Image-based deep learning reveals the responses of human motor neurons to stress and ALS

Although morphological attributes of cells and their substructures are recognized readouts of physiological or pathophysiological states, these have been relatively understudied in amyotrophic lateral sclerosis (ALS) research. In this study we integrate multichannel fluorescence high-content microscopy data with deep-learning imaging methods to reveal - directly from unsegmented images - novel neurite-associated morphological perturbations associated with (ALS-causing) VCP-mutant human motor neurons (MNs). Surprisingly, we reveal that previously unrecognized disease-relevant information is withheld in broadly used and often considered ‘generic’ biological markers of nuclei (DAPI) and neurons (βIII-tubulin). Additionally, we identify changes within the information content of ALS-related RNA binding protein (RBP) immunofluorescence imaging that is captured in VCP-mutant MN cultures. Furthermore, by analyzing MN cultures exposed to different extrinsic stressors, we show that heat stress recapitulates key aspects of ALS. Our study therefore reveals disease-relevant information contained in a range of both generic and more specific fluorescent markers, and establishes the use of image-based deep learning methods for rapid, automated and unbiased testing of biological hypotheses.


INTRODUCTION
While it is still unknown what drives pathological RBPs mislocalization and aggregation in ALS, alteration in liquid-liquid phase separation dynamics and 1 functions have been proposed to underlie this process [5][6][7][8][9] . RBPs are highly dynamic and have been shown to undergo changes in localization in response to various stressors [10][11][12][13][14][15][16] . Notably mitochondrial and oxidative stress are additional recognized and robust phenotypes in ALS pathogenesis in vitro 17 . The role of RBPs in ALS and cellular stress highlights that a diverse and complex interplay exists.
Cell shape and morphology are recognized readouts of a cell's physiological state or phenotype 18 . We previously reported common morphological descriptors that strongly discriminate ALS from control tissue at single cell resolution 19 , further indicating that key information related to cellular state might be contained in cellular shape in ALS. Dystrophic neurites are common pathological features in ALS and disrupted synapse formation have been shown in valosin-containing protein (VCP) mutant human induced pluripotent stem cell (iPSC) cultures of MNs 20 . Taken together, these studies suggest that the neuronal processes (collectively termed neurites or the 'neuritome') may be a good cellular subcompartment to reveal ALS pathomechanisms. However, neurites are challenging to study both in tissue sections (as the arborization of processes is largely lost during sectioning) and in vitro due to difficulty in accurate segmentation and association of neuronal processes with individual cells. Consequently, neuronal processes remain comparatively understudied in ALS and it is still unknown how and to what degree the neuritome is affected in ALS pathogenesis, whether ALS-related stress insults modify this compartment, or if cytoplasmic accumulation of RBPs in ALS MNs relates to other aberrant cellular phenotypes such as dystrophic neurites.
We previously generated a high-content imaging data-set of control and ALS-related VCP-mutant iPSC-derived MNs cultures co-labeled with a combination of three fluorescent markers, specifically: i) a nuclear-specific marker (DAPI), ii) a neuron-specific marker allowing to outline the neurites ( III-tubulin), and iii) an β antibody against one of five RBPs: TDP-43, SPFQ, FUS, heterogeneous nuclear ribonucleoprotein A1 (hnRNPA1) or heterogeneous nuclear ribonucleoprotein K (hnRNPK) 16 . In our previous study we specifically analyzed the spatiotemporal responses of the aforementioned ALS-related RBPs to different stressors (oxidative, heat and osmotic). Here we propose to apply deep-learning methods to this rich imaging data-set to test in an automated fashion 1) whether aberrant cellular morphological phenotypes, including neuronal processes, associate with ALS pathogenesis; 2) whether these morphological phenotypes correlate to aberrant ALS-related RBP phenotypes, and 3) whether extrinsic stress insults in control MN cultures can recapitulate ALS phenotypic changes. Deep learning models such as Convolutional Neural Networks (CNNs) are now widely used to efficiently perform image classification and image segmentation [21][22][23][24][25] . Such methods are able to analyze images without prior image segmentation, feature selection or human-directed training, and automatically extract features from raw data, removing significant bias from this process. Importantly CNN-based image classifier performance largely depends on whether sufficient information is contained in the provided set of images. DAPI and III-tubulin capture complementary and non-overlapping β information related to the nuclear shape and neuronal morphology including the neurite respectively. We propose that comparing the performance of different classifiers trained with iterative combinations of fluorescent images can be used to identify which cellular compartment or specific RBPs is most affected between any two given culture conditions. Additionally we propose that the similitude in phenotypes between different MNs culture conditions can be quantified using the trained model predictions. We demonstrate the utility of this approach, which enables the discovery of novel phenotypes in ALS MN cultures and the identification of the relevant extrinsic stress condition that best approximates ALS pathogenesis. The advantage of our method is that it is highly versatile and can quickly guide the scientist towards the most promising hypothesis for further experimental validation. By providing our fluorescence microscopy raw images together with open-source implementations of the methods and trained models, we aim to allow other researchers to readily apply these methods and test additional hypotheses. In summary, we propose the use of deep learning methods to leverage the power of large image data-bases from ALS-related MN cultures to automatically and rapidly generate testable biological hypotheses, a method that could prove transformational in promoting innovative research directions, diagnostics and therapies.

Repurposing image-based deep learning methods to test biological hypotheses
We previously studied the spatiotemporal responses of ALS-related RBPs to different stressors in control versus ALS-related VCP-mutant iPSC-derived electrically immature MN cultures using image-based analysis ( Fig. 1A and Supplementary Table S1) 16 . These cultures were immunolabeled after one hour of exposure to oxidative stress, heat stress and osmotic stress, along with recovery timepoints from heat stress (two hours) and osmotic stress (one, two and six hours).
Specifically a combination of three specific markers was used: a nuclear marker (DAPI), a neuronal marker allowing precise identification of neurites ( III-tubulin), β and an antibody against one of the  In particular we find that two hours after heat stress, MN cultures still exhibit high heat|TDP-43 and heat|SFPQ model predictions and lower (albeit still elevated) heat|FUS model prediction (Fig. 2E). The results indicate that the TDP-43-and SFPQ-related phenotypes are still present at this stage, and that the FUS-related phenotype is only partially resolved, partly reflecting on our previous study, where we did not detect reconstitution of nuclear TDP-43 and FUS to basal levels following 2 hours of recovery from heat stress 16 . Our previous study also revealed slower nuclear relocalization dynamics for TDP-43 and FUS after osmotic stress, with FUS exhibiting exceptionally aberrant nuclear-to-cytoplasmic distribution as long as 6 hours post-stress 16 . Here we find that TDP-43-related phenotype is fully resolved 2 hours after treatment while FUS-related phenotype is not resolved 6 hours after treatment (Fig. 2F). We also find delayed hnRNPK-related phenotype recovery. Notably we find that the recovery kinetics for most RBPs after both heat and osmotic stress correlate over time with the neuritome-related phenotype,

ALS
We previously reported common morphological descriptors that strongly discriminate ALS from control control tissue at the single cell level 19 , indicating that key information related to ALS cellular state might be contained in cellular shape.
Having found that our approach is suitable to reproduce prior findings related to stress in MNs, we next sought to test whether ALS-related VCP-mutant MNs are We next aimed to understand what information is used by these ALS classifiers to discriminate images from control and VCP-mutant MN cultures.
Integrated gradients (IG) is one popular approach for CNN model interpretation enabling the visualisation of the relevant pixels for a specific image that contribute to its classification 26 . Looking at the IGs of randomly selected images with high ALS|DAPI model predictions showed relevant pixels mostly overlapping with the outline of the nuclei, with some contribution from the pixels located inside the nuclei (Fig. 3D). This indicates that the ALS-related phenotype identified by the ALS|DAPI classifier primarily relates to the nuclear shape (including the size) rather than to other DAPI-related measurements such as texture or intensity. Next looking at the IGs of randomly selected images with high ALS|DAPI:BIII model predictions showed relevant pixels primarily located at the edges of the neurites, indicating that relevant information mostly arises from the outline of the neurites rather than 11 from the texture or the intensity of the III-tubulin immunolabeling (Fig. 3E) randomly selected images with high ALS|RBP model predictions indicated that the relevant pixels in all five ALS|RBPs classifiers are excluded from the nuclear areas as opposed to the most relevant pixels of the ALS|DAPI classifier that are most commonly localized at the inner nuclear membrane or inside the nucleus (Fig. 4B).
This demonstrates that the better the performance of the classifier, the less  Fig. 4A). These However, heat stress induced more than a 10% increase in ALS prediction across 5 out of 7 classifiers (Fig. 4C). ALS|TDP-43 classifier, which performs best in ALS MN classification, is indeed the unique model which leads to significant, however modest, model prediction across the three stress conditions (Fig. 4D and Supplementary Figs. 4B,C). Additionally we find that heat stress is the unique condition which leads to similarly high disease model prediction in control MN cultures and in untreated VCP-mutant MN cultures, given the ALS|FUS and ALS|hnRNPK classifiers (Fig. 4E and Supplementary Figs. 4B,C). Hierarchical clustering of the untreated ALS MN cultures together with the three stress conditions according to the effect size of each classifier (euclidean distance and Ward clustering) eventually confirmed that heat stress induces overall the most similar cellular changes to ALS (Fig. 4F). Altogether these results confirm that MNs  Here we find that iPSC-derived MNs exposed to heat stress, as opposed to

Compliance with ethical standards
Informed consent was obtained from all patients and control controls in this study.
Experimental protocols were all carried out according to approved regulations and guidelines by UCLH's National Hospital for Neurology and Neurosurgery and UCL's Institute of Neurology joint research ethics committee (09/0272). Cell culture, stress treatments, immunohistochemistry and image acquisition were performed as in 16 . Indeed these data are utilised in the current manuscript and no additional experiment was required.

High-content Imaging Dataset
The imaging dataset used in this study consists of fluorescence microscopy images of iPSC-derived motor neurons as previously reported 16 . The neurons either came from control cell lines or cell lines with the ALS-related VCP mutation and underwent experimentation after six days of terminal differentiation. Details of iPSC lines are provided in Table S1. To induce stress the cultures were subject to one hour of oxidative stress, one hour of osmotic stress and one hour of heat stress.
To examine recovery, the cultures were subject to one hour of stress and then returned to untreated conditions for two hours following heat stress and one, two and six hours following osmotic stress. Following stress treatments or recovery,  (Figs. 1A,B). The data-set will be deposited on IDR in a close future.

Image Pre-Processing
All images went through preprocessing steps described in Supplementary Figure 1.
Raw images are 16-bit images. 16-bit raw z-stack images (1080 x 1080 pixels) from the same field of view were first merged using Maximum Intensity Projection, where the pixel with maximum intensity across all z-stacks is selected at each location in the image. Following conversion of MIP images to 8-bit images, channels were merged together to form an RGB image. We created 13 types of RGB images, 20 either composed of one, two or three channels, to train image classifiers with 13 different combinations of immunostained images (Figure 1E). For images with three channels, DAPI was assigned to blue channel, III-tubulin to the red channel and β the RBP to green channel. For images with one or two channels, pitch-black images were assigned to the remaining channels so that the image would still be considered RGB. Images were then enhanced using Python Image Library Pillow from the ImageNet dataset. The last two steps were added in order to fulfill the requirements when using pre-trained models, which expect input images to be normalized in the same way as the dataset on which they were trained.

Data augmentation
In order to improve accuracy and reduce overfitting, we performed five augmentations on each image of the training set as follows and as previously

Model Explainability and Integrated Gradient
The integrated gradient (IG) is a widely used interpretability algorithm that allows to identify what pixels of an image have the strongest effect on the model's predicted class probabilities and therefore allowing to visualize which parts or the image are important for classification 26 , by computing the gradient of the model's prediction output to its input features. We used the Captum Insights method 55 to obtain the IG for randomly selected images associated with high classifier prediction scores.

Model prediction data analysis
We used R and lme4 56      for ALS|DAPI:BIII:RBPs model predictions.

Electronic supplementary material
Supplementary Tables 1-7 can be accessed here.
Table S1 | Description of human sample origin and mutations.