Single-cell data at scale
Where do batch effects actually come from?
The most manual step in the lab1 / 8
Label-free, scripted prep for million-cell datasets, on the robot you already run.
Virtual cell models, perturbation atlases and foundation-scale datasets are built from thousands of samples, and every one arrives with dead cells and debris. Clean them up by hand and the dataset carries biases from the operator, the day and the centrifuge. The LeviCell 96 System removes them with a gentle magnetic field instead, scripted on your liquid handler, identically in every well.
One SALT array: 96 wells, 15,000 to 500,000 cells each, 5 minutes hands-on, 45 minutes from loading to enriched cells.
See it in two minutes
Dead cells and debris out, live cells in one band, the same treatment in every well, 96 samples at a time.
Building a dataset in the thousands of samples?
Book a workflow review2:14 · Sound on · Captions built in · Open on YouTube
Downstream software can regress out a batch effect it can measure. It cannot recover a sample that was never viable, and it cannot see the variability that manual prep adds before the first barcode.
Spins, strainers and bead kits are hands-on steps with hands-on variance. Across a hundred-million-cell program, that variance is a feature the model will learn.
Thaw, electroporation, lipofection, CRISPR editing, drug treatment and dissociation all leave dead and dying cells behind. Their RNA becomes ambient background in every barcode; their fragments become doublets and debris.
Most prep steps leave no record. When a cluster looks odd eighteen months later, nobody can say what the input looked like that day.
Every complex workflow stresses cells: thawing, dissociation, electroporation, editing, sorting, waiting. Stressed cells die, leak RNA into every barcode, and switch on programs that look like biology. The literature is unambiguous: the health of the input sets the ceiling on the quality of the data.
A dying cell lyses, and its RNA is captured in every droplet or well around it. The dirtier the input, the more of your reads describe soup instead of cells.
Handling switches on immediate-early and heat-shock programs within minutes. The genes are real, the biology is not, and no amount of sequencing depth can tell the difference.
In a screen, the perturbation should be the only variable that changed. Damaged cells add a second one, and the model learns both.
Cell-free RNA from lysed cells is captured alongside every cell; contamination is ubiquitous and increases in necrotic or complex samples. Left uncorrected, ambient transcripts show up as differentially expressed genes and spurious pathways in the wrong cell types.
Young & Behjati, GigaScience 2020 (SoupX); Arora et al., PLoS One 2025; Cargnelli et al., Front. Bioinform. 2026A core set of 512 immediate-early, heat-shock and stress genes (FOS, JUN and others) is induced by standard dissociation across tumor types, cell lines and PDX models, and a transcriptomically dying subpopulation appears with it. Cells that should look resting appear activated; stress-altered signals accounted for 60 to 64% of artifact signal in brain and bone-marrow datasets.
O'Flanagan et al., Genome Biology 2019 (155,165 cells); van den Brink et al., Nature Methods 2017; stress-altered transcription study, 2026Including broken or dying cells "inevitably affects data interpretation"; high mitochondrial read fractions mark them, and more than 20 features are needed to catch them reliably after the fact. The reliable fix is upstream: don't load them.
Ilicic et al., Genome Biology 2016Non-viable and dying cells increase ambient RNA and clumping, cause chip clogs and wetting failures, and lower sensitivity. 10x Genomics recommends dead-cell clean-up below about 80% viability, and many cores refuse samples under 70%.
10x Genomics cell preparation guide and technical note CG000093; core-facility guidelines (NGI Sweden, Technion)In a 12-month PBMC cryopreservation study, viability by dye looked stable, yet the number of cells captured in scRNA-seq fell by about 32%: damaged cells that count as live still fail to become data.
Li et al., Frontiers in Immunology 2025Zero-shot evaluation of single-cell foundation models found their embeddings "primarily driven by batch effects," and scGPT's own model card notes that pretraining does not mitigate technical variation. Whatever varies between samples at the bench becomes structure in the model.
Kedzierska et al., Genome Biology 2025; scGPT model cardPerformance figures from LevitasBio specifications, the Discovery Life Sciences study (16 cryopreserved tumor samples; 10 sequenced as paired Parse Evercode WT libraries) and Remsik et al., Nature 2025. Literature statements are from the cited publications, linked under each finding.
The literature says healthy input decides data quality. Here is what levitation has done to that input, in published work, in customer labs and in our own application data. Every number below has a named source.
"We had tried FACS repeatedly and were never able to recover this fragile cell type. The LeviCell EOS's gentler approach was what finally made it possible."
Ronan Chaligné, PhDDirector, Single-cell Analytics Innovation Lab, MSKCC; co-author, Remsik et al., Nature 2025"Without LeviCell, viability at transfection was so low the cells simply didn't survive. With it, even with the viability drop you normally see after transfection, I had material to work with for my downstream analysis."
Erika Castillo, PhDResearch Scientist, Condello Lab, UCSF"After lorlatinib the surviving cells are incredibly fragile; over half the material is debris and dying cells, and standard workflows like FACS or replating just collapse the culture. With LeviCell we recover intact, viable survivors, so what used to take two to three months now shows up in about 48 hours."
