Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction

Kaito Shiku; Kazuya Nishimura; Yasuhiro Kojima; Ryoma Bise

cs.CVSep 27, 2026

Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction

Authors: Kaito Shiku, Kazuya Nishimura, Yasuhiro Kojima, Ryoma Bise

Organizations: Kyushu University, Japan · The University of Osaka, Japan · National Cancer Center, Japan

Abstract

Predicting spatial gene expression from histology images could scale spatial transcriptomics (ST) to image-only cohorts, but conventional histology-based ST prediction is trained and evaluated mainly by per-gene spatial-profile reconstruction. This objective is misaligned with a key downstream use of ST: differentially expressed gene (DEG) discovery, where genes are ranked for a biological or morphology-defined contrast by evidence of between-group expression differences. We formulate image-based differential expression ranking (IDER), which asks whether predicted expression profiles preserve the contrast-specific ranked gene list obtained from measured profiles. IDER compares gene rankings induced by differential-expression statistics, rather than raw expression magnitudes or per-gene spatial correlations. We further introduce a differentiable IDER objective that aligns these statistics across genes and can be trained with morphology-derived proxy contrasts without predefined biological group labels. Experiments on public ST datasets show improved DEG-ranking agreement and pathway-enrichment overlap over conventional reconstruction objectives, including morphology-derived and pathologist-annotated tissue-region evaluations.

Figures & tables

Figure 1 : (a) Conventional evaluation measures spatial-profile agreement within each gene across spots. (b) Image-based Differential Expression Ranking (IDER) ranks genes by contrast statistics for a biological or morphology-defined comparison and evaluates whether predicted profiles preserve this ranked list of DEG candidates.

Objective	Ovary				Lymph Node				Bowel
	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$			SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$			SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$
	SCC DEG	@50	@100	@200	SCC DEG	@50	@100	@200	SCC DEG	@50	@100	@200
MSE [ 4 ]	0.587	0.415	0.446	0.434	0.624	0.407	0.434	0.477	0.709	0.364	0.432	0.468
PCC [ 30 ]	0.584	0.425	0.451	0.451	0.629	0.439	0.478	0.521	0.710	0.384	0.461	0.494
MSE $\&$ PCC [ 37 ]	0.609	0.437	0.465	0.453	0.650	0.477	0.511	0.548	0.714	0.383	0.459	0.496
Poisson	0.460	0.363	0.401	0.422	0.234	0.448	0.493	0.501	0.694	0.440	0.476	0.492

Table 1 : Comparison of DEG ranking performance in the cluster-based DEG setting against methods optimized with conventional objective functions.

Objective	Ovary	Lymph Node	Bowel	Breast
MSE [ 4 ]	0.326	0.382	0.336	0.144
PCC [ 30 ]	0.341	0.414	0.345	0.127
MSE $\&$ PCC [ 37 ]	0.332	0.433	0.314	0.125
Poisson	0.307	0.413	0.362	0.136
NB [ 24 ]	0.289	0.443	0.385	0.143
STRank [ 25 ]	0.302	0.383	0.360	0.147

Table 2 : Analysis of pathway enrichment overlap.

Objective	Average		adipose tissue		connective tissue		immune infiltrate		invasive cancer
Objective	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$
MSE [ 4 ]	0.239	0.084	0.393	0.069	0.206	0.080	0.197	0.176	0.272	0.082
PCC [ 30 ]	0.181	0.073	0.297	0.060	0.143	0.066	0.132	0.199	0.182	0.035
MSE $\&$ PCC [ 37 ]	0.195	0.073	0.317	0.076	0.208	0.048	0.134	0.239	0.216	0.043
Poisson	0.116	0.101	0.123	0.057	0.105	0.109	0.191	0.228	0.114	0.081
NB [ 24 ]	0.137	0.098	0.145	0.038	0.127	0.097	0.190	0.269	0.120	0.073

Table 3 : Comparison of DEG ranking performance in the real annotation-based DEG setting against methods optimized with conventional objective functions, in terms of SCC DEG and

\mathrm{nDCG}^{\mathrm{DEG}}

@200. Due to space limitations,

\mathrm{nDCG}^{\mathrm{DEG}}

@200 is abbreviated as

\mathrm{nDCG}^{\mathrm{DEG}}

Objective			Ovary		Lymph Node		Bowel		Her2st
Objective	U stat.	DEG rank.	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$	SCC DEG	$\mathrm{nDCG}^{\mathrm{DEG}}$
Baseline	✗	✗	0.584	0.451	0.629	0.521	0.710	0.494	0.181	0.073
Ours w/o DEG rank.	✓	✗	0.587	0.454	0.617	0.521	0.710	0.491	0.178	0.077
Ours	✓	✓	0.686	0.531	0.676	0.643	0.731	0.595	0.353	0.247

Table 4 : Ablation experiments of the proposed method .

