Planar Laser Activated Neuronal Scanning (PLANS) System for in vivo Flow Cytometry

Abstract

We present a light-sheet flow cytometer for screening of C. elegans. A machine learning approach is utilized to enable real-time analysis of protein aggregation models.

Alzheimer’s, Huntington’s, and ALS disease pathologies are associated with disruption of protein-folding homeostasis causing progressive neuron degeneration and dysfunction throughout the nervous system. Due to the large number and complex arrangement of neurons in mammals, it is difficult to understand how system-level functions and cellular physiologies change during disease progression. Small model organisms such as Caenorhabditis elegans, Drosophila, and zebrafish provide a testbed to recapitulate human degenerative diseases. Specifically, the nematode C. elegans provides many advantages as a genetic model to delineate age-associated protein-misfolding and formation of toxic aggregates in vivo. This small model exhibits a simple nervous system (302 neurons), stereotype cellular architecture, short life cycle (2-3 weeks), and is easily accessible for fluorescence imaging throughout their development and lifespan [1]. Current flow cytometer analysis methods, such as COPAS Biosorter, can analyze and sort C. elegans with low resolution (20-30 μm) and speeds up to 100 animals per second [2, 3]. While three-dimensional high-resolution fluorescence imaging flow cytometry (Line Excitation Array Detection – LEAD microscopy) has recently become available [4], some models may not require volumetric imaging to resolve individual neurons located on the ventral cord or protein aggregates within the 95 body wall muscle cells. To provide efficient screening with higher resolution than COPAS systems, we developed a microfluidic device and light sheet optical method, Planar Laser Activated Neuronal Scanning (PLANS), which is supported by machine learning algorithms for rapid phenotyping.

Statistical analysis of each drug compound requires rapid interrogation of a large number of C. elegans in a given population, which, due to the time-consuming nature of manual manipulation and anesthetic based on immobilization processes, requires a system to streamline data acquisition and analysis ideally without a need for anesthetics. To enable high resolution screening, microfluidics have been commonly used for manipulation and immobilization of worms [5]. To enable anesthetic free analysis, we developed a microfluidic flow channel for PLANS with a 40 × 40 μm² cross-section, which allows sequential flow of young adult C. elegans through a tailored light sheet to record one-dimensional (1D) fluorescence signal from neurons or protein aggregates (Fig. 1). Here, we demonstrate 3 μm scanning resolution of fluorescence signal from C. elegans flowing at speeds up to 1 millisecond per animal.

Fig 1:

(a) PLANS optical diagram. A 488 nm wavelength continuous wave laser is transmitted through an optical fiber, collimated, and shaped with a cylindrical lens. The shaped beam is focused by a 10×, 0.3 NA objective to form a light sheet tailored to cover the entire cross-section of the flow channel of the microfluidic device. Emitted fluorescence is collected by a second objective and incident upon a single PMT. (b) Microfluidic device design for flowing C. elegans. (c) Expanded view of the light sheet at the focus. The thickness of the light sheet is designed for 3-4 μm 1/e² diameter. Representative PLANS PMT signals (black lines) for polyQ24 (d) and polyQ35 (e) adults as compared to wide-field fluorescence images.

The light-sheet excites a cross-section of each worm as they flow through the microfluidic flow channel, producing an integrated fluorescence signal along the length of each animal. While conventional fluorescence methods such as confocal microscopy provide 3D spatial information, image acquisition is time consuming and computationally expensive, and requires lengthy immobilization methods. Our PLANS flow cytometer enables rapid scanning of 1D data, which allows for real-time analysis. We demonstrate the drug screening capabilities of PLANS with a polyglutamine aggregation (polyQ) model in young adult C. elegans. In this stage, the disease phenotype manifests as age-dependent aggregation of YFP-labeled protein in the body wall muscle cells. Aggregates are distributed along the length of the animal and different states of aggregation are resolved with machine learning data analysis.

To efficiently automate the data processing aspect of PLANS, we must meet two goals: autonomous detection of when an animal enters/exits the light sheet and health status determination of identified C. elegans; here efficient specifies a method that acts on the order of ten to twenty milliseconds such that the processing is near real time. Real-time animal detection can be easily performed via hard coded thresholds as the fluorescence signal of C. elegans in the light sheet far exceeds the variations presented by noise. Multiscale decomposition or Fourier classification techniques are slow, requiring tens to hundreds of milliseconds to classify a single C. elegans. As an alternative to these methods, we use convolutional autoencoders (CAE’s) to enable real-time analysis of PLANS data. These autoencoders have found significant success in a variety of biomedical imaging modalities for problems of denoising, segmentation, and feature learning [6, 7]. The appeal of CAE’s comes from their unsupervised nature and that the weights within the network can be optimized in the following manner:

$$argmi{n}_{\theta ,\varphi }{\Vert {\mathrm{g}}_{\varphi }\left(f_{\theta }(x)\right)-x\Vert }_2^2.$$ (1)

Here, f is the encoder with network weights θ, g is the decoder with network weights φ, and x is the input data. The output of the encoder is often labeled as a latent space or bottleneck, as it is a significantly compressed representation of the input data that can be used for tasks of clustering or classification. The main idea of any autoencoder is that the decoder reconstruction forces latent space to be relevant to the underlying signal; a linear autoencoder converges closely to PCA for example. Despite the ease of this formulation, a relevant latent space does not necessarily lead to a useful metric space where health states are nicely separated. For this reason, we optimize our network according to the following model that draws upon work done for natural image clustering [8]:

$$argmi{n}_{\theta ,\varphi ,\mathrm{v}}{\Vert {\mathrm{g}}_{\varphi }\left(f_{\theta }(x)\right)-x\Vert }_2^2+\lambda {\sum }_{i=0}^{C-1}{y}_i\max (\mid f_{\theta }(x)-v_i\mid -\zeta ,0)+\delta {\sum }_{i=0}^{C-1}{\sum }_{j=0}^{C-1}\max (\xi -\mid v_i-v_j\mid ,0).$$ (2)

This new formulation requires labels be present for all C classes and that every class has an updatable vector v_i of equal size to the latent space. Given that y_i indicates x belongs to the same class as v_i, the second term penalizes a metric space where the distance from a class vector v_i and x′s latent space is greater than margin ζ. The last term penalizes class vectors that are not at least a margin ξ apart. The variables ζ and ξ are constants set at time of training where λ and δ control the degree of regularization. This regularized CAE aides in generation of a metric space that is appropriate for classifying C. elegans health states in under twenty milliseconds per animal.

In summary, PLANS enables screening of model organism C. elegans with real-time analysis. Using a multiplexer system [9], we can deliver individual populations to the PLANS microfluidic chip every few seconds, enabling large-scale studies of system-level effects of age-associated degenerative diseases.