Tomocube

Newsletter

Innovate the way we see life | Tomocube Inc. | April 2026

Pluripotent stem cell quality control (QC) has a paradox at its core.

The assays you trust most — RNA panels, antibody staining — destroy the very cells you want to preserve. Even the gentler alternatives, like fluorescence imaging or label-free microscopy, either damage cells through phototoxicity or lack the resolution and depth to reliably catch early, borderline cases.

A new study from the labs of Profs. YongKeun Park (Department of Physics) and Ki-Jun Yoon (Department of Biological Sciences) at KAIST — built on the HT-X1 system — changes that equation.

DeepHOPE — Deep-learning-guided Holotomography for Pluripotency Evaluation — scores pluripotency directly from 3D refractive index (RI) images of live colonies, with no labels, no fixation, and no waiting. It detects state transitions hours to days before molecular markers shift.

Non-invasive, real-time inference of cellular identity

DeepHOPE integrates two complementary strengths into a single, non-invasive pipeline:

• 3D Holotomography (HT) reconstructs the RI distribution of live human induced pluripotent stem cell (hiPSC) colonies at sub-micrometer resolution in 1–3 minutes per colony — revealing reproducible fine-scale structural and morphodynamic features.
• A deep neural network (DNN) learns to read those structural maps, distinguishing undifferentiated colonies from those entering early differentiation with up to 96.8% accuracy on a blind test dataset spanning all three germ-layer lineages.

Real-time selection of hiPSCs for personalized stem cell therapy

DeepHOPE held up across contexts far beyond its training data — from retinoic acid-induced differentiation to patient-derived hiPSC lines and nascent colonies emerging during somatic cell reprogramming — demonstrating generalizability that matters for real-world deployment.

That generalizability translated to a clinical payoff:

• Enhanced efficiency in midbrain dopaminergic neuron progenitor (mDAP) production for Parkinson's cell therapy — hiPSC lines that DeepHOPE rated highly produced more mDAPs and fewer off-target adrenergic and serotonergic cells.

Beyond classification, the study points to a mechanism:

• Out of 31 structural features extracted from 3D RI data, colony flattening — captured by aspect ratio and average thickness — emerged as a top predictor of pluripotency loss.
• Flattening preceded gene decline across every lineage tested.
• Driver: F-actin redistributes from apical to basal within 24 hours of induction.
• Low-dose blebbistatin alone triggered flattening — traction dropped in 15 min, DeepHOPE saw it within 1 hour, and 14 days reduced pluripotency.
• Cytoskeletal mechanics may sit upstream of pluripotency exit, placing the mechanical signal ahead of the molecular markers conventionally used to define the state.