Objective	Ovary	L. N.	Bowel	Breast
MSE [ 4 ]	0.235	0.171	0.345	0.113
PCC [ 30 ]	0.233	0.172	0.345	0.145
MSE $\&$ PCC [ 37 ]	0.235	0.172	0.341	0.141
Poisson	0.192	0.126	0.300	0.052
NB [ 24 ]	0.195	0.130	0.303	0.056
STRank [ 25 ]	0.190	0.130	0.300	0.065

Table 5 : Evaluation on per-gene spatial-profile PCC. L. N. indicates Lymph Node.

Objective	Ovary	L. N.	Bowel	Breast
MSE [ 4 ]	0.235	0.171	0.345	0.113
PCC [ 30 ]	0.233	0.172	0.345	0.145
MSE $\&$ PCC [ 37 ]	0.235	0.172	0.341	0.141
Poisson	0.192	0.126	0.300	0.052
NB [ 24 ]	0.195	0.130	0.303	0.056
STRank [ 25 ]	0.190	0.130	0.300	0.065

Table 5 : Evaluation on per-gene spatial-profile PCC. L. N. indicates Lymph Node.

Objective		$\mathrm{nDCG}^{\mathrm{DEG}}$
Objective	SCC DEG	@50	@100	@200
ST-Net [ 8 ]	0.181	0.060	0.062	0.073
ST-Net + Ours	0.353	0.212	0.230	0.247
TRIPLEX [ 4 ]	0.299	0.079	0.097	0.115
TRIPLEX+ Ours	0.457	0.190	0.212	0.242

Table 6 : Plug-in experiments for existing methods on the Her2st dataset.

Explore similar work

COAST: Context-Aware Differential Learning for Gene Expression Prediction in Spatial Transcriptomics
Jul 10, 2026Keunho Byeon, Sunhong Park, Jeewoo Lim +1Spatial Transcriptomics Transcriptomics
School of Electrical Engineering, Korea University, Seoul 02841, Republic of Korea
Gene Ontology-Guided Hierarchical Spatial Gene Expression Prediction from Histopathology Images
Aug 1, 2026Zhiwen Xu, Xiaoming Yan, Chengkun Wu +3Spatial Transcriptomics Genome-Wide Association Study
National University of Defense Technology Changsha, China
Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics
Sep 14, 2026Daniela Vega, Paula Cárdenas, Hannah Ceballos +2Spatial Transcriptomics Spatio-Temporal Transformers
Center for Research and Formation in Artificial Intelligence Universidad de los Andes, Colombia

Jul 10, 2026, cs.LGJ/K move · Enter open · S save

COAST: Context-Aware Differential Learning for Gene Expression Prediction in Spatial Transcriptomics

Keunho Byeon, Sunhong Park, Jeewoo Lim, Jin Tae Kwak

School of Electrical Engineering, Korea University, Seoul 02841, Republic of Korea

Spatial Transcriptomics Transcriptomics Digital Pathology Ocean

Open paper AskPDF arXiv

Spatial transcriptomics enables profiling of spatial gene expression but is limited by high cost and low throughput, motivating prediction from H&E histopathology images. Existing context-aware methods mainly supervise absolute expression, while relative expression relationships between spots are rarely used explicitly. We propose COAST, a context-aware differential learning framework for spatial gene expression prediction. COAST conditions the local and global context features with type-specific modulation and aggregates the target and context spot tokens using a Transformer encoder to capture both fine-grained local patterns and slide-level structure. It is trained with a joint objective that combines absolute expression regression with signed differential regression between the target and context spots. Experiments on multiple spatial transcriptomics datasets show consistent improvements in correlation- and distribution-based metrics, demonstrating the effectiveness of context-aware differential learning for histology-based spatial gene expression prediction.

Abstract

Explore similar work

COAST: Context-Aware Differential Learning for Gene Expression Prediction in Spatial Transcriptomics

Gene Ontology-Guided Hierarchical Spatial Gene Expression Prediction from Histopathology Images

Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics