SCSOK 2026: Criteria for Defining Cell States
Hi, I’m Kakyung Kim, a bioinformatics researcher.
Over three days at SCSOK 2026, many talks asked what defines a cell state in single-cell analysis and how far that definition can be trusted. Changing only the random seed changes the number of clusters, T cells with the same phenotype see different antigens, and proteins can appear in a different place from their RNA.
The 2026 SCSOK & KSBMB Single-Cell Society Joint Symposium took place at the Sheraton Incheon Hotel from October 14 to 16. The IBS Center for Genome Engineering and the IBS Korea Virus Research Institute were among the co-organizers, and over the three days I sat through the keynotes and invited talks of the main program in order.
The sessions were organized into stem cells and development, immune diseases, AI and algorithms, multi-omics, and cancer, and this post groups the three days of talks around one question: what defines a cell state, and how far can that definition be trusted?
Spatial Motifs and Lineage Inference in Single-Cell Analysis
The keynote on the second day was “Computing through Space and Time in Single Cell Biology” by Prof. Junhyong Kim (University of Pennsylvania). He had also given a keynote on dendritic transcriptomes at the KOGO international conference in September, so I could hear the two talks back to back (KOGO 2026, last day). If the KOGO talk moved from measurement design to an evolutionary hypothesis, this one dealt with analytical methodology itself.
The talk started from a short history of the field and the sparsity of the data. Around 2006, about 40 single-cell transcripts could be handled, and, as he also noted at KOGO, more than 60% of transcripts are still present at fewer than 10 molecules per cell. In the meantime, the number of single-cell analysis tools has passed 1,000 (Zappia and Theis, Genome Biology 2021), and he called this abundance of tools a major problem in itself.
On the spatial side, each cell and its neighbors are built into a graph and embedded with a variational graph auto-encoder (VGAE). Densely packed subsets in the latent space were treated as spatial motifs that recur across the tissue. Even within the same tumor cell type, the transcriptome differed depending on which motif the cell belonged to. He openly acknowledged that the VGAE does not reconstruct the cell state itself, and added that finding motifs properly requires better distance properties.
The question on the temporal side was cell lineage. Cells of the same molecular type do not necessarily come from the same mother cell, so he asked whether lineage can be inferred backward from the transcriptome alone. CellTreeQM (Kuang et al., ICML 2025) splits the lineage tree into quartets of four cells and uses the four-point condition that a tree must satisfy as its loss function. Even partial knowledge of the lineage yields many quartets as supervision: in a binary tree, knowing only that cells split into two sets makes about 37% of possible quartets usable for training, and with four sets about 70%.
His conclusion was cautious. Geometric inference alone is not enough; the modular structure of biological function has to be learned. That requires vast amounts of context data, which are hard to obtain, so he argued for an inductive bias that reflects modularity. In the Q&A he pointed out that image-processing noise and cell segmentation problems in spatial data are ignored in most analyses.
Cell States That Shift with Analysis Choices
The afternoon AI and algorithms session on day two picked up this point with concrete tools.
Prof. Jae Kyoung Kim (KAIST, IBS) began with the problem that the values users set in scRNA-seq analysis, namely the number of principal components, the clustering resolution and random seed, and the batch correction method, change the results. On a public lung tissue dataset, he showed an example where changing only the seed gave 28 clusters in one run and 27 in another.
scLENS (Kim et al., Nature Communications 2024) uses random matrix theory to count only the components outside the noise eigenvalue distribution as signal, and adds a signal robustness test that keeps only components whose direction barely changes when small noise is added to the matrix. According to the speaker, the direction of a biological signal shifts by 1 to 2 degrees, whereas noise shifts by 45 to 90 degrees. scICE (Kim et al., Nature Communications 2025) repeats clustering with different seeds and measures consistency with an inconsistency coefficient, narrowing down the range of cluster numbers that can be trusted. The gist was that finding a trustworthy signal may have to come before building a large model.
