Proteomic Analysis of BMAC from the Iliac Crest and Humeral Head Using Support Vector Machines

Abstract

The aim of this study was to analyze the proteomic differences in bone marrow aspirate concentrate (BMAC) between extraction sites, specifically, the iliac crest and humeral head. Using support vector machine (SVM) models, SVM­(linear) and SVM­(rbf), we identified six key proteins (MCSF, CD14, CD40, CD163, Ephrin-A4, and Matrilin-3) as essential for distinguishing between these sites. The results indicate that site-to-site proteomic differences are likely influenced by the surrounding tissues of each extraction site. Higher levels of chondrogenic and osteogenic proteins in the humeral head BMAC in addition to immune-modulatory proteins in the iliac crest BMAC suggest site-specific proteomic profiles. This study provides insights that may assist clinicians in selecting site-appropriate BMAC for therapeutic applications and inform future research into BMAC proteomics.

Introduction

In the world of sports and orthopedic medicine, biological therapeutics have become more prevalent for their ability to enhance existing treatments for musculoskeletal injuries and defects. − One of the common therapies is bone marrow aspirate concentrate (BMAC) which involves aspirating and concentrating down small volumes of bone marrow and injecting the product into the target site. ,− BMAC is known to have high concentrations of anti-inflammatory and anabolic growth factors such as platelet-derived growth factor-B (PDGF-B) as well as bone morphogenic proteins 2 and 7 (BMP) and interleukin-1 receptor antagonist (IL-1RA), which have been shown to inhibit the catabolic activity of IL-1, mesenchymal stem cells, and other progenitor cells. , Several studies have demonstrated the effectiveness of BMAC with respect to the need for revision surgeries; − one such example comes from a 2022 metadata analysis conducted at the Mayo Clinic which analyzed the Mariner data set from the PearlDiver patient records repository and found that across a 114 patient group, when administered at the time of a rotator cuff repair (RCR), BMAC was associated with a significant decrease in revision surgeries.

Bone marrow aspiration procedures are typically performed in an operating room, under anasthesia. In many cases, bone marrow is aspirated from the iliac crest; this is largely due to its relative abundance of marrow, yield of progenitor cells, and association with preferred patient outcomes. − While the iliac crest is often the preferred location for aspirating bone marrow for BMAC preparations, there may be patients for whom aspiration from a different location may be preferable.

In the case of a full thickness rotator cuff tear, the humeral head becomes an accessible source of bone marrow which is local to the site of repair. ,, Typically, extracting marrow from the humeral head is challenging due to the musculature of the rotator cuff, in addition to the other surrounding tissue which ordinarily blocks the humeral head. However, when a full thickness rotator cuff tear occurs, a gap opens at the musculotendinous junction. , This gap can be used to attach a surgical anchor for repair as well as aspirate marrow from the humeral head. With this, surgeons may not need to add a second procedure to extract marrow from the iliac crest; rather, they could perform the full surgery (repair and BMAC) in a single surgical site. Our work specifically examines BMAC samples originating from either the iliac crest or the humeral head, with the aim of characterizing and comparing the proteomes of each marrow to determine the degree of similarity they possess. As BMAC can be produced from both sites, it is important to understand the site-dependent biological characteristics of the marrow. This could enable treatment decisions that are more patient and condition specific.

To our knowledge, there are no pre-existing works which examine how the extraction site of BMAC changes its proteomic profile and how that affects its regenerative and therapeutic potential. The differences between the proteomes of the iliac-crest-derived BMAC and the humeral-head-derived BMAC were characterized with a robust support vector machine (SVM) learning analysis performed to highlight the nuances of each sample.

The aim of this work was to use SVM to determine the degree of separability or classifiability between the patient BMAC samples based on their proteomes. This information allowed us to analyze the biochemical comparability of the marrow’s regenerative potential based on its proteomic profiles with an emphasis on the relative levels of anti-inflammatory, anabolic, and catabolic cytokines between the iliac crest and humeral head derived BMAC samples (Figure ).

1.

Schematic overview of the experimental workflow used in this study. Bone marrow aspirate (BMA) samples were collected from patients and processed into a bone marrow aspirate concentrate (BMAC). Samples were then further separated into serum and cellular fractions, and the serum component was analyzed by using antibody microarrays. Raw fluorescent data were processed in Python and Excel to build a proteomic data set, which was then analyzed using inferential statistics and machine learning models to identify site-specific proteomic signatures. Created in BioRender, https://BioRender.com/14yza7k.

Experimental Procedures

Sample Acquisition

Bone marrow samples were aspirated (BMA) from patients (n = 105) presenting with symptoms of joint pain (Table ). BMA was processed into bone marrow aspirate concentrate (BMAC) by Sachdev Orthopedics as part of surgical intervention related to the onset of patient pain using a commercially available centrifuge (Magellan, ISTO Biologics) to isolate the buffy coat containing the MSCs. The procedure was performed as a medical operation with informed consent. Leftover materials were deidentified and anonymized prior to being used as part of the subsequent analysis.

