NON-TISSUE CONSTITUENTS AS POTENTIAL CONFOUNDERS IN VIBRATIONAL SPECTROSCOPY ASSESSMENT OF TISSUE ENGINEERED CONSTRUCTS
Introduction
In tissue engineering, constructs
development is typically heterogenous in nature, where substantial differences
exist among constructs grown in the same batch under the same conditions(1,2). This necessitates non-destructive methods to
determine the readiness of the individual engineered tissue before implantation.
However, current golden standard for the analysis of tissue engineered
constructs are destructive, which is based on the analysis of samples that not
necessarily reflect the whole batch (1,3).
NIR spectroscopy coupled with machine
learning algorithm can be used as non-destructive assessment method of tissue
engineered constructs. This can
provide real-time information about the construct, which can be employed for in
situ monitoring of the tissue growth (3). Being a label free method is one of the main
advantages of NIR spectroscopy that enables it to be non-destructive. However,
being a label free method means that it measures everything in the sample, even
those constituents not related to the tissue, which makes every non-tissue
constituent a potential confounder.
In statistics, a
confounder is a variable that affect both the independent and dependent variables
i.e., in terms of machine learning, a variable affecting both the input and
output of the model. Thus, in biomedical spectroscopy, we can define a
confounder as a variable changing among the samples and affect the spectra of
the samples, and this effect can be detected by the employed machine learning algorithm.
Controlling
confounders effect is important as they can lead to misleading conclusions compromising
the study validity. Confounders can be avoided in the study design by restriction
where all variables are kept constant across the samples except for the target variable.
However, due to the high cost and the long time required by tissue engineering,
using sample from multiple experiments might be necessary to provide the high
number needed for a typical machine learning algorithm (1,4). Thus, spectra of constructs with varying
non-tissue constituents might be employed in the training and testing machine
learning algorithms for the analysis of a specific characteristic in the
construct. These non-tissue constituents include growth factors, reagents,
& regulator proteins. These,
This study addresses the hypothesis that non-tissue constituents of tissue-engineered constructs are associated with changes in the patterns of NIR Spectra of Engineered constructs that can be detected by multivariate analysis techniques. Such changes will make these constituents potential confounders that need to be controlled. To test this hypothesis
Alt1: “three specific aims have been addressed. The first aim is to determine the accuracy of partial least square discriminant analysis (PLS-DA) in classifying the NIR spectra of tissue-engineered constructs based on the presence of different non-tissue constituents independent of maturity of the constructs (based on incubation duration). The Second aim is to determine the accuracy of PLS-DA in classifying the NIR spectra of mature and immature tissue-engineered constructs (based on their incubation time: 7 days vs 28 days) in presence of diverse non-tissue constituents. While the third aim is to determine the impact of controlling the confounders on the PLS-DA models.”
Alt2: “the influence of 4 non-tissue constituents on the spectra of the constructs have been investigated by: 2 growth factors (BMP-9 and TGFb), M-PER reagent, & CDH2 protein.
Assess the performance of PLS-DA in classifying the spectra of the constructs according to the presence of these non-Tissue constituents
Methods
Confounders on the model and performance
To further elaborate the importance of controlling for the confounders, selected transformations (1st derivative followed by vector normalization) were applied to the whole 241 spectra and then exported to Unscrambler 10.4 X (Camo Software, Oslo, Norway). The 241 spectra were used to train and test a controlled and uncontrolled PLS-DA binary classifiers for the duration of incubation.
The 36 spectra of the 12 control samples were included in both training subsets of the uncontrolled and controlled PLS-DA models. In the uncontrolled PLS-DA training subset, the Day28 arm included all samples with M-PER, while they were excluded from the Day7 arm; the rest of the samples were selected randomly in both arms in the training subset and in the testing subset. In the controlled PLS-DA models, the confounders have been controlled by propensity score matching using the package MatchIt in R, where the propensity scores have been estimated with logistic regression and the scores were used for nearest neighbour matching without replacement(5). Each matched pairs were then included together in either the training or testing subsets and the propensity scores have been included in the model to ensure the that the confounders are controlled.
