A neural network model for cell classification based on single-cell biomechanical properties

Deep learning strategy for small dataset from atomic force microscopy mechano-imaging on macrophages phenotypes

The cytoskeleton is involved during movement, shaping, resilience, and functionality in immune system cells. Biomarkers such as elasticity and adhesion can be promising alternatives to detect the status of cells upon phenotype activation in correlation with functionality. For instance, professional immune cells such as macrophages undergo phenotype functional polarization, and their biomechanical behaviors can be used as indicators for early diagnostics. For this purpose, combining the biomechanical sensitivity of atomic force microscopy (AFM) with the automation and performance of a deep neural network (DNN) is a promising strategy to distinguish and classify different activation states. To resolve the issue of small datasets in AFM-typical experiments, nanomechanical maps were divided into pixels with additional localization data. On such an enlarged dataset, a DNN was trained by multimodal fusion, and the prediction was obtained by voting classification. Without using conventional biomarkers, our algorithm demonstrated high performance in predicting the phenotype of macrophages. Moreover, permutation feature importance was employed to interpret the results and unveil the importance of different biophysical properties and, in turn, correlated this with the local density of the cytoskeleton. While our results were demonstrated on the RAW264.7 model cell line, we expect that our methodology could be opportunely customized and applied to distinguish different cell systems and correlate feature importance with biophysical properties to unveil innovative markers for diagnostics.

scDLC: a deep learning framework to classify large sample single-cell RNA-seq data

Background

Using single-cell RNA sequencing (scRNA-seq) data to diagnose disease is an effective technique in medical research. Several statistical methods have been developed for the classification of RNA sequencing (RNA-seq) data, including, for example, Poisson linear discriminant analysis (PLDA), negative binomial linear discriminant analysis (NBLDA), and zero-inflated Poisson logistic discriminant analysis (ZIPLDA). Nevertheless, few existing methods perform well for large sample scRNA-seq data, in particular when the distribution assumption is also violated.

Results

We propose a deep learning classifier (scDLC) for large sample scRNA-seq data, based on the long short-term memory recurrent neural networks (LSTMs). Our new scDLC does not require a prior knowledge on the data distribution, but instead, it takes into account the dependency of the most outstanding feature genes in the LSTMs model. LSTMs is a special recurrent neural network, which can learn long-term dependencies of a sequence.

Conclusions

Simulation studies show that our new scDLC performs consistently better than the existing methods in a wide range of settings with large sample sizes. Four real scRNA-seq datasets are also analyzed, and they coincide with the simulation results that our new scDLC always performs the best. The code named “scDLC” is publicly available at https://github.com/scDLC-code/code.

Supplementary Information

The online version contains supplementary material available at (10.1186/s12864-022-08715-1).

Category encoding method to select feature genes for the classification of bulk and single-cell RNA-seq data

Bulk and single-cell RNA-seq (scRNA-seq) data are being used as alternatives to traditional technology in biology and medicine research. These data are used, for example, for the detection of differentially expressed (DE) genes. Several statistical methods have been developed for the classification of bulk and single-cell RNA-seq data. These feature genes are vitally important for the classification of bulk and single-cell RNA-seq data. The majority of genes are not DE and they are thus irrelevant for class distinction. To improve the classification performance and save the computation time, removal of irrelevant genes is necessary. Removal will aid the detection of the important feature genes. Widely used schemes in the literature, such as the BSS/WSS (BW) method, assume that data are normally distributed and may not be suitable for bulk and single-cell RNA-seq data. In this article, a category encoding (CAEN) method is proposed to select feature genes for bulk and single-cell RNA-seq data classification. This novel method encodes categories by employing the rank of sequence samples for each gene in each class. Correlation coefficients are considered for gene and class with the rank of sample and a new rank of category. The highest gene correlation coefficients are considered feature genes, which are the most effective for classifying bulk and single-cell RNA-seq dataset. The sure screening method was also established for rank consistency properties of the proposed CAEN method. Simulation studies show that the classifier using the proposed CAEN method performs better than, or at least as well as, the existing methods in most settings. Existing real datasets were analyzed, with the results demonstrating superior performance of the proposed method over current competitors. The application has been coded into an R package named "CAEN" to facilitate wide use.