1. Demographic Characteristics of Patients from Whom Bone Marrow Aspirate Concentrate (BMAC) Samples Were Acquired .

Groups	Iliac Crest (65)	Humeral Head (40)
Sex, (M/F)	27/38	21/19
Age, (Years)	61.12 ± 10.73	62.6 ± 9.57
Smoking Status. (current, former, never)	10, 13, 42	6, 11, 22
BMI (m/kg2)	30.18 ± 7.23	30.59 ± 6.13

^aData are presented for the iliac crest (n = 65) and humeral head (n = 40) cohorts, including sex distribution, age, BMI, and smoking status (current, former, never).

Ethical Statement on Samples

The specimens in question were collected under a routine clinical procedure and not as part of a clinical trial. The specimens were marked as waste and further deidentified prior to receipt. As such, Lehigh University did not classify this as Human Subjects Research.

Sample Preparation

BMAC samples were separated into their serum and cellular components via centrifugation (1500g × 5 min). Each BMAC serum sample was further processed using an Albumin & IgG Depletion SpinTrap (Cytiva) to remove excess solutes, immunoglobulins, and albumin proteins. Once stripped, each BMAC serum sample’s total protein concentration was quantified using a bicinchoninic acid assay (BCA) using the following kit Micro BCA Protein Assay Kit (Thermo Scientific, 23227). All kits were used in accordance with their manufacturer protocols. After stripping and determining concentrations of each BMAC serum sample, samples were diluted at a final concentration of 100 ug/mL to standardize the concentration for microarray analysis.

Microarray Slide Preparation and Scanning

BMAC serum samples at a final concentration to 100 ug/mL were loaded into four antibody microarrays (Table ) from Ray Biotech: Human immune response array GS1 (GSH-IMR-1-2), Human inflammation response GS3 (GSH-INF-3-2), Custom G-series Select “Osteoclast” Array (GSH-CUST), and Custom G-series Select “Cartilage” Array (GSH-CUST). The protocol for processing each set of arrays was provided by Ray Biotech.

2. Map of Each Array Used to Characterize the Patient Bone Marrow Aspirate Concentrate Samples Based on 109 Unique Proteins .

Array Protein Maps
Human Inflammation Array GS3					Human Immune Response Array GS1
BLC	Eotaxin	Eotaxin-2	G-CSF	GM-CSF	CD14	CD40	CD163	CRP	E-Selectin
I-309	ICAM-1	IFNg	IL-1a	IL-1b	Fas	FAS L	G-CSF	ICAM-1	IL-1a
IL-1ra	IL-2	IL-4	IL-5	IL-6	IL-1b	IL-2	IL-2 Ra	IL-4	IL-6
IL-6sR	IL-7	IL-8	IL-10	IL-11	IL-8	IL-10	IL-12p70	IL-13	IL-18
IL-12p40	IL-12p70	IL-13	IL-15	IL-16	Lipocalin-2	MCP-1	MCP-2	MIF	MIP-1a
IL-17	MCP-1	MCSF	MIG	MIP-1a	MIP-1b	OPN	PAI-I	PF4	Procalcitonin
MIP-1b	MIP-1d	PDGF-BB	RANTES	TIMP-1	RAGE	Resistin	ST2	Thrombomodulin	TNFa
TIMP-2	TNFa	TNFb	TNF RI	TNF RII	TREM-1	Troponin I	uPAR	VCAM-1	VEGF

Array Protein Maps
Custom G-series Select “Osteoclast” Array					Custom G-series Select “Cartilage” Array
BMP-4	CD109	EphA2	Ephrin-A4	FLRG	Aggrecan	bFGF	bIG-H3	BMP-2	BMP-5
Flt-3L	FSH	IFNg	IL-12 p40	IL-17	BMP-7	BMP-8	BMP-9	BMPR-IA	BMPR-IB
IL-20	IL-23	IL-23 R	ILT2	MCSF	BMPR-II	CHI3L1	FGF-4	FGF-6	FGF-9
M-CSF RA	MIP-1a	PTH	RANK	TGF-b1	IL-17F	Leptin	LRP-6	Lumican	Matrilin-3
TLR3	TLR4	TNFa	TRANCE	TREM-2	MBL	MMP-13A	NOV	sFRP-3	Thrombospondin-5

^aThe maps are broken down into the “Human Inflammation Array GS3”, “Human Immune Response Array GS1”, “Custom G-series Select ‘Osteoclast’ Array”, and the “Custom G-series Select ‘Cartilage’ Array”.

