maximum intensity reached after the first minute of deposition (Fig. S2). In positive ion mode, however, there are only slight variations of intensities between the different times of spray (Fig. S3). To minimize the sample preparation time and improve homogeneity of the matrix layer, shorter spray times and lower viscosity of glycerol solutions were tested (data not shown) revealing that 10s and 20% glycerol/IPA are sufficient to obtain a homogeneous layer, reproducible and high intensity signals (up to 10e5 TIC). To reinforce the importance of analyzing non-dewaxed FFPE samples, we then investigated the influence of tissue dewaxing on the detected species. After tissues were dewaxed and rehydrated according to standard protocols used in histology, mass spectra were recorded and compared to FFPE tissues analyzed without any sample treatment (Fig. 2). Dewaxed spectra (bottom) show a very different molecular profile compared to the non-dewaxed (top) with a similar trend observed in both positive (Fig. 2a) and negative (Fig. 2b) ion mode. In the dewaxed samples no signals attributed to the phospholipids were observed and only a few and low intensity signals are still found corresponding to fatty acids (e.g. signals at m/z 283.2 and 255.2). This confirms that the dewaxing procedure removes the residual lipids left behind by the dehydration process, making further lipid profiling pointless. We conclude that rapid and highly sensitive detection of lipids directly from FFPE tissue is possible using 20% glycerol in IPA and a spray time reduced down to 10s. Therefore, the complete process (sample preparation + analysis) requires no more than a few minutes.
Establishing a rapid FFPE diagnostic screening MS-based platform using WALDI-MS
Based on the aforementioned observations, we have created a platform for rapid lipidomic screening of FFPE tissues without deparaffinization and minimal sample preparation targeting applications in cancer diagnostics. The platform is operated in a two-stage workflow (Fig. 3). First (Fig. 3a), a molecular database using histologically annotated samples is constructed. The FFPE blocks from tumor tissue banks are sectioned in consecutive slides. The first section is stained and annotated by histopathology professionals; whilst the following section is analyzed by the SpiderMass system and the recorded mass spectra are subjected to data processing. First, non-supervised analysis by Principal Component Analysis (PCA) enables to check the reproducibility of the data and any potential issues associated with instrument drift or batch effects. Then, following the annotation of MS Spectra according to their histological classification, supervised analysis is performed by Linear Discriminant Analysis (LDA) to build-up classification models for the user-defined classes. Finally, these classification models are used to perform real-time screening of unknown tissues (few minutes) (Fig. 3b). The first part of the workflow (training) is crucially important for obtaining accurate results and must be achieved following well established procedures(16, 17). This part can take up to a few weeks (according to the number of samples, their availability and their annotation by an expert pathologist); though if tissues are available >100 samples can easily be analyzed per day. Then, after the training part is finalized, the identification of any sample becomes possible within <5 minutes using the platform.
Evaluating SpiderMass FFPE analytical performances from clinical samples
We then further evaluated the performances of the newly established methodology for the classification of real clinical cancer samples. This was achieved by analyzing a cohort of samples (29 patients; 76 samples) from canine veterinary patients with sarcoma. Canines, like humans, spontaneously develop soft tissue sarcomas, suffer from spontaneous tumor formation and mirror morphological appearances, clinico-pathological presentation, and phenotypes, which makes them an excellent comparative patients (29, 30). Especially, as for humans, canine sarcoma types and sub-types are numerous and the discrimination of certain sub-types represents a challenge even for an expert sarcoma pathologist. The canine FFPE classification model was constructed from the lipidomic molecular signatures, following the optimized sample preparation protocol. The collected FFPE blocs followed four main sarcoma classes from diverse breeds and tumor origin: Angiosarcoma, Leiomyosarcoma, Fibrosarcoma and Malignant tumor of peripheral nerve sheaths (Table S2). Following the collection of spectra in positive ion mode we were interested to see, if the lipido-molecular analysis of FFPE tissue would be able to discriminate between the sarcoma subtypes. The results are shown in Fig. S4. The PCA analysis (Fig. S4a) did not show any evident separation between the subtypes, however, PC2 shows a better separation between Leiomyosarcoma (red) and Angiosarcoma (grey), while, Fibrosarcoma (blue) and the Malignant tumor of peripheral nerve sheaths (violet) seem to have more similarities with angiosarcoma. The 3D-PCA model obtained reflects to intra- and inter patient heterogeneity and was in part associated with the different location of the initial tumor (e.g. skin, liver, spleen…). The selected PCAs were then subjected to supervised LDA analysis(13, 14) with the spectra grouped into 4-classes according to the sarcoma sub-types (Fig. S4c). A substantial discrimination is observed in LD1 from leiomyosarcoma compared to the remaining subtypes, however, they are also separating from each other in LD2 and LD3. The model gives excellent accuracies 99.05 % using “5-fold” and 98.45% using “leaving-one-patient out” method (Table 1). Since some of the canine sarcoma subtypes were under-represented, we reconstructed the model including only angiosarcoma and leyiomyosarcoma (Fig. S4b) and Fig. S4d)). In this case the cross-validation accuracy slightly improved to 99.51 % using “5-fold” and 99.45% using “leaving-one-patient out” method (Table 1). These first promising results show that it is possible to construct a sarcoma tissue-typing classification model based on the detected -molecular features from non-processed FFPE tissue samples.