Nate Merrill, PhDResearch Assistant Professor, University of MichiganMouse leptomeningeal immune cells were stained with a 198-plex CITE-seq panel, cleaned with LeviCell and loaded on 10x: 3.15% mean mitochondrial reads across the dataset, 85% of 54,781 detected cells retained as high quality, and no population co-expressing JUN and FOS, the two genes that rise fastest under handling stress. The fragile dendritic cells and neutrophils the paper turned on survived to be sequenced.
Remsik et al., Nature 2025 (doi 10.1038/s41586-025-09012-z); figures from the preprint Methods and the authors' GenomeWeb webinarOn low-viability gut tissue and epithelial organoids where flow sorting had stalled a project, label-free levitation with CD45 depletion gave up to five times more high-quality cells (about 2× on organoids), with lower mitochondrial content and less ambient RNA even among cells that already passed QC, and all cell types preserved.
A. Antanaviciute, University of Oxford, public GEN webinar 2023; published use in McGregor et al., Nature 2026Fresh iPSC-derived HSPCs rose from 60–75% to 91–96% viability in every paired sample, with about 85% of live cells recovered, which made nucleofection and gene editing possible where the cells had not previously survived. Clearing dead cells also kept differentiating cells from drifting toward death or activation.
E. Castillo, Condello Lab, UCSF; customer-reported, quote approvedThawed primary leukemia samples as low as 9% viable by flow; 22 paired samples levitated in one day, every one above 50% and most above 80%. A BCL-2 knockout culture that looked entirely dead yielded a viable band for study, and levitated cells engrafted mice from 42,000–100,000 cells instead of the usual 1–2 million.
E. Diaz-Flores, UCSF; levitation in the Methods of Pariury et al., Haematologica 2023; recovery figures customer-reportedNasal swabs that had clogged the 10x ran successfully after levitation, with more quality-filtered cells and unchanged cell-type composition; levitation was used on 58 of 132 samples in the resulting pediatric study. In plant protoplasts, levitation retained 3.5× more large (>20 µm) cells than flow-based clean-up, so the atlas stayed representative.
J. Ordovás-Montañés, Boston Children's Hospital (Walsh et al., bioRxiv 2026); C. Grones, Wageningen University, plant webinar and three publicationsLevitation uses less than 1 psi of pressure and binds nothing. Macrophages, among the most reactive cell types, showed no change in activation state after processing; a LeviPrep-LeviCell brain workflow delivered homeostatic microglia with 2–6× lower activation markers than alternative isolation methods; neutrophils and mast cells, which react to surface binding, come through closer to their native state.
LevitasBio application notes and research snapshots; M. Galeas-Pena, TulaneEnriching bladder cancer cells before a standard 10x run gave 3× the cells detected and 6× the average genes per cell. Lung DTC viability rose from 32% to 87% as dead cells fell from 67% to 12%. In a head-to-head on a split ovarian sample, levitation beat a bead-based dead-cell kit on both yield and viability.
LevitasBio research study and application notesDiscovery Life Sciences levitated 106 cryopreserved samples across 21 indications, with LevitasBio involved in neither processing nor analysis: samples above the 70% viability bar went from 29 to 92, average viability from 61% to 80%, and 105 of 106 improved, on inputs from 5,500 to 2.9 million cells. A London core enriched 130 frozen PBMC samples in 13 pooled runs, lifting viability from 60–70% to 90%.
LevitasBio–DLS case study 90-00123C (2023); Imperial BRC Genomics Facility research snapshot 90-00284A (2024)Levitation appears in more than 15 peer-reviewed publications, including Nature (2025, 2026), Nature Communications (2024, 2025), Nature Microbiology, Cell Reports (2025, 2026), PNAS, JCI Insight and EMBO Journal, across human tumors, immune tissue, kidney, heart, brain and plants.
levitasbio.com/publicationsCustomer figures are reported by the laboratories named and attributed to them, not presented as LevitasBio specifications. Quotes used with permission.
Levitation separates cells by their physical properties, so it does not care how the damage happened. It belongs wherever cells have just been thawed, transfected, edited, treated or dissociated, and right before they are fixed, barcoded or loaded.
Half the cells can be dead at thaw, and the debris is in the tube with them. Enrich live cells at thaw instead of waiting for a recovery culture that adds time and variance.
Measured on cryopreserved tumor DTCs: viability 52% → 75%; sequencing-ready samples from 1 in 16 → 10 in 16.Electroporation and lipofection kill a large share of cells in the first days, and the dying cells and fragments are still there when you harvest. Remove them label-free, right before fixation or barcoding, with no spins on fragile cells.
Timing: enrich at harvest, not after a recovery step that selects for the fittest clones.Dead cells and debris inflate ambient RNA and doublets, and both corrupt guide assignment and perturbation calls. Clean input makes the perturbation the only variable that changed.
Measured: 21% less ambient RNA and six-fold fewer doublets after one run, lower in every cell population.Cytotoxic conditions produce apoptotic fractions by design. Enrich the surviving cells so their transcriptomes describe the response, or collect defined fractions when death is the readout.
Label-free, so no antibody or dye biases which cells are kept.Differentiated cells and dissociated organoids are fragile and carry dissociation debris. A passive magnetic field, with no shear, centrifugation or filtering, keeps them intact and their transient states unchanged.
Cells, nuclei, organoids and spheroids; fresh, thawed or fixed.Nuclei preps carry cytoplasmic debris and clumps that read as background in every barcode. Clean up the nuclei suspension on the same array: debris and aggregates removed, nuclei collected ready to load.
Label-free clean-up of nuclei preps; no spins, no filters.The study figures above are from cryopreserved dissociated tumor cells, the hardest common input. The separation is physical and label-free, so the same step applies to any suspension; the 96-sample standardization run measures it on your cell type.
Cells, nuclei, organoids and spheroids, and particles. Fresh, thawed and fixed-cell workflows. Same chemistry from the LeviCell 12 to the LeviCell 96, so methods developed at small scale transfer without re-optimizing. The LeviCell 96 System ships with the INTEGRA liquid handler; the SALT array uses the SLAS/SBS 96-well footprint.
Where batch effects come from, why software can't fix them after the fact, and what a scripted, label-free prep step does about it. Eight slides; they advance on their own, or tap through.
Every one of these is a hands-on step with hands-on variance. Across a hundred-million-cell program, that variance is in the training set, and the model will learn it as biology.
Their RNA leaks into every cell's barcode as ambient RNA. Their fragments become doublets and debris. In the Discovery Life Sciences study, the most contaminated tumor sample spent 68% of its reads on soup, not cells.
You cannot recover a sample that was never viable, and you cannot see the variance that entered before the first barcode. The fix has to be physical, and it has to be the same every time.
A scripted protocol does the same thing to every well and leaves a run log behind. Pair it with viability counts before and after, and every sample in the dataset comes with a record of how it was prepared, and an answer when a cluster looks odd eighteen months later.
Building a perturbation atlas, a virtual cell training set or a screening program in the thousands of samples? We'll map the LeviCell 96 onto your robot. One page back, fifteen minutes, free.
Discovery Life Sciences ran 16 cryopreserved dissociated tumor samples, the input most likely to arrive half dead, through one levitation step, then sequenced ten of them with the Parse Evercode WT assay, untouched and levitated.
110of 16
Median viability 52% to 75%; 89% of dead cells removed from a typical sample; 21% less ambient RNA and six times fewer doublets in the paired libraries; 85 of 100 marker genes carried over, so the biology came through unchanged. Ten times the datasets from the same tissue.
Full results, charts and cost model
"We successfully integrated LevitasBio's Levitation Technology into our workflows for a variety of single cell approaches. Experiments with very sensitive cell types, or samples that are not suitable for cell sorting, benefit from this gentle, fast method to exclude dead cells and debris."
Stefan Günther, PhDHead of Next Gen Sequencing Facility, Max Planck Institute"The LeviCell EOS system allows our lab to process several low-viability samples simultaneously into high-viability samples, which is absolutely critical for single cell sequencing analyses."
Greg Golden, PhDPostdoctoral Researcher, University of PennsylvaniaThe Discovery study is 16 samples at bench scale and 10 at sequencing scale, with a confirmatory same-run study in progress. Batch-effect reduction across a multi-month program is the expected consequence of identical, scripted processing; it is what the 96-sample standardization run below is designed to measure on your samples, not something we ask you to take on faith.
Pick the rung that matches where your program is. Each one produces something you can put in front of your team.
Tell us your workflow, sample types and throughput, and whether you run a liquid handler. We send back a one-page plan: where levitation fits, timing per plate and the protocols that apply. Free, no instrument required.
Request a workflow reviewSend 96 samples to us, or run them at a partner service lab. You get viability and debris counts for every sample before and after, as in the Discovery study, and a side-by-side against your current prep on the same material.
Plan a runDevelop methods on a LeviCell 12, then move to the LeviCell 96 System without re-optimizing. Flexible purchase options, so scaling up is a schedule decision, not a budget cliff.
Talk to an automation specialistAn illustrative model you can run on your own numbers. It uses this study's viability gain and your costs; replace any input with your own quotes.
0
| To reach the target | Current clean-up | With LeviCell 96 |
|---|---|---|
| Samples to prep and sequence | 0 | 0 |
| Usable cells per sample | 0 | 0 |
| Clean-up cost | $0 | $0 |
| Downstream cost | $0 | $0 |
| Total | $0 | $0 |
| Cost per usable cell | $0 | $0 |
Book a workflow review, plan a 96-sample run, get your cost model or just ask a question. One short form goes straight to the LevitasBio sales team.
Only your name, email and organization are required. Details about your robot and samples help us come back with something specific.
Prefer email? Write to sales@levitasbio.com.
Sharing this with your team? Download the one-pager (PDF).
If you're building a perturbation atlas, a virtual cell training set or a screening program that runs into the thousands of samples, we'd like to map LeviCell 96 onto your robot. It takes fifteen minutes and costs nothing.