3 million cells per minute: parallel microdevice with AI-powered single-cell analysis

Many applications—from drug discovery and diagnostics to cell engineering and gene modulation—require delivering biomolecules into large numbers of cells and rapidly evaluating the outcomes. The challenge is twofold: achieve intracellular delivery at scale across diverse cells and cargos, and obtain quantitative results fast enough to keep pace with that delivery.

Researchers at the Indian Institute of Technology (IIT) Madras and Toyohashi University of Technology (TUT) have developed an integrated platform that advances both fronts simultaneously. The platform combines two modules: a massively parallel through‑hole cell‑squeezing mechanoporation device for high‑throughput intracellular delivery, and an automated single‑cell image‑cytometry pipeline built on Mask R‑CNN.

The device guides cells through an array of up to 62,000 tiny through-holes that are narrower than the cell diameter. A brief, gentle squeeze creates transient membrane pores that admit biomolecules into the cell interior, then reseal, allowing cells to recover. In validation studies, the team delivered gene‑silencing RNA (siRNA) and plasmid DNA across multiple cell types, including human gingival fibroblasts (hGFs), demonstrating broad utility for cell engineering and personalized therapies.

To keep up with the fast delivery, an automated analysis system uses the same microscope images labs already acquire. A deep‑learning model looks at those images and, in a single pass, reports four readouts: cell size, the fraction of cargo‑positive cells, the fraction of viable cells, and per‑cell fluorescence intensity. What used to take hours of manual counting now takes about 83 seconds on a representative dataset of roughly 500 cells, with accuracy comparable to human review. Large cohorts were processed—for example, 1,980 cells (6‑FAM siRNA) and 1,184 (EGFP plasmid). These sample sizes make the results statistically robust and turn high‑throughput delivery into high‑confidence decisions.

“Our goal was simple: get molecules inside many cells quickly and gently,” says first author Pulasta Chakrabarty. “Seeing the device work across different cell types points to real potential in cell engineering and personalized therapies.” Co‑author Ryoma Suzuki adds, “The automated model looks at the same images labs already use. It does the counting and measuring—cell size, delivery efficiency, viability, and per‑cell fluorescence—in one pass, so the evaluation keeps pace with the experiments.” “High throughput alone isn't enough,” notes Prof. Moeto Nagai. “What matters is the speed at which results can be trusted. By unifying high‑throughput delivery with automated quality control, we move from proofs‑of‑concept to practical workflows—and closer to systems that prepare a patient’s cells on‑site.”

By coupling scalable intracellular delivery with rapid, automated evaluation, the platform expands what can be accomplished in a single day—from large‑scale screening and diagnostics to practical cell manipulation and future point‑of‑care gene‑editing workflows.

Figure 1: End‑to‑end workflow: 2D cell squeezing for intracellular delivery and AI image cytometry for single‑cell readout

Top row (delivery and imaging). Cells and cargo are introduced into a microfluidic chip whose base integrates a porous SU‑8 membrane with vertical through‑holes and a PDMS chamber. As individual cells pass the sub‑cell‑sized through-holes, a brief 2D squeeze forms transient membrane pores, allowing cargo to enter; after passage, the membrane reseals, and cells exit to the outlets. Treated cells are then collected and imaged under a microscope (bright‑field and fluorescence).

Bottom row (automated single‑cell analysis). The same image set is processed by image cytometry using a Mask R‑CNN–based instance‑segmentation model: (i) bright‑field for cell localization, (ii) green fluorescence for cargo delivery, (iii) red fluorescence for viability/dying cells (e.g., PI or Calcein red‑orange), and (iv) instance‑segmentation overlays for per‑cell measurements. A simple rule-based map classifies each cell by the presence/absence of green and red signals into four states (delivered‑live, delivered‑dead, undelivered‑live, undelivered‑dead), enabling automatic calculation of delivery efficiency, viability, and per‑cell delivered amount. The diameter of each cell in an image is evaluated from the instance segmentation masks.

Figure 2: Intracellular delivery in human gingival fibroblasts (hGFs) using the 2D cell‑squeezing device.

(a) 6‑FAM siRNA; (b) EGFP plasmid. For each condition, (i) bright‑field to locate cells, (ii) green fluorescence reporting intracellular cargo, (iii) Calcein red‑orange as the live/dead readout, and (iv) image‑cytometry overlays showing per‑cell segmentation and delivery/viability class. Scale bar: 20 µm.

(c) Summary bars of delivery efficiency (green‑positive fraction) and cell viability (Calcein‑positive fraction), reported as mean ± SD across fields/replicates.

(d) Single‑cell scatter plots linking signal to state: (i) 6‑FAM siRNA (N=1,980) and (ii) EGFP plasmid (N=1,184). X‑axis, green fluorescence (a.u., delivered amount proxy); Y‑axis, Calcein red‑orange (a.u., viability proxy). These panels connect raw images to cohort metrics and per‑cell quantification in one workflow.

“High Throughput Intracellular Delivery Using a 2D Cell-Squeezing Mechanoporation Device and Its Analysis by a Deep Learning Model”

Pulasta Chakrabarty, Abinaya R, Ryoma Suzuki, Srikanth Vedantam, Suresh Rao, Moeto Nagai, Tuhin Subhra Santra

First published: 21 August 2025

https://doi.org/10.1002/adhm.202502472