Table 1
Cross-validation results. Cross-validation results of 4-class /2 -class dog sarcoma and 4-class/ 3-class human sarcoma models with model type, cross-validation type, N of classes, N spectra, N of passes as well as N of failures, N outliers and % of correct classification accuracies after cross-validation.
Model type
|
Cross-validation type
|
N
classes
|
N
spectra
|
N
passes
|
N
failures
|
N outliers
|
Correct classification accuracies (%)
|
Excluding outliers
|
Including outliers
|
Canine sarcoma
|
“5-fold”
|
4
|
235
|
208
|
2
|
25
|
99.05%
|
88.51%
|
“One- patient- out”
|
4
|
235
|
191
|
3
|
41
|
98.45%
|
81.28%
|
“5-fold”
|
2
|
204
|
202
|
1
|
1
|
99.51%
|
99.02%
|
“One- patient- out”
|
2
|
204
|
182
|
1
|
21
|
99.45%
|
89.28%
|
Human sarcoma
|
“5-fold”
|
4
|
116
|
104
|
10
|
2
|
91.23 %
|
89.66%
|
“One- patient- out”
|
4
|
116
|
78
|
28
|
10
|
73.58%
|
67.24%
|
“5-fold”
|
3
|
101
|
97
|
3
|
1
|
97.00%
|
96.04%
|
“One- patient- out”
|
3
|
101
|
77
|
19
|
5
|
80.21%
|
76.24%
|
Histological classification of human sarcoma FFPE tissues
In the next step, a conceptual databank was created using a cohort of human sarcoma FFPE samples (29 patients; 37 samples). Human sarcomas are known to be a very heterogeneous cancer presenting many challenges for accurate diagnostics (31, 32). FFPE sarcoma samples were classified (Fig. 4) based on two 2 different sarcoma types and 2 different classes (malignant and benign) (Fig. 4a). The samples were collected from patients with different age, sex, and tumor origin as shown in Table S4. Sarcomas are very rare lesions (1% of malignant tumors); therefore, it is difficult to obtain an equivalent number of samples of the same sub-type. The acquired spectra for each subtype are shown in Fig.S5. The PCA analysis (Fig. S6) shows that the first 3 PCs explain more than 80% of the total variance in the dataset (PC1 45.5 %, PC2 25.12 %, PC3 14.4 %). Similarly, to canine sarcoma samples, the PCA shows a spread in the acquired data across the plane, due to intra and inter patient heterogeneity. This is particularly true for lipoma and liposarcoma FFPE samples which have a considerable number of subtypes and tumor origin (Table S3). Selected PCs were than further used for supervised LDA analysis represented by the 3D plot in Fig. 4b. A good separation is observed among the four subtypes. The first component clearly discriminates between the two sarcoma types (LD 1) while the second between the lipoma and liposarcoma cancer types (LD 2). The third (LD3) discriminates between leiomyoma and leiomyosarcoma. The loadings plot of LD2 is shown in Fig. 4c. The LD2 loadings spectra shows the most discriminatory lipid peaks (blue stars) and correspond to m/z 855.8, 879.8, 881.8, 897.8, 900.8 and 907.8. MS/MS analysis was performed in order to elucidate the identity of the selected molecular species (Table S4 and Fig. S7). Lipids contributing to the discrimination mostly correspond to a mixture of triglycerides (TG). The model gave considerably good cross-validation accuracies 91.23% using “5-fold” and 73.78% using “leave-one-patient out” method (Table1) for the undersized and unbalanced number of human samples. However, if we remove the under-represented leiomyoma samples and reconstruct the 3 class-model again the cross-validation accuracies improve to 97% for “5-fold” and to 80% for “leaving-one patient out”. In both models few benign samples were miss-classified to their malignant counterpart and vice versa. In order to improve the cross-validation accuracies and the model robustness, more samples would be needed, which is difficult for sarcoma since it is a rare disease and certain subtypes are difficult to obtain. However, further expansion of the database for clinical purposes can be achieved at any time new samples are added to the model and reprocessed.