To avoid the replicates trap, the technical replicates were ensured to be together in the same subset (training or testing) and in the testing subsets they were averaged before testing i.e., the mean spectra of the 49 testing samples have been tested instead of the individual spectra. The numbers of preprocessed spectra in the training and testing subsets were based on the testing subset size that can provide a statistical significance at the acceptable accuracy level. A testing set of at least 47 samples will provide the study with power of at least 80% to detect a 20% increase in the accuracy of the model compared to the null hypothesis (70% compared to the null of 50%), with the use of a two-sided one-sample test of proportion at the 0.05 significance level.
The PLS-DA models were trained with a maximum 20 latent variables and using uncentred data. The Y matrix of both models was comprised of the Day7, Day28, and Control columns (with binary variables 0 & 1 indicating “No” and “Yes”, respectively); and the X matrix was comprised of the spectra only in the uncontrolled PLS-DA model, while it also included the propensity scores of the samples in the controlled PLS-DA model. Cross-validation was performed using 20 random segments.
Results & Discussion
Constructs characteristics and spectra
The 69 test samples included 3 biological replicates, where each sample included one or more of the non-tissue constituents that
Impact of controlling the confounders on the model interpretation and performance
The scores plot of the first 2 factors of the uncontrolled PLS-DA model (Fig 3A) shows a better differentiation of the samples according to the incubation duration compared to the covariate-adjusted PLS-DA model trained on the matched dataset (Fig 3B). However, this better differentiation is misleading where the scores plot of the uncontrolled model is actually differentiating the samples with M-PER from those without M-PER, as the samples with M-PER in the training dataset are only those incubated for 28 days.
Selected Info to be added in discussion:
Conclusion
PLS-DA can
classify NIR spectra of tissue-engineered constructs with different non-tissue
constituents independent of the incubation duration > Confounders
PLS-DA can
classify NIR spectra of tissue-engineered constructs based on incubation
duration (D7 & D28) in presence of different Non-tissue
constituents with better performance when confounders are controlled.
Confounders
should be considered in NIR based assessment of tissue engineered constructs.
Controlling the confounders by PS matching + Model Adjustment using Ps Shows
best performance.
All information
about the constructs need to considered in NIR-Based
Assessment of Tissue Engineered constructs
to create a
trustworthy NIR spectroscopy-based model for the assessment of tissue
engineered constructs, potential confounders need to be identified and
controlled in the model.
Tables
Table 1 The distribution of the non-tissue constituents in the tissue engineered constructs incubated for 7 and 28 days
|
Non-tissue constituent |
Day7 (n=41) |
Day28 (n=28) |
|
|
M-PER (%) |
No |
30 (73.2) |
18 (64.3) |
|
Yes |
11 (26.8) |
10 (35.7) |
|
|
N-Cadherin MP (%) |
No |
17 (41.5) |
12 (42.9) |
|
Yes |
24 (58.5) |
16 (57.1) |
|
|
BMP-9 (%) |
No |
19 (46.3) |
16 (57.1) |
|
Yes |
22 (53.7) |
12 (42.9) |
|
|
TGFb (%) |
No |
27 (65.9) |
21 (75.0) |
|
Yes |
14 (34.1) |
7 (25.0) |
|
Each sample may contain one or more of the non-tissue constituents
Table 2 The performance of the PLSDA Monte-Carlo double cross-validation in classifying the diverse non-tissue contents in the constructs