Data Preprocessing

Once processed, each array was loaded into Genepix 4000B Microarray scanner (Molecular Devices) and imaged using the “GenePix Pro 6 Microarray Acquisition & Analysis Software”. Proteins for each array type (Table ) were analyzed in quadruplicate. The fluorescent signal intensity values from ‘F532 Mean – B532’ for a given protein were included if the value of the spot did not exceed 1.2× the median value of the collective spots for a protein. The average number of replicates included for the Inflammation, immune response, cartilage, and osteoclast arrays across the entire data set were 3.5, 3.4, 3.5, and 3.3, respectively (max is n = 4). All spots which met the inclusion criteria were averaged, and a log2 transform was applied. Final values were normalized (eq ) to the average of two control samples (Human Serum Cellect, Fisher BioReagents, BP2657-100) run on opposite ends of each array. All data processing took place in Spyder IDE using Python 3.12 and Microsoft Excel.

$$\mathrm{normalized\; samples}=\frac{\mathrm{averaged\; raw\; sample}}{\mathrm{averaged\; raw\; control}}$$ 1

Comparative Inferential Statistical Analysis

The BMAC data set was stratified by anatomical source: Iliac crest (n = 65 samples) and humeral head (n = 40 samples). All 109 proteins were tested for normality using Shapiro–Wilk tests (α = 0.05). Due to non-normality in >70% of the proteins in both groups, nonparametric Mann–Whitney U (MWU) tests (two-tailed, α = 0.05) were employed for group comparisons. To control multiple comparisons (109 proteins = 109 comparisons), a Benjamini–Hochberg false discovery rate (FDR) correction was applied. Effect sizes were calculated using Cohen’s r, where positive (+) values indicate higher expression in the humeral head and negative (−) values the iliac crest.

Support Vector Machine Learning Analysis

The data set was comprised of 105 patient samples defined by 109 features with a single binary classification target variable (Bone marrow source). − To perform this task, a support vector machine (SVM) was imported “from sklearn.svm import SVC”, and two versions of this model were used: SVC­(kernel = ‘linear’, C = 1, class_weight = ‘balanced’, probability = True, gamma = ‘auto’) and SVC­(kernel = ‘rbf’, C = 1, class_weight = ‘balanced’, gamma = 0.37). The parameters C and gamma were assigned based on the grid search for theoretical optima and live testing. K-fold cross validation where K = 10, samples were shuffled, and the random state was fixed for reproducibility to address model overfitting. − ,−

Model Performance Evaluation

Accuracy scores were used to evaluate the model’s performance at a baseline. However, each model’s performance was further evaluated with a combination of precision, recall, F1-scores, and receiver operating characteristic curve with the area under the curve (ROC+AUC), as well as Mathew’s correlation coefficient. Each of these were calculated from a confusion matrix and verified using modules from the sklearn.metrics library.

Feature Importance Identification

To identify the significance of each feature (protein) with respect to the SVM­(linear), we examined the model coefficients (Table and Figure ). These model coefficients act as guides for the relationship of a given feature to a class output (humeral head or iliac crest). The sign of the value indicates the class that the feature likely belongs to, and the magnitude of the value ranks it among the other proteins. Permutation feature importance scores were used to evaluate the SVM­(rbf), similar to the SVM­(linear); the magnitude of the score indicates how important the feature was to the model (Table and Figure ).

5. Top 10 and Bottom 10 Protein Features Based on SVM (Linear) Model Coefficients .

Humeral Head Bias		Iliac Crest Bias
Feature	Model Weight ± Std Deviation	Feature	Model Weight ± Std Deviation
CD14	1.309 ± 0.114	BMP-5	–0.664 ± 0.089
CD40	1.284 ± 0.134	Lipocalin-2	–0.669 ± 0.183
CD163	1.228 ± 0.094	IL-18	–0.685 ± 0.076
IL-23 R	0.834 ± 0.095	IL-10	–0.781 ± 0.126
Lumican	0.805 ± 0.085	TIMP-1	–1.106 ± 0.159
Matrilin-3	0.792 ± 0.111	MCSF	–1.212 ± 0.061
CD109	0.770 ± 0.098	MIF	–1.224 ± 0.133
IL-17	0.730 ± 0.142	bFGF	–1.238 ± 0.291
CHI3L1	0.669 ± 0.138	Ephrin-A4	–1.268 ± 0.130
M-CSF R	0.622 ± 0.090	ST2	–1.348 ± 0.197

^aPositive coefficients indicate association with humeral head samples and negative coefficients association with iliac crest samples. Coefficients were averaged over 10-fold cross-validation; standard deviations reflect variability across folds.

4.

SVM­(linear): mean normalized fluorescent signal intensities with standard deviations of the 20 proteins that the SVM (linear) model coefficients suggested had the greatest impact on the model’s classification of a BMAC sample being labeled as humeral head or iliac crest. The graph shows the levels that each protein expressed at in each BMAC group.