Blind screening in real-time of FFPE samples
To demonstrate the efficiency of our platform and human sarcoma model in a clinical environment, we simulated the tissue classification in real-time. Two FFPE human sarcoma samples of different subtypes were randomly selected by the pathologist and analyzed in blind using SpiderMass accordingly to the optimized analytical workflow. The previously 4-class generated human sarcoma PCA-LDA model was loaded into AMX recognition software(13). Each acquisition is automatically predicted, and the software gives out the readout of different color-coding and percent of certainty: Leiomyoma (Red), Leiomyosarcoma (Green), Lipoma (Turquoise) and Liposarcoma (Purple). The results of the two blind acquisitions are shown in Fig. 5. The interrogation software correctly predicted in real-time with over 95% probability value of the two subtypes of human sarcoma. In the first case, the software predicted the sample to be a benign Lipoma for all three acquisitions (Fig. 5a). The scan 53-55 was predicted with 99.25 % probability, scan 62-64 with 99.28 % probability and scan 71-72 was predicted with 98.25%. In the second case, the software predicted the sample to be a benign leiomyoma (red) (Fig. 5b). The prediction was with 96.21 % certainty for scan 11-12 and 97.82 % probability for scan 23-24, and finally 95.38% for scan 33-34. Following the fast screening, the results were compared to the histological annotations of the pathologist confirming the correct classification in both cases obtained in real-time with the SpiderMass platform. These results show it is possible to provide correct classification accuracy for blind sample, above 95%, already from a rather small training cohort of samples.
Effect of FFPE conservation time on molecular profiles
A particularly important aspect of the analysis of FFPE samples is to study the dependence of molecular profiles on the FFPE protocol as well as on the age of the samples. It is known that FFPE tissues undergo chemical changes with the conservation time. In order to check how this can affect our analysis, we investigated the effects of the embedding as well as age of the sample on the lipid profiles. In the first example, a recently embedded (2019) human lipoma sample was analyzed before and after sample preparation (Fig. 6a and Fig. 6b). The mass spectra following the deposition of 20% glycerol in IPA spray show a significant increase (x100) in signal intensities of some lipid species mainly assigned as the previously identified TG (52:4) / TG(50:1)Na+ at m/z 855.8, TG (54:6) / TG(52:3)Na+ at m/z 879.8, TG (54:5) / TG(52:2)Na+ at m/z 881.8 and TG (56:6) / TG(54:3)Na+ at m/z 907.8. For comparison we also added a spectrum of an un-treated fresh-frozen human lipoma sample (Fig. 6c). While most lipid peaks between m/z 800-950 are retained in FFPE tissue, several lipids in the mass range m/z 700-800 are not observed after fixation and paraffin embedding. The identity of the m/z 881.8 species was confirmed in fresh-frozen and FFPE tissue by MS/MS analysis (Fig. 7). In the second example, we performed a sample age dependence study, where we acquired MS spectra from Leiomyosarcoma samples at different years of embedding from 2014-2019. The results are shown in Fig. 8. First, we compared the lipid profiles obtained from the year 2014, 2015 and 2019 of embedding Fig. 8a. There is a clear drop in intensity observed when comparing the lipid profiles. In the lipid mass range m/z 600-900 the absolute intensities of the peaks in 2014 and 2015 are only about 10% of the value in 2019 (Fig. S8). Furthermore, PCA analysis of the collected samples (Fig. 8b) was performed to get an overview of the impact of conservation time. The PC 1 describes 62.27% variance; PC 2 describes 12.47% and PC 3 11.29%. There is a progressive separation of the data along PC 1 from the 2019 to a bundle at 2014, 2015 and 2016 with 2016 and 2017 in between. The loading-mass plot of PC1 indicated 4 most significant peaks which contribute to the variance m/z 643.6, 853.8, 879.8 and 881.8 (data not shown). Statistical tests were performed on the normalized intensities for each of peak and expressed as box plots with Tukey whisker definition Fig. 8c. The data show there is a significant difference between the intensities of the m/z 853.8, 879.8 and 881.8 between the samples embedded 2016 and 2019 and only m/z 881.8 between the 2015 and 2019. No significant difference is observed between the embedding years 2014 and 2019. The statistical tests, therefore, indicated that only a small fraction, the 3 of the most variant peaks (in the year 2016), show a significant difference. However, differences are observed, with losses of absolute intensities for certain signals, when analyzing the material embedded after a period (max 3 years).