NOVA

Cell painting and high-content screening usually reduce a microscopy image to hand-designed features — object counts, intensities, texture descriptors — computed after segmenting cells and organelles. That pipeline is brittle in neurons, whose axons and dendrites defeat border detection, and it can only measure what someone thought to define. The Neuronal Organellomics Vision Atlas (NOVA) takes the opposite approach: learn a representation of organelle topography — the joint localization and morphology pattern visible in an immunofluorescence channel — and quantify perturbation effects as distances in that latent space.

NOVA was built by Sagy Krispin, Nancy Yacovzada, Eran Hornstein, and colleagues at the Weizmann Institute of Science, with collaborators at Johns Hopkins University, the National Institute on Aging, the National Institute of Neurological Disorders and Stroke, and the University of Aberdeen. It is a vision transformer trained in two sequential stages, and it is paired with a statistical layer — mixed-effects meta-analysis over experimental repeats — that converts embedding distances into effect sizes with confidence intervals and multiplicity correction.

Applied to human iPSC-derived neurons carrying ALS-associated mutations, the framework placed more than 30 organelle markers on a common scale and surfaced a link between cytoplasmic mislocalized TDP-43 and processing bodies, which the authors then confirmed by independent imaging in patient-derived neurons and human neuropathology.

#Key Features

• Segmentation-free phenotyping: The model reads raw two-channel tiles and never requires cell borders, organelle detection, or a predefined feature list.
• Two-stage training: A supervised localization-encoding stage teaches the encoder organellar topography; a contrastive stage adapts it to perturbation effects while preserving that knowledge through selective layer freezing.
• Repeat-aware contrastive sampling: Positives and negatives are drawn from different wells and experimental repeats, so the model is pushed toward reproducible perturbation signal rather than plate or differentiation artifacts.
• One scale for every organelle: Effect size is the latent-space distance between perturbed and baseline states, computed per repeat and pooled by mixed-effects meta-analysis, yielding a ranked forest plot across all organelles studied.
• Generalizes to unseen conditions: The framework was applied to cell lines and organelles absent from training, including patient-derived motor neurons.

#Technical Details

The first stage trains a ViT from scratch on the OpenCell dataset — 1,100,253 cropped images covering 1,311 endogenously tagged human proteins in HEK293T cells — with a linear head and cross-entropy loss over protein identity, reaching 87% mean accuracy on organelle localization categories after 44 epochs. Configuration is compact: 192-dimensional embeddings, patch size 14, two input channels (target marker plus nucleus), and 100×100 pixel tiles.

The second stage initializes from those weights and switches to a supervised contrastive objective. An angle metric compares each layer of the fine-tuned model against its pretrained counterpart; layers deviating less than the median are frozen, and the rest are retrained. The linear head is replaced by a 192→768→384→128 multilayer projection. Each training step draws an anchor tile, one positive from the same organelle and perturbation, and five negatives from the same organelle under different perturbations, all from different wells and repeats, and applies InfoNCE loss. This stage used roughly 3.2 million images of perturbed cultured human neurons generated in-house, part of an 11.5-million-image corpus spanning more than 30 organelle markers.

Downstream, effect sizes are estimated per perturbation × organelle × repeat and combined in a mixed-effects meta-analysis modeling within-repeat sampling variance and between-repeat heterogeneity. In a sodium arsenite oxidative stress experiment across three repeats and over 590,600 images, stress granules showed the largest combined effect size; in a FUS R495* line across six repeats, 29 organelles were analyzed against both a genetically corrected revertant and a wild-type control.

#Applications

The framework fits laboratories running arrayed imaging screens where the readout is "did this perturbation change anything, and where" rather than a single predefined phenotype: chemical and genetic screens in iPSC-derived neurons, disease-variant characterization, and comparisons across patient cohorts. Because the statistical layer models experimental repeats explicitly, it is well matched to iPSC work, where differentiation batch is a dominant source of variance and each repeat is effectively a separate study.

#Impact

NOVA's contribution is the pairing of a learned organelle representation with a meta-analytic framework rigorous enough that a discovery made in latent space — the TDP-43/P-body relationship — could be carried into orthogonal validation, including FRAP measurements of P-body dynamics and human post-mortem neuropathology. It offers a template for turning representation learning into reportable effect sizes rather than UMAP plots. The work is a preprint awaiting peer review, and the preprint is posted with all rights reserved. Training and analysis code is public on GitHub, the pretrained checkpoint is downloadable from Hugging Face, and the image corpus is hosted in a public S3 bucket, but the repository carries no license file, which leaves reuse terms for the code and weights unsettled.