6. Permutation Feature Importance (PFI) Scores for the SVM (RBF) Model .

Feature	Permutation Score ± Std Deviation	Feature	Permutation Score ± Std Deviation
FAS L	0.025 ± 0.013	IL-17F	0.002 ± 0.002
Matrilin-3	0.013 ± 0.007	PAI-I	0.002 ± 0.002
Ephrin-A4	0.011 ± 0.008	sFRP-3	0.001 ± 0.001
MIP-1a	0.011 ± 0.011	ICAM-1	0.001 ± 0.006
BMP-4	0.007 ± 0.009	IL-11	0.001 ± 0.002
MCSF	0.004 ± 0.007	VEGF	0.001 ± 0.002
CD14	0.003 ± 0.012	CD163	0.001 ± 0.003
Resistin	0.003 ± 0.002	IL-6	0.001 ± 0.002
IL-1b	0.002 ± 0.004	CD40	0.001 ± 0.004
IL-12p40	0.002 ± 0.003	IL-23	0.001 ± 0.003

^aScores reflect the change in model performance when a feature’s values were randomly permuted. Higher positive values indicate that the feature had a stronger influence on the model accuracy. Reported values are the mean ± standard deviation across 10 cross-validation folds.

5.

SVM­(RBF): mean normalized fluorescent signal intensities with standard deviations of the 20 proteins that the SVM (rbf) PFI scores suggested had the greatest impact on the model’s classification of a BMAC sample being labeled as humeral head or iliac crest. The graph shows the levels that each protein expressed at in each BMAC group.

Identification of Differentiating Proteome Features

After analyzing the feature importance of each SVM model, the overlap between what appeared to be the most important features in the context of BMAC sample classifiability and their functions identified as they related to BMAC was compared and analyzed (Figure and Table ).

6.

Venn diagram depicting the unique and overlapping top features for each SVM model. The overlapping features were deemed significant enough to investigate further.

7. Comparison of Average Normalized Fluorescent Signal Intensities for Six Proteins Identified as Being Important by Both SVM Models .

	Average Normalized Fluorescent Signal Intensity
Protein	Humeral Head	Iliac Crest	% Difference
CD163	0.735	0.684	7.24
CD14	0.946	0.749	23.25
Matrilin-3	0.571	0.544	4.77
CD40	0.724	0.656	9.88
Ephrin-A4	0.631	0.549	13.99
MCSF	1.501	1.584	5.39

^aPercent differences reflect the relative expression between humeral head and iliac crest BMAC samples, highlighting site-specific proteomic trends.

Results and Discussion

A total of 105 patient-derived bone marrow samples were characterized using four distinct antibody microarrays (Table ): inflammation, immune response, osteoclast, and chondrogenic arrays. These arrays were selected for their diverse analyte coverage and overlap with proteins implicated in BMAC’s therapeutic mechanisms. Selection prioritized proteins with established roles in BMAC function (e.g., inflammatory cytokines and osteogenic growth factors) while balancing practical constraints on sample volume and throughput. Although other arrays could have been informative in isolation, our selection captures many of the canonical pathways associated with BMAC. Notably, the design of this study allows for future horizontal expansion (additional proteins) and vertical expansion (additional patient observations).

Conventional methods for analyzing biological data typically involve comparative inferential statistics. − We elected to perform said analysis in conjunction with our SVM directed approach. Shapiro–Wilk tests revealed that only 24.8% (27/109) of the proteins in the iliac crest group and 25.7% (28/109) in the humeral headgroup followed a normal distribution, with just 7 proteins common between groups. , Given ∼75% of the proteins followed a non-normal distribution, MWU tests identified 24 proteins with a raw p-value <0.05 and after FDR correction 7 proteins adjusted p-value <0.05. The MWU test’s raw p-value reflects a 5% probability that the observed differences between the iliac crest and humeral head were by chance; with 109 proteins being tested, FDR correction was required to limit false discoveries. Cohen’s r effect sizes contextualize the biological relevance where the magnitude of the value indicates stronger group separations, and the direction of that separation is indicated by the positive or negative signs aligned with humeral head and iliac crest, respectively (Figure and Table ).

2.

Volcano plot of differential protein expression between iliac crest and humeral head BMAC samples. The y-axis shows the −log10 transformed p-values after Benjamini–Hochberg false discovery rate (FDR) correction (n = 109 samples). Proteins above the significance threshold (y = 1.0, FDR-adjusted p < 0.05) are statistically significant. The x-axis represents the Cohen’s r effect sizes where positive values denote higher expression in humeral head and negative values denote higher expression in iliac crest. Black diamonds indicate FDR-Significant proteins (p < 0.05), gray circles denote nonsignificant proteins.