Prof. Jeongbin Park (Pusan National University College of Medicine) reversed the usual order of drawing cell boundaries first in spatial transcriptomics with SSAF-T. He sees the way cell boundaries are defined as the bottleneck that alters molecule assignment and cell type calls. SSAF-T first infers cell types from the density field of mRNA signals, then uses that result as a constraint to build cell boundaries afterward. Annotation is handled by CeLLMan, an LLM agent grounded in a marker knowledge graph compiled from human and mouse literature. What impressed me most was the principle that when marker evidence conflicts, or there is not enough evidence to subdivide further, a cell stays at the parent cell type, and a new state absent from the knowledge graph is refused assignment and left as an unknown type.
In the immune diseases session on the morning of day two, Prof. Heeju Ryu (Sungkyunkwan University School of Medicine) raised the same problem from another angle. High-dimensional analysis divides T cell states finely but often discards the information on which antigen the cell sees, and she pointed out that even with the same phenotype, specificity can split into tumor, viral, self-antigen, and bystander cells that respond to unrelated antigens. Her team combined combinatorial tetramer staining, which uses up to 120 tetramers at once through combinations of metal barcodes, with mass cytometry (CyTOF) to obtain more than 150 dimensions. In Merkel cell carcinoma patients, she reported that progression-free survival differed depending on whether virus (MCPyV)-specific CD8 T cells in the blood were detectable before treatment. In long COVID, not only SARS-CoV-2-specific cells but also CMV- and EBV-specific CD8 T cells skewed toward terminal differentiation and exhaustion (Cell Reports Medicine 2026).
In the same session, Prof. Jeong Seok Lee (KAIST) combined longitudinal blood multi-omics from 76 people, including myocarditis patients and controls, with heart tissue, and presented that the CCR2hi classical monocyte program in blood converges with the inflammatory niche in heart tissue. Both talks showed that reading disease and treatment response from circulating immune cells requires information beyond the surface phenotype.
Layers Read Together from One Sample
If the transcriptome alone cannot settle a cell state, one option is to read other layers from the same sample. Several talks over the three days went in this direction.
Dr. Chung-Chau Hon (RIKEN) presented, on the first day, work that reads transcription start sites separately from 5’ scRNA-seq data to build a map of transcribed cis-regulatory elements (tCREs). Applying tools such as SCAFE to data produced with the standard 10x 5’ protocol gives one more line of evidence, enhancer activity, from the same sample. According to the abstract, the map covers about 3 million cells from 30 human tissues, with 131 cell types and 687 cell states, and the team is preparing to release the single-cell heritability enrichment results through a web resource, SCREATE. The abstract also reports that Chromix, a DNA model trained on 83 genomes, outperformed the comparison models in prioritizing GTEx eQTLs.
Prof. Dongsung Lee (Seoul National University College of Medicine) summarized the application of snm3C-seq, which reads DNA methylation and 3D chromatin structure together in a single nucleus (Lee et al., Nature Methods 2019), to the developing human brain (Heffel et al., Nature 2024). Methylation remodeling in the hippocampus preceded that in the prefrontal cortex, and the timing of changes in methylation and chromatin structure during development was clearly distinct. Asked which comes first, he answered that during differentiation, chromatin structure changes were faster than methylation. He added, though, that proteins bind poorly to methylated sites, so methylation may act like a lock that fixes a lineage, meaning the two influence each other.
Prof. Kyunggon Kim (University of Ulsan College of Medicine) introduced single-cell and spatial proteomics based on mass spectrometry (LC-MS). His starting point was the low correlation between RNA and protein in single cells, and he named as an advantage that mass spectrometry, which does not depend on antibodies, directly measures drug targets and post-translational modifications. He showed both single-cell proteomics, where cells are dispensed one by one and quantified, and spatial proteomics, where regions of a tissue section are cut out by laser and quantified.
In the keynote on the third day, Prof. Yutaka Suzuki (University of Tokyo) showed a workflow that reads RNA on the same slide with a Xenium 2k panel and then reads additional protein targets with PhenoCycler. There were cases where the two layers disagreed. For CXCL13, the RNA was seen in B cells, but the protein is secreted and appeared in a different location. For the RA and RO isoforms of PTPRC (CD45), RNA could hardly separate them, while the protein distinguished memory T cells from naive T cells.
Virtual Cells Moving from Observation to Perturbation
The virtual cell came up repeatedly across sessions.
Prof. Byungjin Hwang (Yonsei University College of Medicine) began with the premise that the life sciences are moving from observation to perturbation. His argument is that learning cellular context, such as causality between genes, requires context-specific perturbation data.
SCITO-Perturb-seq, built for this purpose, uses combinatorial indexing to avoid the limit of multiple cells overlapping in a single droplet, raising throughput, and reads genome-scale CRISPRa perturbations together with surface proteins. According to the abstract, 201 surface proteins were measured in 3.6 million primary human CD4 T cells (preprint: Wang et al., bioRxiv 2026). Perturbation effects were decomposed into five protein modules by semi-NMF, and when compared with external functional screens, the modules performed better than individual markers. He also covered the remaining limit in predicting unseen perturbations.
Prof. Joon-Yong An (Korea University) trained a single-cell foundation model on NOCAP, an atlas of 501 brain organoid samples and 3.6 million cells, to predict genome-wide in silico perturbation responses. He then combined the response clusters with rare variant association results to prioritize candidate genes for autism spectrum disorder (Koh et al., bioRxiv 2026). One question in the Q&A got at the core of this approach: whether simple co-expression clustering such as cNMF or WGCNA would recover the same risk gene clusters. I heard it as a request to show the foundation model’s gain against a simple baseline.
Prof. Joonha Kwon (National Cancer Center Korea) introduced ReverseAGE, which uses the virtual cell in reverse, as the title of the talk (“Decoding the Aging-Cancer Nexus and Inverse Virtual Cells for Rejuvenation”) suggests. According to the abstract, it fixes a target state and uses CRISPRa and CRISPRi atlases to search backward for the combination of perturbations that reaches that state. The direction is to turn the virtual cell, which had stayed at forward prediction, into a design tool.
Cell Maps from Organoids to Patient Tissue
The first-day keynote, “Organoids Zoo” by Prof. Bon-Kyoung Koo (IBS), introduced a platform that maintains adult stem cell-derived intestinal organoids from 26 vertebrate species with a single commercial medium. According to the abstract, comparing species-specific transcriptomes identified Slc12a2 and Cdk6 as cross-species conserved intestinal stem cell markers. He also showed results from infecting organoids of several species with viruses to compare species-specific susceptibility. In the Q&A, he relayed the experience of an outside researcher who found that organoids have a wider dynamic range than cell lines in neutralizing antibody assays, making them worth using as a primary screen despite the cost.
Prof. Sangho Lim (KAIST) generated M cells, a rare epithelial cell type, in human intestinal organoids and reconstructed the differentiation trajectory from SPIB+ early, to ICAM2+ immature, to GP2+ mature cells. Unlike in mice, human M cells carried a dendritic cell-like program with constitutive MHC II expression, and when treated with gluten peptides they activated antigen-specific CD4 T cells. Defining a rare population stage by stage with marker combinations and confirming it by function read to me as one answer to the cell state definition problem discussed above.
The scale changes with patient tissue. The Asia-Pacific Spatial Translational Research Alliance (ASTRA), which Prof. Suzuki introduced, is led by the Garvan Institute and the University of Tokyo and aims to build a pan-cancer spatial atlas of more than 2,000 cases in the Asia-Pacific region. He explained the motivation for the launch with data showing that most GWAS participants are of European ancestry (Fatumo et al., Nature Medicine 2022). The data are targeted for deposit in DDBJ in the first quarter of 2027, and he answered that patient-level clinical information can be shared only within the collaboration under the terms of the ethics review.
