KOGO 2026, Last Day: Measuring Rare Signals
Hi, I’m Kakyung Kim, a bioinformatics researcher.
The talks on the last day of KOGO 2026 covered multi-cancer early detection, subcellular sequencing, long-read transcriptomics, and cellular senescence, yet most of them dealt with the same problem: how to measure a signal far rarer than the background noise. Tumor DNA is well under 0.1% of the DNA in blood, and more than half of the transcripts in a neuron are present at fewer than 10 molecules per cell.
The 35th International Conference of the Korean Society for Genomics (KOGO) took place at the Sejong University Convention Center from September 3 to 5. I attended on the last day, September 5, and sat through Session 13, the keynote, and Session 16. Session 15 ran at the same time in the next hall, so I listened to its recording separately.
Senescent cells are rare in living tissue too, and no single marker works in every setting. This post goes through the day with one problem as the thread: the signal you want to measure is far rarer than the background noise.
Multi-Cancer Early Detection Today
Session 13, co-organized with the Korean Cancer Association (KCA), covered multi-cancer early detection (MCED): finding several cancers from a single blood test.
Prof. Tae-You Kim (Seoul National University College of Medicine) began with the bar MCED has to clear. The nine criteria that Braunstein and Ofman proposed in 2021 call for sensitivity at early stages, a fixed low false-positive rate, targeting of high-risk groups, prediction of the tissue of origin, detection weighted toward lethal cancers, prospective population-level validation, and cost-effectiveness, among others. Throughout the talk he used the five-phase development pathway of the U.S. National Cancer Institute’s Early Detection Research Network (EDRN) as a baseline.
The data on the commercial test Galleri stayed with me longest. In CCGA, a case-control study of 6,689 participants, overall sensitivity was 44.2% and specificity 99.3%. By stage, sensitivity was 18% for stage I, 43% for stage II, 83% for stage III, and 92% for stage IV. The later a cancer is found, the better the test catches it, which sits uneasily with the goal of early detection. He also noted that the 140,000-person NHS-Galleri randomized controlled trial was reported at ASCO 2026 as having missed its primary endpoint, a reduction in stage III-IV cancers. Screen-detected cancers were still four times as frequent as in the control arm. His starting point was that even with commercial products on the market, sensitivity at early stages remains an unsolved problem.
Prof. Junseok W. Hur (Korea University College of Medicine) put numbers on why early stages are hard. Circulating tumor DNA (ctDNA) from early-stage disease makes up well under 0.1% of circulating cell-free DNA (cfDNA), and ctDNA from a 1 cm³ lung tumor has been estimated at a variant allele frequency (VAF) of about 0.022% (Avanzini et al., Science Advances 2020). The trouble is that two bottlenecks are intertwined.
- Molecule recovery. ctDNA molecules are lost while the sequencing library is built. Hybrid-capture methods lose material at every step (end repair, ligation, pull-down, washing), so recovery stays at 30% or below, and 30 to 50% even after optimization.
- Error suppression. Below a VAF of 0.3-0.5%, sequencing errors are hard to tell apart from true variants. Unique molecular identifiers (UMIs) can push detection down to 0.1%, but they require ultra-deep coverage of more than 100,000x, which drives up cost.
Existing technologies tend to be good at only one of the two. His group presented MUTE-Seq, which uses a high-fidelity Cas9 to cut and remove wild-type cfDNA so that only mutant DNA remains (Ye et al., Advanced Materials 2025). The analogy was to burn down the haystack instead of searching it for the needle. Because the guide RNAs that eliminate the wild type have to be designed in advance, though, how to extend this to screening, where the variants are not known beforehand, was not addressed in the talk.
Prof. Yoon-Kyoung Cho (UNIST) presented work on extracellular vesicles (EVs) using a lab-on-a-disc, which moves samples by centrifugal force. I liked how she summarized the shared difficulty of circulating biomarkers: they are “rare, heterogeneous, fragile, and set against a large background.” Ultracentrifugation collects EVs at 100,000 g, whereas nanofiltration on the disc works at 500 g. She also showed a digital detection method that partitions 10 μL of plasma into picoliter droplets and counts the droplets that light up when EVs bind (ACS Sensors 2026). The talk went beyond diagnostics, to the mechanisms by which EVs prepare organ-specific metastasis and to their use as drug carriers.
Prof. Seung-Min Park (Nanyang Technological University) presented two lines of research that stood out. The first links magnetic capture with a nanowell single-cell array to read gene expression and mutations in individual rare cells. Even within one patient, mutant and wild-type signals were mixed from cell to cell, which exposed the cell-to-cell heterogeneity directly. The second is smart healthcare that continuously measures biosignals in daily life: a smart toilet that analyzes urine and stool (Nat Biomed Eng 2020) and smart glasses that record food intake. The smart-glasses study is under review, so I only describe the concept here. The shift from collecting a sample once to measuring continuously over time connected with the aging talks running in the parallel session.
Subcellular Sequencing of Neuronal Dendrites
The keynote was Prof. Junhyong Kim (University of Pennsylvania) on “Beyond Single Cells: Subcellular Sequencing Reveals Mosaic Evolution of the Dendritic Transcriptome in Rodent Neurons.” It read less as a methods introduction than as the formulation and testing of one evolutionary hypothesis.
The starting point was RNA localization in neuronal dendrites. Learning and memory require the protein composition of thousands of synapses to be tuned one by one, and sending mRNA made in the nucleus to specific positions in the dendrites is the solution. Yet when ten earlier studies were compared, the number of reported localized genes ranged from 60 to 4,500, with little agreement between studies. The underlying cause he gave was sparsity: even at 30 million reads per cell, more than 60% of the transcriptome is present at fewer than 10 molecules per cell.
His team separated the soma and the dendrites of a single mouse hippocampal neuron and sequenced them individually. Comparing within the same cell controls for cell-to-cell variation. They found about 1,800 genes consistently detected in more than 90% of dendrites (Middleton et al., BMC Biology 2019), and a rat-versus-mouse comparison showed that the dendritic transcriptome had diverged between the species much faster than the somatic or whole-cell transcriptome.
For interpretation he brought in Tomoko Ohta’s concept of system drift: when a trait arises from the interaction of many components, the components can change while the function stays the same. His analogy was clear, that what is conserved between a mechanical calculator and an electronic one is not the parts but the function of calculating. He translated this into a multidimensional quantitative genetics model and closed with the hypothesis that the more complex a trait system is, the more it is conserved evolutionarily, and the simpler it is, the more it drifts. A single keynote contained measurement technology, data analysis, and evolutionary theory.
Long Reads and Single Cells in Genomic Informatics
Afternoon Session 16 dealt with informatics that connects multi-omics data to the clinic.
Dr. Jonathan Göke (Genome Institute of Singapore, A*STAR) addressed the quantification problem in long-read RNA sequencing. He released SG-NEx, a benchmark dataset in which seven cell lines were read with five RNA-seq protocols plus spike-in controls, and compared the major isoforms called by short and long reads. For genes where both chose the same isoform, the quantification correlation was 0.93; it fell to 0.75 where only long reads called the isoform and to 0.49 where only short reads did. The difference persisted even when long reads were computationally cut into short ones, so read length alone did not explain it, and independent validation by digital PCR agreed with the long-read side (Chen et al., Nature Methods 2025). He said Bambu, the tool for long-read transcript discovery and quantification, is being extended to single-cell and spatial data.
Prof. Yoo-Jin Jiny Ha (Hanyang University) presented “Advanced Somatic Mutation Detection in Human Tissues with Multi-platform Sequencing Data.” The talk included unpublished collaborative work and the speaker asked that it not be photographed, so I list only the title. It dealt with detecting low-frequency somatic variants in normal tissue across several sequencing platforms.
Dr. Yayoi Natsume-Kitatani (National Institutes of Biomedical Innovation, Health and Nutrition, NIBN) showed two studies that combine clinical information with omics. One linked medical records, blood tests, and CT reports of interstitial pneumonia patients with the serum EV proteome, used association rule learning to find combinations that define particular patient groups, and went on to infer drug candidates. The other was EV-STAR-Protein, a Bayesian model that infers from which organ plasma EV proteins originate; in a preprint, the fraction of pancreas-derived EVs alone separated pancreatic cancer from healthy controls with an AUC of 0.931. It was a different route to tissue-of-origin estimation, which the morning session had named as a major challenge for MCED.
Prof. Hyobin Jeong (Yonsei University) introduced methods that read structural variants and epigenetic state together in single cells. scTRIP, based on Strand-seq, calls structural variants at haplotype resolution in a single cell, and scNOVA reads nucleosome occupancy from the same data (Nat Biotechnol 2023). In patients with T-cell acute lymphoblastic leukemia, only the clone carrying chromothripsis had an altered NOTCH1 pathway, and a NOTCH1 inhibitor lowered the proportion of that clone. In aged hematopoietic stem cells, she showed a case where an X chromosome inversion raised the activity of the androgen receptor gene (Nat Genet 2024). The core of the talk was that a single-cell map of structural variants extends to clone-specific drug vulnerabilities.
Heterogeneity of Senescent Cells
Session 15 in the neighboring hall covered cellular senescence. I worked from an automatic transcription of the recording, so I keep the numbers to a minimum and describe only the flow.
- Prof. Takuya Yamamoto (Kyoto University) tracked secondary senescence, in which inflammatory factors secreted by senescent cells (SASP) push neighboring cells into senescence, using chimeric mice, and reported that the spread depends on IL-1 signaling.
- Prof. Chanhee Kang (Seoul National University) explained, through selective autophagy and splicing regulation, how senescent cells lower overall protein synthesis while selectively increasing inflammatory proteins.
- Prof. Ok Hee Jeon (Korea University College of Medicine) summarized that senescent cells diverge along several trajectories even under the same stimulus, that no single senescence marker works under every condition, and that consistency shows up at the pathway level rather than for individual genes.
- Dr. Chuna Kim (KRIBB) measured human visceral preadipocytes at five time points over 20 days and showed decoupling, with the transcriptome and proteome moving in opposite directions at certain time points. A single measurement at the endpoint would not have revealed it.
Three Threads Across the Day
Rereading my notes, I pulled out three threads.
A single metric cannot make the call. MCED has to satisfy sensitivity, specificity, and tissue-of-origin prediction all at once. In senescence research there was no single marker that worked in every condition, and in dendritic localization no sequence motif proved necessary and sufficient. All three fields are moving their criteria toward combinations of metrics or toward the pathway level.
Some things only show up when measured as trajectories. The 20-day senescence study, the single-cell structural variant work that places clonal evolution on a time axis, and smart healthcare with continuous measurement in daily life all pointed the same way. A measurement at one time point cannot reveal the order of changes or their turning points.
When the signal is rare, the measurement design decides the result. The choice of read type changed which isoform was called the major one, the trade-off between recovery and error suppression set the detection limit for ctDNA, and a within-cell comparison design reduced the noise in the dendritic transcriptome. As someone who analyzes data, I was reminded that the measurement design has to be examined before the model is chosen.
Next
At the closing ceremony, the KOGO Winter Symposium was announced for February 1 to 3, 2027. About six weeks later, in mid-October, I attended the SCSOK symposium on single-cell and spatial omics, and I plan to write it up in the next post. Prof. Junhyong Kim, today’s keynote speaker, also gave a keynote at that symposium, so the two meetings can be read together.