3. Differentially Expressed Proteins between Iliac Crest and Humeral Head BMAC Samples .

Protein	Raw p-value	Cohen’s r	FDR p-value
BMP-7	0.0463	–0.1945	0.2134
BMP-9	0.0471	–0.1938	0.2134
CHI3L1	0.0002	0.3613	0.0163
CD14	0.0012	0.3120	0.0262
CD40	0.0017	0.3039	0.0309
CD163	0.0250	0.2183	0.1515
CRP	0.0097	0.2511	0.1135
IL-4	0.0255	–0.2177	0.1515
IL-10	0.0264	–0.2164	0.1515
IL-12p70	0.0456	–0.1951	0.2134
IL-18	0.0003	-0.3471	0.0163
Procalcitonin	0.0105	–0.2486	0.1135
ST2	0.0010	-0.3191	0.0262
Thrombomodulin	0.0250	–0.2183	0.1515
TREM-1	0.0025	-0.2924	0.0389
BLC	0.0005	-0.3349	0.0182
Eotaxin	0.0092	–0.2531	0.1135
GM-CSF	0.0125	–0.2428	0.1135
TIMP-1	0.0123	–0.2434	0.1135
TNFb	0.0346	–0.2061	0.1886
CD109	0.0259	0.2170	0.1515
EphA2	0.0463	0.1948	0.2134
M-CSF RA	0.0210	0.2247	0.1515
TRANCE	0.0255	–0.2180	0.1515

^aRaw p-value: Mann–Whitney U test results (unadjusted). FDR p-value: Benjamini–Hochberg corrected significance (adjusted). Cohen’s r: effect size magnitude/direction where positive values denote higher expression in humeral head and negative values denote higher expression in iliac crest. Rows in boldface type indicate proteins that remained statistically significant after FDR correction (p < 0.05). Proteins sorted by raw p-value (most to least significant).

While conventional inferential statistics (e.g., Mann–Whitney U tests) identify individual protein differences, they assume independence between features and struggle to model complex, high-dimensional relationships which are innately part of proteomic data. Given the moderate sample size of our data set (n = 105 total samples) and the correlative nature of protein expression patterns, we employed an alternative support vector machine (SVM) analysis.

Comparison of the BMAC proteome in iliac crest and humeral head samples could be considered as looking at the separability of the proteomes of both groups from one another. Ultimately, this becomes a classification task to identify the similarity (or similar proteins) between each BMAC sample. The support vector machine (SVM) machine learning model was selected for its classification ability in conjunction with use cases in similar biological classification problems. − ,

Fundamentally, the SVM looks to plot a decision boundary between two classes of data, which allows for classification from one class to another. The decision boundary is often termed a hyperplane and is oriented in such a way that maximizes the distance of the boundary from the closest data points of each class (support vectors). When a linear decision boundary (Figure A) is insufficient in classifying the data (usually due to high dimensionality), a kernel function (Figure B) can be implemented to expand the classification ability of the model by projecting the features into higher dimensional space. ,,,

3.

Schematic depicting the support vector machine classification decision boundary without the addition of kernel functions. (A) Base case where the data being classified are linearly separable. (B) How nonlinearly separable data can be handled via a kernel function. Figure was created in BioRender, https://BioRender.com/0pogcqy.

When fitting the model to a set of data, the parameters “C” and “Gamma” are used to optimize the fit of the model with respect to the data being analyzed. C refers to the regularization parameter, which controls the tradeoff between the setting the distance between the decision boundary and the support vectors (maximizing the margin) and minimizing the classification errors. A low value for C allows for some misclassification in the base data set, leading to better generalizability in novel data sets; higher values of C will prioritize correctness at the risk of overfitting. Typically values of C range from 1 to 10, where 10 is the standard. Gamma is used to tune how much influence a single observation has over the SVM decision boundary. Lower gamma values result in smoother decision boundaries that risk underfitting the data because the kernel has a wide decision radius. Conversely, a high gamma narrows the decision radius of the kernel, meaning that the nearby points have the greatest effect. Baseline gamma values are typically decided by taking the inverse of the number of features in the data set (1/109 in our case). It is important to note that when tuning gamma, it should be done with respect to C; high values for both C and gamma create overly complex decision boundaries which typically result in overfitting.

Two variations of the SVM with different kernel functions (linear and rbf) were used. The average of a 10-fold cross validation accuracy assessment of each of these models yielded an average accuracy of 0.829 ± 0.111 for the SVM­(linear) and 0.847 ± 0.122 for the SVM­(rbf). This result indicates that the SVM models were able to predict the origin of the testing BMAC samples with room for generalizability to novel data.

The average confusion matrices were then extracted for each model. A confusion matrix is a 2 × 2 with four components true positive (TP), false positive (FP), false negative (FN), and true negative (TN) (eq ). These components are the backbone behind more comprehensive analytics such as precision/recall scores, F1 scores, or Mathew’s correlation coefficients. , Take for example accuracy; this is a metric which can be broken down into eq :

$$\left[\begin{array}{ll}\mathrm{T}\mathrm{N}& \mathrm{F}\mathrm{P}\\ \mathrm{F}\mathrm{N}& \mathrm{T}\mathrm{P}\end{array}\right]\leftrightarrow \left[\begin{array}{ll}\mathrm{true\; HH}& \mathrm{false\; IC}\\ \mathrm{false\; HH}& \mathrm{true\; IC}\end{array}\right]$$ 2

$$\mathrm{accuracy}=\frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}+\mathrm{F}\mathrm{P}}$$ 3

Thus, one could evaluate each model’s performance in a more comprehensive manner by looking at the TP, TN, FN, and FP components as seen in eq , rather than solely looking at the accuracy score of a model.