In the cancer session on the third day, many studies read bulk cohorts through single-cell references. Prof. Semin Lee (UNIST) presented the first single-cell atlas of ampullary carcinoma (Jeong et al., Biomarker Research 2026). From 34,672 cells of 8 primary tumors and 4 normal samples, cNMF split malignant epithelium into four programs (Int-Wnt, Int-Hypoxia, PB-KRAS, Cycling). When 62 independent bulk samples were deconvolved, the PB-KRAS fraction was higher in recurrent tumors (p=0.011). In a multivariable Cox model, the hazard ratio for overall survival with PB-KRAS was 116.6, but the 95% confidence interval was wide, from 5.4 to 2,496.6 (N=61), so I read it for the direction of the effect rather than its size. In the Q&A, someone asked whether the top genes of an NMF program, which are sets of co-expressed genes rather than cell markers, can be used as a deconvolution signature, and the speaker acknowledged that active programs can be mixed in, so they are not entirely identical to cell types.
Prof. Jinhyuk Bhin (Yonsei University College of Medicine) deconvolved gastric cancer bulk cohorts with single-cell and spatial references and found 42 cell states that replicate across 11 microenvironment cell types. Among them, a fibroblast state emerged as the key that separates prognosis and anti-PD-1 response, and one of the ecotypes, grouped by how cell states co-occur, was associated with improved survival.
In the same session, Prof. Hae Ock Lee (Catholic University of Korea College of Medicine) said in the abstract that she integrated genomic, transcriptomic, single-cell, and spatial data from 56 resected early-stage non-small cell lung cancers to derive a lymphoid niche signature that complements TMB and PD-L1. Prof. Subin Lim (Ajou University) looked at tumor heterogeneity at several levels, within tumors, between tumors, and across cancer types, and introduced a line of work that started from a lung cancer prognostic index built from a set of extracellular matrix genes (matrisome) (Nature Communications 2017).
Methods for rereading archived fixed tissue at single-cell resolution appeared in the multi-omics session on day two. Prof. Chang Ho Sohn (KAIST) presented FX-seq, a single-nucleus transcriptomic method that exchanges the fixative of heavily fixed tissue to restore reverse transcription accessibility and combines it with combinatorial indexing; it is available as a preprint (Park et al., bioRxiv 2026).
Three Threads Across the Three Days
Three threads remained after I sorted the talks.
The definition of a cell state moves with analysis choices. A single seed changes the number of clusters, the order of segmentation changes cell type calls, and NMF programs can mix cell types with activity. Prof. Junhyong Kim’s remark that having over 1,000 tools is itself the problem belongs to the same context. Designs that withhold judgment when evidence is insufficient were one response: CeLLMan leaves a cell at the parent type, and scLENS keeps only components that pass the robustness test as signal.
The transcriptome alone cannot explain a cell. T cells with the same phenotype see different antigens, and in cases like CXCL13, RNA and protein appear in different places. Methods that read transcription start sites, methylation and 3D structure, and protein together from the same sample were filling this gap.
Perturbation data and simple baselines are both needed. The virtual cell was at the center of the move from observation to perturbation, and studies that generate context-specific perturbation data directly continued to appear. At the same time, there was a question asking to see the foundation model’s gain against simple co-expression clustering, and an argument to find a trustworthy signal before building a larger model. I came away feeling that the baselines to compare against have to be in place before a model is scaled up.
Next
On the evening of the second day, Dr. Yoshinari Ando (RIKEN) introduced the Human Cell Atlas Asia. The 2026 HCA Asia meeting will be held in Perth, Australia, from November 2 to 4, and the 2027 meeting was announced as tentatively planned for Jeju from October 11 to 13. The first data release from ASTRA was also announced for the first quarter of 2027. Once the spatial data from Asian patients is released, I plan to apply the analysis methods seen in these talks to real data.