Unsupervised machine learning identifies latent pyrenoid states linked to mitotic remodeling defects and CO 2 -dependent growth
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Biomolecular condensates that persist through cell division must be reorganized and inherited, yet it remains unclear whether subtle defects before division are associated with later organelle or growth phenotypes. We examined the Chlamydomonas reinhardtii pyrenoid, a liquid-like condensate that concentrates ribulose-1,5-bisphosphate carboxylase/oxygenase (Rubisco), the photosynthetic CO 2 -fixing enzyme. As part of the algal CO 2 -concentrating mechanism, the pyrenoid raises CO 2 availability around Rubisco. We generated an RBCS1-mGold Rubisco reporter and developed an unsupervised image-analysis pipeline combining a convolutional autoencoder and a one-class support vector machine. Using 4,905 wild-type single-cell images, augmented 22-fold to 107,910 image instances, we defined the range of normal pyrenoid morphology. A combined machine-learning and visual screen of approximately 21,000 insertional mutants yielded 17 pyrenoid integrity mutants ( pim1 – pim17 ). Differential reconstruction-error maps highlighted local deviations from the wild-type reference, including phenotypes difficult to classify by eye. Four-dimensional live imaging showed defects in matrix dispersal, partitioning of Rubisco-containing foci, or pyrenoid recondensation in multiple pim strains. Growth assays identified broad defects and phenotypes that became more apparent as CO 2 supply decreased. Insertion-site mapping nominated candidate loci, including STT7 , which encodes a chloroplast kinase best known for regulating photosynthetic light harvesting. Independent STT7 -edited lines lacked detectable STT7 accumulation and showed pyrenoid-region reconstruction-error patterns, supporting an association between impaired STT7 function and altered pyrenoid morphology. These findings show that unsupervised image screening can extend forward genetics to subtle pyrenoid phenotypes accompanied by mitotic remodeling or growth defects.
Significance statement
Membraneless organelles lack enclosing membranes, yet their organization must be maintained or rebuilt as cells divide or conditions change. We developed an unsupervised machine-learning screen for the Chlamydomonas pyrenoid, a phase-separated photosynthetic condensate that supports CO 2 fixation. The screen recovered pyrenoid integrity mutants, including strains whose interphase pyrenoids appeared nearly normal but showed defects during mitotic remodeling or growth as CO 2 supply decreased. Static images thus helped identify strains with subsequent mitotic or growth phenotypes, and insertion-site mapping nominated associated candidate loci. The approach extends forward genetics to dynamic cellular compartments whose informative phenotypes are difficult to define in advance.