A constriction channel analysis of astrocytoma stiffness and disease progression

Studies have demonstrated that cancer cells tend to have reduced stiffness (Young's modulus) compared to their healthy counterparts. The mechanical properties of primary brain cancer cells, however, have remained largely unstudied. To investigate whether the stiffness of primary brain cancer cells decreases as malignancy increases, we used a microfluidic constriction channel device to deform healthy astrocytes and astrocytoma cells of grade II, III, and IV and measured the entry time, transit time, and elongation. Calculating cell stiffness directly from the experimental measurements is not possible. To overcome this challenge, finite element simulations of the cell entry into the constriction channel were used to train a neural network to calculate the stiffness of the analyzed cells based on their experimentally measured diameter, entry time, and elongation in the channel. Our study provides the first calculation of stiffness for grades II and III astrocytoma and is the first to apply a neural network analysis to determine cell mechanical properties from a constriction channel device. Our results suggest that the stiffness of astrocytoma cells is not well-correlated with the cell grade. Furthermore, while other non-central-nervous-system cell types typically show reduced stiffness of malignant cells, we found that most astrocytoma cell lines had increased stiffness compared to healthy astrocytes, with lower-grade astrocytoma having higher stiffness values than grade IV glioblastoma. Differences in nucleus-to-cytoplasm ratio only partly explain differences in stiffness values. Although our study does have limitations, our results do not show a strong correlation of stiffness with cell grade, suggesting that other factors may play important roles in determining the invasive capability of astrocytoma. Future studies are warranted to further elucidate the mechanical properties of astrocytoma across various pathological grades.

Engineering approaches for characterizing soft tissue mechanical properties: A review

From cancer diagnosis to detailed characterization of arterial wall biomechanics, the elastic property of tissues is widely studied as an early sign of disease onset. The fibrous structural features of tissues are a direct measure of its health and functionality. Alterations in the structural features of tissues are often manifested as local stiffening and are early signs for diagnosing a disease. These elastic properties are measured ex vivo in conventional mechanical testing regimes, however, the heterogeneous microstructure of tissues can be accurately resolved over relatively smaller length scales with enhanced spatial resolution using techniques such as micro-indentation, microelectromechanical (MEMS) based cantilever sensors and optical catheters which also facilitate in vivo assessment of mechanical properties. In this review, we describe several probing strategies (qualitative and quantitative) based on the spatial scale of mechanical assessment and also discuss the potential use of machine learning techniques to compute the mechanical properties of soft tissues. This work details state of the art advancement in probing strategies, associated challenges toward quantitative characterization of tissue biomechanics both from an engineering and clinical standpoint.

Predicting cancer cell invasion by single-cell physical phenotyping

The physical properties of cells are promising biomarkers for cancer diagnosis and prognosis. Here we determine the physical phenotypes that best distinguish human cancer cell lines, and their relationship to cell invasion. We use the high throughput, single-cell microfluidic method, quantitative deformability cytometry (q-DC), to measure six physical phenotypes including elastic modulus, cell fluidity, transit time, entry time, cell size, and maximum strain at rates of 102 cells per second. By training a k-nearest neighbor machine learning algorithm, we demonstrate that multiparameter analysis of physical phenotypes enhances the accuracy of classifying cancer cell lines compared to single parameters alone. We also discover a set of four physical phenotypes that predict invasion; using these four parameters, we generate the physical phenotype model of invasion by training a multiple linear regression model with experimental data from a set of human ovarian cancer cells that overexpress a panel of tumor suppressor microRNAs. We validate the model by predicting invasion based on measured physical phenotypes of breast and ovarian human cancer cell lines that are subject to genetic or pharmacologic perturbations. Taken together, our results highlight how physical phenotypes of single cells provide a biomarker to predict the invasion of cancer cells.

