Joint profiling of cell morphology and gene expression during in vitro neurodevelopment

We have performed a comprehensive characterization of iPSC-derived cortical neurons from a widely used protocol where we combined image-based readouts of cellular morphology with single-cell transcriptomics. We describe a proof-of-concept approach, where we applied CP to a heterogeneous, dynamic system – cortical neurodevelopment – and explored how the joint analysis of imaging and gene expression-based profiles of single cells can contribute to identifying new cellular phenotypes. While scRNA-seq alone allowed us to classify a diverse array of cell types produced in vitro, the dynamic nature of developing neurons involves alterations beyond the transcriptome such as in cellular size, shape, and complexity. In neuronal cell types, CP has previously been applied to iPSC-derived neuronal progenitors (McDiarmid et al., 2023) as a readout for drug treatment, rather than as a characterization approach. Application of CP in our study throughout the developmental trajectory offered novel insights on the dynamics of morphological readouts, as observed, for example, with mitochondrial channel intensity (Figure 3b, c).

Previous studies have attempted to link bulk gene expression and cell morphology assuming shared biological information between the two, and have found that changes in image-based features are associated with the expression of a subset of genes, often related to the cellular components and organelles stained in the assay (Haghighi et al., 2022; Nassiri and McCall, 2018). The generalizability of the CP assay (Willis et al., 2020) means that such direct links of gene expression to cellular components, originally observed in a human osteosarcoma cell line U2OS, are likely reproducible in different cell types. This hypothesis prompted us to link these two modalities in our neuronal system. While our sample size is small and limits our ability to generalize our findings beyond our specific set of differentiations, donor-specific variation and the large number of individual cells in our dataset boosted the capacity of the predictive model to link variation in gene expression to CP features. For example, we identified functional terms linked to CP features that recapitulate known or expected biology, such as those linked to ER, supporting our model’s performance. To enhance the interpretability of novel associations between image features and their potential biological function, we applied a linear model rather than a machine learning approach, as in previous cross-modality CP-based approaches (Way et al., 2021).

A common challenge in modeling neurodevelopment using pluripotent stem cells, regardless of the applied protocol, lies in determining whether the cell types generated in vitro accurately mirror the transcriptional signatures and developmental trajectories observed in vivo. Previous studies in organoids have linked in vitro-specific cell states to aberrant oxidative or glycolytic stress (Bhaduri et al., 2020; Uzquiano et al., 2022). These alterations in energy-associated pathways are frequently associated with the limitations of the culture media to provide essential nutrients to cells, especially within the necrotic cores of organoids (Uzquiano et al., 2022). In a recent study (He et al., 2024), transcriptomic differences between cells from human neural organoids (comprising 26 protocols) and developing human brain (first-trimester) (Braun et al., 2023) were associated with the upregulation of canonical glycolysis and mitochondrial ATP synthesis-coupled electron transport in organoids. Similarly, we observed that low-quality cells in our 2D culture system showed distinct metabolic and energetic states compared to high-quality cells of the same type. For instance, activation of glycolysis differed among progenitors, and activation in steroids biosynthesis among excitatory neurons. A fraction of these low-quality cells could be linked to an identity exclusively seen in organoids, most of them being either pan-neuronal or pan-radial glial. If not accounted for, mis-annotating these low-quality cells could potentially bias downstream findings in disease models.

The morphological profiling of low-quality cell types within CP data is currently not addressed in our study due to the model limitations. Replacing generic dyes with those specifically targeting neurons could improve cell type granularity (Laber et al., 2023), and increasing image magnification could better resolve individual organelles, potentially enhancing cell type identification by the assay. Further, to capture morphological heterogeneity, we analyzed CP readouts at a single-cell level instead of averaging CP features by well after cell segmentation as is often done, which came at the cost of increased noise (Caicedo et al., 2022). Although lower seeding density in culture may improve the accuracy of segmentation, we have previously observed that the viability of developing neurons is compromised in sparser culture conditions. It is also uncertain whether the 384-well format of the CP assay, as compared to 35 mm dishes used for scRNA-seq, impacts cell type production. In the future, new techniques that allow deriving CP readouts and transcriptomes from the same individual cell would offer a ground truth for model validation.

In general, a multi-modal perspective of any system offers insights that may not be visible from any one assay alone (Way et al., 2022). Here, CP revealed donor-specific ER changes in healthy donor lines corresponding to transcriptomic differences in ribosomal genes of inhibitory versus excitatory neurons. Understanding how interneuron production is altered at the level of both gene expression and cellular processes is key to uncovering the mechanisms implicated in NDDs, as suggested by recent findings implicating the ER and cytoskeleton in interneuron development and migration (Meng et al., 2023). In our model system, the 11 morphologically distinct CP clusters broadly separated progenitors from maturing neurons. In terms of cell type identification, single-cell transcriptomics remains a far more in-depth tool, particularly when cell types are closely related. However, the differences in cellular features between the various clusters suggest that CP provides information outside of the paradigm of cell types. To unravel what this biological signal might instead be capturing, we performed stratified LD score regression on the CP feature space, via the genes predicting them. While indirect, this analysis highlighted the disease relevance of morphological features in capturing heritability of brain-related traits, looking beyond cell types. Neuronal size, structure, and density have been previously shown to be impacted in various neurodevelopmental and psychiatric disorders (Kathuria et al., 2018; Purcell et al., 2023) and we show proof of concept that such changes can be captured with the CP assay.

To further validate the ability of CP to detect disease-relevant phenotypes, we applied our framework to neural progenitors derived from KS iPSC lines. Even without matched transcriptomic data, CP revealed morphologically distinct clusters enriched or depleted in KS samples, associated with altered cell cycle activity and accelerated neuronal differentiation – features previously reported in KS (Carosso et al., 2019) and other NDDs (Jhanji et al., 2024). While not performed here, such signatures could in principle be validated through orthogonal assays such as immunocytochemistry. In the future, multiplexing CP dyes with antibody-based quantification of neuronal markers could further enhance the utility of the assay in uncovering novel disease phenotypes. Nonetheless, these findings underscore the capacity of CP-derived features to capture subtle, disease-specific shifts in cell state. Moreover, linking morphological phenotypes to predicted transcriptional programs reinforced the biological relevance of these observations. Together, this highlights how image-based profiling can complement scRNA-seq by providing an orthogonal and scalable readout of developmental perturbations in both healthy and disease contexts.

In conclusion, we have performed an in-depth characterization of a widely used and highly disease-relevant cortical model using joint profiling of cell morphology and gene expression. We show that cell morphology profiling is able to capture disease-specific cellular phenotypes, complementary to those captured by scRNA-seq, and provide a novel framework for combining image-based phenotypes with transcriptomics in single cells. We highlight how joint molecular and morphological phenotyping can be widely applicable in disease modeling (Chandrasekaran et al., 2023; McDiarmid et al., 2023) and anticipate this approach to have great potential in linking genetic variants with biological processes underlying diseases.