Precision can therefore be used to look at the proportion of real true positive cases out of all positive (true and false) cases in the data set (eq ). , Under ideal circumstances, precision will be 1, where the number of FP is 0. In the case of the models used here, the SVM­(linear) had a precision score of 0.808 ± 0.173, whereas the SVM­(RBF) scored 0.792 ± 0.164, meaning they were respectively able to identify marrow from the iliac crest ∼80.8% and ∼79.2% of the time.

$$\mathrm{precision}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}$$ 4

Recall (alternatively known as sensitivity) examines out of all of the perceived instances of a positive case, how many of those were real (eq ). , Again, an ideal value for this would be 1. The SVM was able to correctly identify marrow as originating from the iliac crest at 0.825 ± 0.275 (linear) and 0.825 ± 0.195 (RBF).

$$\mathrm{recall}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$ 5

Both the precision and recall can be used in tandem with one another to create an F1-score (eq ) which takes the harmonic mean (balances the scores) of the precision and recall scores and outputs a value ranging between 0 and 1. The closer the F1-score is to 1, the better the model is at maximizing positive cases and minimizing false cases, and therefore the better it is at binary classification. , The SVM models produced F1-scores of 0.767 ± 0.180 (linear) and 0.801 ± 0.164 (RBF).

$$\mathrm{F1‐score}=2\times \frac{\mathrm{precision}\times \mathrm{recall}}{\mathrm{precision}+\mathrm{recall}}$$ 6

Another means of validating the model performance, aside from the F1-score, is by plotting the rate of false positive cases versus the rate of true positive cases; this is known as a receiver operating characteristic curve (ROC). Once plotted, the ROC curve is typically evaluated by finding the area under the curve (AUC) where the values range from 0 to 1. A score of 0.5 is the equivalent of random guessing whereas a score of 1 is perfect classification; the SVM models scored 0.842 ± 0.118 (linear) and 0.823 ± 0.167 (RBF). , However, the limitations of the ROC AUC are such that it does not provide the user with information about the positive predictive value or the negative predictive value used by the classifier. Adding the Mathew’s correlation coefficient (MCC) (eq ) overcomes this shortfall. This maximizes the TP, TN, FP, and FN values, giving us a score between −1 and 1. Where 1 corresponds to perfect prediction, 0 corresponds to a prediction made by random chance, and −1 corresponds to inversely perfect prediction, meaning all of the positive predictions were actually negative; the SVM models scored 0.685 ± 0.124 (linear) and 0.686 ± 0.261 (RBF). A summary of these results can be found in Table .

$$\mathrm{M}\mathrm{C}\mathrm{C}=\frac{\mathrm{T}\mathrm{P}\times \mathrm{T}\mathrm{N}-\mathrm{F}\mathrm{P}\times \mathrm{F}\mathrm{N}}{\sqrt{(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P})\times (\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N})\times (\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P})\times (\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N})}}$$ 7

4. Performance Metrics for SVM Models Used to Classify BMAC Samples by Anatomical Source .

Metrics/Models	SVM(linear)	SVM(RBF)
Accuracy	0.829 ± 0.111	0.847 ± 0.122
Precision	0.808 ± 0.173	0.792 ± 0.164
Recall	0.825 ± 0.275	0.825 ± 0.195
F1 Score	0.767 ± 0.180	0.801 ± 0.164
ROC+AUC	0.828 ± 0.124	0.842 ± 0.132
Matthews Correlation Coefficient (MCC)	0.685 ± 0.209	0.686 ± 0.261

^aMetrics include accuracy, precision, recall, F1-score, ROC+AUC, and Matthew’s correlation coefficient (MCC), with each presented as mean ± standard deviation over 10-fold cross-validation. Higher values indicate better classification performance.

When looking at a machine learning model’s performance overfitting is always a concern; this refers to a case when training a model, the parameters of the model are adjusted (tuned) too specifically to training data, resulting in a loss of generalizability to novel data. To combat the overfitting problem, we elected to perform a 10-Fold Cross Validation (K = 10) because with smaller data sets a larger K value helps achieve more stable and reliable model estimates. , The addition of K-fold cross validation (KCV) helps to reduce the model’s dependence on a specific training-testing split for a classification prediction, because every sample is used in training and testing at some point, which serves to improve the robustness of the model by averaging the results of the K folds together for its final score. ,

Additionally, the rationale for using a combination of SVM models was that the linear kernel has a low computational cost for training and testing the model, allowing for quick tuning and adjustments, which could be further applied to the RBF kernel. In addition to its efficiency the linear kernel has the added benefit of providing users with the feature model coefficients (Table and Figure ), ultimately providing better interpretability of the model, which helped to decide if a given sample is categorized as originating from the iliac crest or humeral head. Despite this, the linear kernel should be outperformed by a well-tuned RBF kernel as the features extend into higher dimensional space; however, that trade in performance comes at the cost of not having access to the model coefficients.