|
Classified Content |
Optimum |
Mean Accuracy |
Mean Specificity |
Mean Sensitivity (95% CI) |
|
M-PER |
Uvn |
0.953 (0.945-0.96) |
0.981 (0.975-0.988) |
0.875 (0.848-0.902) |
|
BMP-9 |
D2S11Uvn |
0.649 (0.628-0.669) |
0.706 (0.675-0.738) |
0.585 (0.548-0.621) |
|
N-Cadherin
MP |
Sm11D2S3Uvn |
0.6 (0.579-0.622) |
0.595 (0.56-0.63) |
0.63 (0.596-0.665) |
|
TGFb |
Sm5Uvn |
0.702 (0.684-0.72) |
0.86 (0.837-0.882) |
0.224 (0.184-0.263) |
Table 3 The performance of the PLSDA Monte-Carlo double cross-validation in classifying the constructs based on their culture duration (28 Days or 7 Days) with diverse methods for controlling the effect of the non-tissue constituents
|
Controlling Method |
Optimum Preprocessing |
Mean Accuracy (95% CI) |
Mean Specificity (95% CI) |
Mean Sensitivity (95% CI) |
|
Non |
D1S3Uvn |
0.814 |
0.84 |
0.769 |
|
Model Adjustment using Covariates |
Sm11D1S3 |
0.844
|
0.869 |
0.802
|
|
Propensity Score Matching |
Sm11Uvn |
0.864 |
0.956 |
0.575 |
|
Propensity Score Matching + Model Adjustment using Covariates |
Sm11Snv |
0.831
|
0.88
|
0.676
|
|
Propensity Score Matching + Model Adjustment using Propensity Scores |
D2S11 |
0.96 |
0.973 |
0.918 |
Table 4 The performance of the PLSDA Monte-Carlo double cross-validation in classifying the diverse non-tissue contents in the constructs after Propensity Score Matching followed by model adjustment using Propensity Scores
|
Classified Content |
Optimum |
Mean Accuracy |
Mean Specificity |
Mean Sensitivity-(95% CI) |
|
M-PER |
Sm5D2S3Uvn |
0.98(0.975-0.986) |
0.99(0.986-0.994) |
0.918(0.886-0.949) |
|
BMP-9
|
D1S9 |
0.974(0.965-0.983) |
1(1-1) |
0.944(0.925-0.963) |
|
N-Cadherin |
Sm11D2S3 |
0.64(0.612-0.669) |
0.551(0.501-0.602) |
0.716(0.675-0.757) |
|
TGFb |
Sm11Uvn |
0.98(0.975-0.984) |
0.999(0.997-1) |
0.859(0.829-0.889) |
Figures
Figure 1 mean raw spectra of the samples grouped according to incubation duration (A), incubation duration with controls as independent group (B), TGFb presence (C), BMP9 presence (D), N-Cadherin mimetic peptide presence (E), & M-PER presence (F)
Figure 2 Mean preprocessed spectra (1st derivative + Vector Normalization) of the samples grouped according to incubation duration (A), incubation duration with controls as independent group (B), TGFb presence (C), BMP9 presence (D), N-Cadherin mimetic peptide presence (E), & M-PER presence (F)
Figure 3 The scores plot of the first 2 factors of the uncontrolled PLS-DA (A) and the controlled PLS-DA model (B)
Figure 4 The cumulative variances explained by the 20 factors of the uncontrolled (A) and controlled (B) PLS-DA
Figure 5 The Root mean square of error of the cross-validation of the uncontrolled (A) and controlled (B) PLS-DA
References
1. Castro NJ, Babakhanova G, Hu J,
Athanasiou KA. Nondestructive testing of native and tissue-engineered medical
products: adding numbers to pictures. Trends Biotechnol. 2022;40(2):194–209.
2. Querido W, Falcon JM, Kandel S, Pleshko
N. Vibrational Spectroscopy and Imaging: Applications for Tissue Engineering.
Analyst. 2017 Nov 7;142(21):4005.
3. Kandel S, Querido W, Falcon JM, Reiners
DJ, Pleshko N. Approaches for In Situ Monitoring of Matrix Development in Hydrogel-Based
Engineered Cartilage. Tissue Eng Part C Methods. 2020 Apr 1;26(4):225–38.
4. Beleites C, Neugebauer U, Bocklitz T,
Krafft C, Popp J, Kra C, et al. Sample size planning for classification models.
Anal Chim Acta. 2013;760:25–33.
5. Ho DE, Imai K, King G, Stuart EA.
MatchIt : Nonparametric Preprocessing for Parametric Causal Inference. J Stat
Softw. 2011;42(8).