Learning biophysical determinants of cell fate with deep neural networks

Abstract

Deep learning is now a powerful tool in microscopy data analysis, and is routinely used for image processing applications such as segmentation and denoising. However, it has rarely been used to directly learn mechanistic models of a biological system, owing to the complexity of the internal representations. Here, we develop an end-to-end machine learning approach capable of learning an explainable model of a complex biological phenomenon, cell competition, directly from a large corpus of time-lapse microscopy data. Cell competition is a quality control mechanism that eliminates unfit cells from a tissue and during which cell fate is thought to be determined by the local cellular neighborhood over time. To investigate this, we developed a new approach (τ -VAE) by coupling a probabilistic encoder to a temporal convolution network to predict the fate of each cell in an epithelium. Using the τ -VAE's latent representation of the local tissue organization and the flow of information in the network, we decode the physical parameters responsible for correct prediction of fate in cell competition. Remarkably, the model autonomously learns that cell density is the single most important factor in predicting cell fate -a conclusion that is in agreement with our current understanding from over a decade of scientific research. Finally, to test the learned internal representation, we challenge the network with experiments performed in the presence of drugs that block signalling pathways involved in competition. We present a novel discriminator network that, using the predictions of the τ -VAE, can identify conditions which deviate from the normal behaviour, paving the way for automated, mechanism-aware drug screening.

3D single-cell shape analysis using geometric deep learning

“…On average, cells treated with Nocodazole had low scores of PC 1, while cells treated with Blebbistatin had high scores of PC 1. Similar to methods in [25], sampling cell shapes from the dataset with a range of values for different principal components visually confirmed the correlations with classical features (Figure 4 F-H shows this for PC 1, PC 4, and PC 10 as these were the drivers of the classification between Blebbistatin and Nocodazole. To assess the relationship between classical features and the PC of the DL features, we fit local regression curves using weighted linear least squares regression (lowess) (Figure 4 F-H).…”

Learning dynamic image representations for self-supervised cell cycle annotation

“…cell fate determination, signalling pathways, morphogen gradients, etc. (Soelistyo et al, 2022;Shakarchy et al, 2023;Gallusser et al, 2023), in biology and beyond.…”

“…Previous analyses addressing the morphodynamic transition challenges have sourced from latent image representations, however, these methods either rely on compressing the entire trajectory with variable durations into single, fixedlength embedding (Wu et al, 2022), or truncate the timeresolved tracks to an identical, pre-defined length (Soelistyo et al, 2022). Conversely, the length-invariant approaches are either driven by hand-crafted features (El-Labban et al, 2014) or yield poorly interpretable continuous latent space clustering (Ji et al, 2017), which complicates discrete categorical cell cycle phase annotation (Rappez et al, 2020).…”

Visual interpretability of image-based classification models by generative latent space disentanglement applied to in vitro fertilization

“…Interpretability of image-based classification models is absolutely necessary in biomedical domains where mechanistic understanding and transparency are crucial. Established attribution-based (Barnett et al 2021, Kraus et al 2017, Graziani et al 2018, Wu et al 2018, Singh et al 2020, Zhang et al 2021) or counterfactual-explanation based ( Singla et al 2023 , Thiagarajan et al 2022, Mertes et al 2022, Narayanaswamy et al 2020, Soelistyo et al 2022, Zaritsky et al 2021, Lamiable et al 2023, Kraus et al 2017) methods were applied, out-of the box or after some adaptations, to interpret a variety of biomedical image-based classification tasks. DISCOVER’s classification-driven and disentanglement representations overcome the inherent limitations in these methods and enabled us to quantitatively confirm non-trivial interpretations, rather than relying on qualitative explanations of representative images, and to systematically perform quantitative instance-specific interpretations.…”