The model scores across the board indicate a strong performance in being able to perform binary classification between BMAC samples from the humeral head and iliac crest. To better understand the interplay of the proteome in marrow classification, the individual proteins themselves (rather than the model as a whole) were examined. To do so, the model weights were extracted from the SVM­(linear) and ranked based on magnitude and sign. These model coefficients act in two ways: first, they show how a feature weight influences the model’s prediction based on its sign. The SVM­(linear) model identified the top 10 ranking features most likely to be associated with the humeral head BMAC samples as [‘CD14’, ‘CD40’, ‘CD163’, ‘IL-23 R’, ‘Lumican’, ‘Matrilin-3’, ‘CD109’, ‘IL-17’, ‘CHI3L1’, ‘M-CSF R’], and the top 10 features most likely to be associated with the iliac crest BMAC samples as [‘BMP-5’, ‘Lipocalin-2’, ‘IL-18’, ‘IL-10’, ‘TIMP-1’, ‘MCSF’, ‘MIF’, ‘bFGF’, ‘Ephrin-A4’, ‘ST2’] (Table and Figure ). These 20 proteins (features) provided insight into how exactly the model was making its predictions and which features it deemed important with respect to classifying the marrows.

Unlike the SVM­(linear), the SVM­(RBF) does not give us access to the model coefficients, and the decision boundary cannot be cleanly visualized. Evaluating the SVM­(RBF) is more challenging because what is gained in its enhanced classification ability is lost in the model’s interpretability. ,,, To visualize how the various features might be important to the SVM­(RBF), the permutation feature importance (PFI) method was applied to each protein in our data set; PFI randomly shuffles a single feature value, evaluates how this affects the model performance, and then assigns a score to the shuffled feature. This is then repeated across all the features in the data set (Table and Figure ). A high PFI score indicates that permutating a protein’s value caused the model to perform worse, whereas a low PFI score indicates the model performed better. The intent of using this PFI method is to find which features are important to SVM­(RBF) for making its predictions. It should be noted that the scores of a given feature do not speak to the degree of biological relevance that the protein (feature) holds; rather, the scores enable assessment of the values of the feature as it pertains to the classification task. For example, while IL-23 has a low permutation score relative to Fas-L, this does not mean that IL-23 is not important. Rather, this indicates that when the value associated with IL-23 was permuted, this perturbation did not significantly impact the model’s ability to classify the BMAC samples. This distinction becomes important because the model is not speaking to the function of the protein in a system but rather how alterations affect the biological fingerprint of the bodies in which they exist in. For this data set, the top 20 PFI scores were for: [‘FASL’, ‘Matrilin-3’, ‘Ephrin-A4’, ‘MIP-1a’, ‘BMP-4’, ‘MCSF’, ‘CD14’, ‘Resistin’, ‘IL-1b’, ‘il-12p40’, ‘IL-17F’, ‘PAI-I’, ‘sFRP-3’, ‘ICAM-1’, ‘IL-11’, ‘VEGF’, ‘CD163’, ‘IL-6’, ‘CD40’, ‘IL-23’].

Of the 40 (out of 109) proteins examined between the SVM­(linear) model and the SVM­(RBF), there were 34 unique proteins identified as important for classification, with an overlap of 6 proteins between the models (Figure ). Those 6 proteins are Matrilin-3, Ephrin-A4, MCSF, CD163, CD40, and CD14.

Matrilin-3 is a well characterized extracellular matrix protein that helps in the formation of collagen networks. , Additionally, it plays a key role in cartilage development and ossification by helping modulate mesenchymal differentiation and chondrocyte differentiation. ,− Matrilin-3’s role in anabolic and catabolic activities appears to be contingent on its concentration; at concentrations of 100–200 ng/mL, it is shown to be able to modulate the activity of IL-1Ra, even if IL-1B is present in chondrocytes. Additionally, it has been shown to suppress MMP-13 activity in chondrocytes. , It is also able to inhibit BMP2 which is well-known for its role in the hypertrophic differentiation of chondrocytes; ,, however, it was shown when bound by Matrilin-3 this activity diminishes. , At higher concentrations of 5–50 ug/mL Matrilin-3 increased the expression of MMPs, IL-1B, 6 and 8 in chondrocytes. , Further investigation into the expression of Matrilin-3 could prove useful in determining how much of an effect its expression has on the downstream targets.

