Encoding Cell Phenotype from Label-Free Imaging Flow Cytometry with Unsupervised Deep Learning

Abstract

Phenotype characterization with single-cell resolution can enable deep and nuanced insights into microbiological systems. Currently, Flow Cytometry and Imaging Flow Cytometry (IFC) offer numerous advantages, but are marred by barriers to accessibility: (1) high instrument costs; (2) labor-intensive, technically demanding sample preparation; and (3) reliance on consumable reagents (i.e., fluorescent labels). To achieve phenotype characterization without these constraints, we evaluated the low-cost, low-input ARTiMiS IFC as a potential alternative instrument technology. To demonstrate this approach, we used intracellular lipid content in microalgae, an important phenotype for production of biofuels and high-value bioproducts, as the phenotype of interest. Variational Auto-Encoder (VAE) unsupervised deep learning methodology was implemented to encode phenotype variation from un-annotated training data. The VAE embeddings were compared with other label-free predictor modalities to evaluate the stability of VAE data encoding across replicates and its predictive power to estimate the target phenotype. The VAE embeddings were robust and consistent between culture batches, and yielded accurate, consistent predictions of the demonstration phenotype in a high-throughput, non-destructive, dye-free methodology. In this proof-of-concept study, we demonstrate that VAE-enabled ARTiMiS IFC may serve as a viable alternative for cell phenotype characterization while overcoming several of the key drawbacks of traditional high-fidelity techniques. SynopsisLabel-free Imaging Flow Cytometry data was processed by a Variational Auto-Encoder to accurately predict lipid content in microalgal cells.

Code and data availability

The supplied blocks describe a VAE-based label-free imaging flow cytometry study on microalgal lipid content, but contain no data availability, code deposit, or supplement statements. No public dataset, image collection, analysis code, or trained model with an authors' URL is mentioned; only cited prior work and the bX

No evidence-backed public reproduction asset is currently recorded.