Towards rapid prediction of drug-resistant cancer cell phenotypes: single cell mass spectrometry combined with machine learning

Combined single cell mass spectrometry and machine learning methods is demonstrated for the first time to achieve rapid and reliable prediction of the phenotype of unknown single cells based on their metabolomic profiles, with experimental validation. This approach can be potentially applied towards prediction of drug-resistant phenotypes prior to chemotherapy.

High-Throughput Assessment of Cellular Mechanical Properties

Traditionally, cell analysis has focused on using molecular biomarkers for basic research, cell preparation, and clinical diagnostics; however, new microtechnologies are enabling evaluation of the mechanical properties of cells at throughputs that make them amenable to widespread use. We review the current understanding of how the mechanical characteristics of cells relate to underlying molecular and architectural changes, describe how these changes evolve with cell-state and disease processes, and propose promising biomedical applications that will be facilitated by the increased throughput of mechanical testing: from diagnosing cancer and monitoring immune states to preparing cells for regenerative medicine. We provide background about techniques that laid the groundwork for the quantitative understanding of cell mechanics and discuss current efforts to develop robust techniques for rapid analysis that aim to implement mechanophenotyping as a routine tool in biomedicine. Looking forward, we describe additional milestones that will facilitate broad adoption, as well as new directions not only in mechanically assessing cells but also in perturbing them to passively engineer cell state.

Hydrostatic pressure promotes the proliferation and osteogenic/chondrogenic differentiation of mesenchymal stem cells: The roles of RhoA and Rac1

Our previous studies have shown that hydrostatic pressure can serve as an active regulator for bone marrow mesenchymal stem cells (BMSCs). The current work further investigates the roles of cytoskeletal regulatory proteins Ras homolog gene family member A (RhoA) and Ras-related C3 botulinum toxin substrate 1 (Rac1) in hydrostatic pressure-related effects on BMSCs. Flow cytometry assays showed that the hydrostatic pressure promoted cell cycle initiation in a RhoA- and Rac1-dependent manner. Furthermore, fluorescence assays confirmed that RhoA played a positive and Rac1 displayed a negative role in the hydrostatic pressure-induced F-actin stress fiber assembly. Western blots suggested that RhoA and Rac1 play central roles in the pressure-inhibited ERK phosphorylation, and Rac1 but not RhoA was involved in the pressure-promoted JNK phosphorylation. Finally, real-time polymerase chain reaction (PCR) experiments showed that pressure promoted the expression of osteogenic marker genes in BMSCs at an early stage of osteogenic differentiation through the up-regulation of RhoA activity. Additionally, the PCR results showed that pressure enhanced the expression of chondrogenic marker genes in BMSCs during chondrogenic differentiation via the up-regulation of Rac1 activity. Collectively, our results suggested that RhoA and Rac1 are critical to the pressure-induced proliferation and differentiation, the stress fiber assembly, and MAPK activation in BMSCs.

Screening of hyaluronic acid-poly(ethylene glycol) composite hydrogels to support intervertebral disc cell biosynthesis using artificial neural network analysis

Hyaluronic acid (HA)-poly(ethylene glycol) (PEG) composite hydrogels have been widely studied for both cell delivery and soft tissue regeneration applications. A very broad range of physical and biological properties have been engineered into HA-PEG hydrogels that may differentially affect cellular "outcomes" of survival, synthesis and metabolism. The objective of this study was to rapidly screen multiple HA-PEG composite hydrogel formulations for an effect on matrix synthesis and behaviors of nucleus pulposus (NP) and annulus fibrosus (AF) cells of the intervertebral disc (IVD). A secondary objective was to apply artificial neural network analysis to identify relationships between HA-PEG composite hydrogel formulation parameters and biological outcome measures for each cell type of the IVD. Eight different hydrogels were developed from preparations of thiolated HA (HA-SH) and PEG vinylsulfone (PEG-VS) macromers, and used as substrates for NP and AF cell culture in vitro. Hydrogel mechanical properties ranged from 70 to 489kPa depending on HA molecular weight, and measures of matrix synthesis, metabolite consumption and production and cell morphology were obtained to study relationships to hydrogel parameters. Results showed that NP and AF cell numbers were highest upon the HA-PEG hydrogels formed from the lower-molecular-weight HA, with evidence of higher sulfated glycosaminoglycan production also upon lower-HA-molecular-weight composite gels. All cells formed more multi-cell clusters upon any HA-PEG composite hydrogel as compared to gelatin substrates. Formulations were clustered into neurons based largely on their HA molecular weight, with few effects of PEG molecular weight observed on any measured parameters.

Electromechanical and elastic probing of bacteria in a cell culture medium

Rapid phenotype characterization and identification of cultured cells, which is needed for progress in tissue engineering and drug testing, requires an experimental technique that measures physical properties of cells with sub-micron resolution. Recently, band excitation piezoresponse force microscopy (BEPFM) has been proven useful for recognition and imaging of bacteria of different types in pure water. Here, the BEPFM method is performed for the first time on physiologically relevant electrolyte media, such as Dulbecco's phosphate-buffered saline (DPBS) and Dulbecco's modified Eagle's medium (DMEM). Distinct electromechanical responses for Micrococcus lysodeikticus (Gram-positive) and Pseudomonas fluorescens (Gram-negative) bacteria in DPBS are demonstrated. The results suggest that mechanical properties of the outer surface coating each bacterium, as well as the electrical double layer around them, are responsible for the BEPFM image formation mechanism in electrolyte media.

Cellular mechanical properties reflect the differentiation potential of adipose-derived mesenchymal stem cells

The mechanical properties of adipose-derived stem cell (ASC) clones correlate with their ability to produce tissue-specific metabolites, a finding that has dramatic implications for cell-based regenerative therapies. Autologous ASCs are an attractive cell source due to their immunogenicity and multipotent characteristics. However, for practical applications ASCs must first be purified from other cell types, a critical step which has proven difficult using surface-marker approaches. Alternative enrichment strategies identifying broad categories of tissue-specific cells are necessary for translational applications. One possibility developed in our lab uses single-cell mechanical properties as predictive biomarkers of ASC clonal differentiation capability. Elastic and viscoelastic properties of undifferentiated ASCs were tested via atomic force microscopy and correlated with lineage-specific metabolite production. Cell sorting simulations based on these "mechanical biomarkers" indicated they were predictive of differentiation capability and could be used to enrich for tissue-specific cells, which if implemented could dramatically improve the quality of regenerated tissues.

Neural network analysis identifies scaffold properties necessary for in vitro chondrogenesis in elastin-like polypeptide biopolymer scaffolds

The successful design of biomaterial scaffolds for articular cartilage tissue engineering requires an understanding of the impact of combinations of material formulation parameters on diverse and competing functional outcomes of biomaterial performance. This study sought to explore the use of a type of unsupervised artificial network, a self-organizing map, to identify relationships between scaffold formulation parameters (crosslink density, molecular weight, and concentration) and 11 such outcomes (including mechanical properties, matrix accumulation, metabolite usage and production, and histological appearance) for scaffolds formed from crosslinked elastin-like polypeptide (ELP) hydrogels. The artificial neural network recognized patterns in functional outcomes and provided a set of relationships between ELP formulation parameters and measured outcomes. Mapping resulted in the best mean separation amongst neurons for mechanical properties and pointed to crosslink density as the strongest predictor of most outcomes, followed by ELP concentration. The map also grouped formulations together that simultaneously resulted in the highest values for matrix production, greatest changes in metabolite consumption or production, and highest histological scores, indicating that the network was able to recognize patterns amongst diverse measurement outcomes. These results demonstrated the utility of artificial neural network tools for recognizing relationships in systems with competing parameters, toward the goal of optimizing and accelerating the design of biomaterial scaffolds for articular cartilage tissue engineering.