Ephrin-A4 has well characterized roles in skeletal development as a negative regulator of osteoclast activity in addition to helping form the perichondrium. , Notably it acts as a synergistic regulator of IGF1, which has downstream effects on cartilage matrix degradation, and vascularization through upregulated expression of VEGF. ,,

MCSF has roles in the proliferation, differentiation, and survival of bone marrow progenitor cells. This was found to be particularly true for osteoclasts which help to modulate bone mass; MCSF was found to be essential for the entire process of osteoclast differentiation spanning from the generation of precursor cells to the formation of mature osteoclasts. −

CD14 is a monocyte/macrophage differentiation antigen on the surface of myeloid lineage cells such as monocytes, macrophages, and dendritic cells. , Notably CD14+ monocytes play a central role in bone remodeling by way of differentiating into osteoclasts in the presence of MCSF and RANKL. Monocytes will circulate out of the bone marrow in both inflammatory and homeostatic conditions and once in circulation can differentiate into osteoclasts. ,

CD40 is a costimulatory signaling molecule in adaptive immunity that binds CD154 (CD40L) to provide secondary signals for T-cell dependent immune responses. Additionally, it has essential function in B-cell maturation and antibody class switching within the bone marrow. In the context of inflammatory responses CD40 and its ligand increase the expression of cell adhesion molecules, pro-inflammatory cytokines, chemokines, and MMPs according to Ziegler et al.

CD163 is a well characterized scavenger receptor primarily expressed on macrophages; it modulates oxidative stress and inflammation by clearing hemoglobin-haptoglobin complexes and acts as a marker for M2-polarized macrophages. , With respect to bone marrow aspirate concentrate, CD163+ macrophages contribute to immunomodulation by promoting a regenerative microenvironment by secretion of anti-inflammatory cytokines such as IL-10 and proangiogenic factors like VEGF. ,

Matrilin-3 is involved in cartilage development and extracellular matrix modulations; − ,− , Ephrin-A4, MCSF, and CD14 are involved in bone remodeling and osteoclast regulation. ,,,− CD40 and CD163, on the basis of the literature, are more involved in immunomodulation and inflammatory responses. − We speculate that the site-to-site protein expression levels (Table ) could be attributed to the tissue types surrounding the iliac crest and humeral head. The humeral head articulates in the glenohumeral joint which is a ball and socket joint, stabilized by the rotator cuff musculature, and is lined with articular cartilage. , However, the iliac crest serves as an attachment site for muscles and connective tissue; hence, it lacks a meaningful amount of cartilage. , Matrilin-3 had greater levels of expression in the humeral head BMAC samples, as did CD14, CD40, CD163, and Ephrin-A4. MCSF was slightly (∼5.38%) more expressed in the iliac crest population. Given the general roles of each of these proteins and their relative expression across each BMAC sample, we believe that the tissues surrounding the bone marrow extraction sites ultimately influence the local proteomes of each marrow. Notably between the results of the Mann–Whitney U comparison test and the SVM models, two proteins, CD14 and CD40, were mutually significant in each method.

Conclusion and Future Direction

Analysis of this data set reveals that while the feature importance determined by the models does not speak to the biological relevance of said feature, it can provide us with the insight necessary to look at the relationship between these proteins and bone marrow. This work was able to take BMAC samples from the humeral head and the iliac crest and identify ways in which they differ by inferential statistics. These nuances between the BMACs were further investigated with two flavors of the SVM machine learning model to get a better grasp on which proteins in the data set might be relevant. Based on varying feature performance metrics, we were able to isolate potentially key proteins which can be used to separate BMAC from the humeral head from that of the iliac crest. The results of this study can inform clinicians of the site-specific bone marrow nuances and can help guide future proteomic studies.

To the best of our knowledge and current review of the literature, there are no studies that perform a direct proteomic analysis comparing BMAC samples derived from the iliac crest to the humeral head. Typically, when BMAC proteomic analyses are performed, they revolve around a handful of key proteins (VEGF, BMPs, IL-1ra, IL-6, IL-8, TNF-a, TGF-B). Our study provides two things. First, we screen a wide selection of proteins which have varying degrees of function within the niches of inflammatory pathways, immunomodulation, osteocyte, and chondrocyte modulation. Second, our study provides a framework for others to come and apply a similar SVM directed approach to these BMAC samples (derived from any location, not just the iliac crest or humeral head) and perform a deeper analysis by looking at the interactions between a myriad of proteins rather than single observations. We would like to expand upon this work by delving deeper into the interplay between patient demographic factors and relative protein expression with an emphasis on site to stie differences.

The data underlying this study are openly available in Mendeley Data at https://data.mendeley.com/datasets/zdyx4zsvw9/1. DOI: 10.17632/zdyx4zsvw9.1

The specimens in question were collected under a routine clinical procedure, not as part of a clinical trial. Ethics approval and individual consent is not applicable.

The authors declare no competing financial interest.