An Information Limit for Assigning Shared-Gene Transcripts in Overlapping Cells of Imaging Spatial Transcriptomics

When an immune cell lies over a tumour cell in a tissue section, transcripts of the two cells fall inside the same segmentation boundary. For genes expressed by both cells, a single molecule carries no label of its origin. We ask how much of that origin can be recovered by any method. Modelling transcripts as a Poisson point process with linear mixing of cell types, we specialise the Bayes-optimal recoverability ceiling to overlapping cells and shared genes: where the intensity ratio of the two cells is constant over the overlap, no method assigns shared-gene molecules better than the majority class ratio, and only the fraction is recoverable. We test the bound on semi-synthetic overlaps built from real transcripts of two 10x Genomics Xenium sections (renal and breast cancer), where the origin of every molecule is known. In kidney, a calibrated learned classifier matches the class ratio in accuracy (0.760 against 0.759) yet carries 0.023 bits per molecule and lowers count error (0.181 against 0.194); the information lies in probabilities, not in hard labels. In breast it gains 0.076 in accuracy and 0.103 bits, about 10% of the label entropy. Hard assignment, including the segmentation tool Proseg, counts worse than soft assignment in both tissues, and the sign of Proseg's count bias changes between tissues (-0.38, +0.23). Oracles show that the count error is limited by cell-to-cell expression variation rather than geometry, and a pre-registered attempt to estimate per-cell expression did not reduce it. On natural overlaps typed by protein, soft counting reduces the dependence of tumour-cell RNA on CD45 protein. We also adapt a multichannel phase-randomisation control, used earlier for cosmological density fields, to separate second-order from higher-order spatial information and show that naive cross-evaluation overstates the latter by 98.8 to 99.8%. Preprint draft, version 1.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24375362
Primary Topic
Single-cell and spatial transcriptomics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

An Information Limit for Assigning Shared-Gene Transcripts in Overlapping Cells of Imaging Spatial Transcriptomics

Nguyễn Vũ Trọng Nhân
Zenodo (CERN European Organization for Nuclear Research)
Single-cell and spatial transcriptomics
preprint

An Information Limit for Assigning Shared-Gene Transcripts in Overlapping Cells of Imaging Spatial Transcriptomics

Nguyễn Vũ Trọng Nhân
preprint en

Abstract

When an immune cell lies over a tumour cell in a tissue section, transcripts of the two cells fall inside the same segmentation boundary. For genes expressed by both cells, a single molecule carries no label of its origin. We ask how much of that origin can be recovered by any method. Modelling transcripts as a Poisson point process with linear mixing of cell types, we specialise the Bayes-optimal recoverability ceiling to overlapping cells and shared genes: where the intensity ratio of the two cells is constant over the overlap, no method assigns shared-gene molecules better than the majority class ratio, and only the fraction is recoverable. We test the bound on semi-synthetic overlaps built from real transcripts of two 10x Genomics Xenium sections (renal and breast cancer), where the origin of every molecule is known. In kidney, a calibrated learned classifier matches the class ratio in accuracy (0.760 against 0.759) yet carries 0.023 bits per molecule and lowers count error (0.181 against 0.194); the information lies in probabilities, not in hard labels. In breast it gains 0.076 in accuracy and 0.103 bits, about 10% of the label entropy. Hard assignment, including the segmentation tool Proseg, counts worse than soft assignment in both tissues, and the sign of Proseg's count bias changes between tissues (-0.38, +0.23). Oracles show that the count error is limited by cell-to-cell expression variation rather than geometry, and a pre-registered attempt to estimate per-cell expression did not reduce it. On natural overlaps typed by protein, soft counting reduces the dependence of tumour-cell RNA on CD45 protein. We also adapt a multichannel phase-randomisation control, used earlier for cosmological density fields, to separate second-order from higher-order spatial information and show that naive cross-evaluation overstates the latter by 98.8 to 99.8%. Preprint draft, version 1.

Zenodo (CERN European Organization for Nuclear Research)
Vietnam Aviation Academy (VN)
Single-cell and spatial transcriptomics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon