PaperlayerView exact revision on arXiv

Source-linked AI summary

Understanding Gene Expression-to-Image Learning: Spatial Structure and Information Fidelity

Xiaoqing Lian, Pei Liu, Jiahui Tan, Quan Zou, Chen Lin, Xiangxiang Zeng

arXiv:2610.17699v1cs.AIcs.LGq-bio.GN

TL;DR

Learning representations from high-dimensional, unordered gene-expression profiles remains challenging, and how GE2I layouts affect learning is unclear. The paper analyzes full-transcriptome GE2I under shared network and training conditions and finds that preserving gene-level information can improve prediction even as image smoothness decreases.

  • Problem

    Gene-expression profiles are high-dimensional and unordered, and the effects of GE2I layout choices on learning remain unclear.

    Source: §1 · pp. 1–2

  • Method

    The study measures image properties, tracks training behavior, relates descriptors to prediction with beta regression, and tests pixel-assignment interventions under fixed network and training conditions.

    Source: §1 · p. 2, §3.4 · p. 4

  • Results

    One-to-one gene-to-pixel assignment improves prediction even when image smoothness decreases.

    Source: §1 · p. 2, §5.3 · p. 8, §6 · p. 10

  • Takeaways & Limitations

    Transcriptomic image design should balance spatial organization with gene-level information preservation rather than treat smoothness or early optimization as a proxy for prediction quality.

    Source: §6 · p. 10

    The experiments focus mainly on cell-type prediction with a shared CNN and training protocol, so they do not measure the best performance attainable through method-specific tuning.

    Source: §6 · p. 10

Abstract

from arXiv · show

Single-cell RNA sequencing (scRNA-seq) enables cell-level analysis of biological heterogeneity, but learning effective representations from high-dimensional, unordered gene expression profiles remains challenging. Gene Expression-to-Image (GE2I) offers an effective learning strategy by arranging gene values on pixel grids for convolutional neural networks, but its underlying mechanisms remain unclear. We study GE2I on full-transcriptome data while keeping the network and training procedure fixed. Our results reveal a trade-off between spatial smoothness and gene-level fidelity. Structured mappings produce smooth, low-frequency patterns that support faster early optimization, whereas assigning multiple genes to the same pixel obscures expression differences and can reduce predictive performance. This trade-off explains why random mappings can compete with structured layouts when many genes are represented. One-to-one gene-to-pixel assignment improves prediction across the tested visual backbones even when image smoothness decreases. We further find that the benefits of pretraining and augmentation depend on the backbone and mapping. These findings provide practical guidance for pixel allocation and the transfer of vision priors to gene expression data.

1 Introduction

GE2I arranges unordered gene-expression values on a pixel grid so a CNN can learn their patterns, but structured layouts affect both gene relationships and image smoothness. The paper examines how that trade-off relates to early learning and final prediction, and tests pixel-assignment changes and vision-training choices.

Approach

  • GE2I places unordered gene values on a 2D pixel grid for CNNs to learn high-order statistics and nonlinear correlations.The paper notes that placing correlated genes together may combine weak signals into clearer biological patterns.Source: §1 · pp. 1–2, Figure 1 · p. 2

Research question

  • Structured layouts change both gene relationships and spatial smoothness, leaving unclear which property accounts for better learning.The paper investigates this question on full-transcriptome data while holding network and training conditions fixed.Source: §1 · p. 2

Mechanistic analysis

  • Image measurements, training dynamics, and beta regression relate spatial smoothness and gene-level fidelity to early optimization and final prediction.The analysis distinguishes how the two properties affect these outcomes.Source: §1 · p. 2

Assignment interventions

  • One-to-one gene-to-pixel assignments improve prediction even when they make images less smooth.The authors also introduce a greedy allocation procedure and compare assignment effects across visual backbones.Source: §1 · p. 2

Backbone evaluation and transfer guidelines

  • The paper compares assignment effects across visual backbones and tests pretraining and augmentation to clarify when vision priors help transcriptomic learning.The supplied introduction does not state which backbone or mapping benefits from those choices.Source: §1 · p. 2

2 Related Work

Single-cell representation methods either model expression profiles directly or first arrange gene measurements as images for visual models. GE2I work varies in how layouts are constructed, while this paper studies how smoothness and gene-level fidelity relate to learning under shared conditions.

2.1 Representation Learning in Single-Cell Genomics

  • Variational autoencoders and foundation models learn cell representations from expression profiles without first assigning genes to a 2D grid.scANVI can incorporate available cell-type annotations; foundation models use large-scale pretraining.Source: §2.1 · pp. 2–3

2.2 The Gene Expression-to-Image (GE2I) Paradigm

  • GE2I makes gene arrangement an additional representation-design choice before applying a visual model.Existing layouts use dimensionality reduction, Fourier transforms, local search, distance-rank matching, or optimal transport to encode gene relationships.Source: §2.1 · p. 3, §2.2 · p. 3
  • This work relates spatial smoothness and gene-level fidelity to early optimization and final prediction under shared network and training conditions.Pixel-assignment interventions test the proposed explanations; prior final-score comparisons alone do not explain how construction choices affect learning.Source: §2.2 · p. 3

3 Analytical Framework

The analytical framework represents each cell's gene-expression vector as a 2D image and characterizes both spatial frequency structure and gene allocation. A common CNN probe and beta regressions connect these descriptors to optimization and prediction outcomes.

3.1 Problem Formulation

  • A coordinate mapping assigns genes to an H × W grid, and a permutation-invariant operator aggregates values at each occupied pixel.The mapping is shared across cells; even when the grid has at least as many pixels as genes, multiple genes may share a pixel.Source: §3.1 · p. 3

3.2 Spectral Descriptors (Quantifying Smoothness)

  • Spectral entropy measures the spread of power across radial frequencies, while higher low-frequency energy ratio indicates smoother spatial patterns.The radial profile is computed by averaging the centered 2D power spectrum within integer-radius rings and normalizing it.Source: §3.2 · pp. 3–4

3.3 Information Fidelity (Quantifying Resolution)

  • APPG and collision rate describe potential information loss through allocation, rather than directly measuring task-relevant information.APPG = 1 means each gene occupies a unique pixel for single-view assignments; zero collision rate means no genes share a pixel. For Genomap, collision rate denotes soft mixing despite APPG = 1.Source: §3.3 · p. 4

3.4 The Convolutional Probe

  • The ResNet-18 probe uses early loss reduction to characterize optimization and accuracy and F1-Macro to measure final prediction.Beta regressions relate all four image descriptors jointly to these outcomes, with Full and HVG inputs and the two prediction metrics modeled separately; pixel-assignment interventions test whether changing allocation alters prediction.Source: §3.4 · p. 4

4 Experimental Protocols

The experiments evaluate cell-type classification across five single-cell datasets and gene-to-protein prediction on CITE-Seq, using multiple image mappings and external methods for context. Main GE2I comparisons share a ResNet-18 and training protocol, with additional tests of backbones, pretraining, and augmentation.

4.1 Datasets and Benchmarks

  • Cell-type classification is evaluated on five datasets spanning disease-specific cohorts and cross-tissue atlases, with gene-to-protein prediction also assessed on NeurIPS 2021 CITE-Seq.Main analyses use full-transcriptome inputs; complementary highly variable gene evaluations are reported in the appendix.Source: §4.1 · p. 4

4.2 Gene Expression-to-Image Transformation Strategies

  • GE2I strategies include projection-based layouts that average values at shared pixels, adaptive methods with one-to-one assignments, and random unique coordinates.Genomap uses optimal transport-weighted pixel sums, while IGTD and REFINED optimize one-to-one assignments; Random Mapping does not use gene similarity and fixes coordinates across cells.Source: §4.2 · p. 5

4.3 Baseline Comparisons

  • External methods provide performance context rather than part of the controlled GE2I comparison.They include XGBoost on PCA features, Seurat, scANVI with and without scArches, and zero-shot scGPT and CellFM embeddings; a tuned raw-expression MLP is an appendix reference.Source: §4.3 · p. 5

4.4 Training Protocol

  • Main GE2I comparisons use the same ResNet-18 backbone and training hyperparameters without mapping-specific network tuning.Means and standard deviations are reported across three random seeds; mapping construction is timed separately, and additional experiments test visual backbones, ImageNet pretraining, and augmentation.Source: §4.4 · p. 5

5 Results

GE2I supports competitive prediction, but no layout wins across datasets. Structured layouts aid early optimization, while collision-free gene assignment improves prediction despite lower smoothness; training choices also depend on the mapping and backbone.

5.1 How Well Do Gene Expression Images Support Cell-Type Prediction?

  • GE2I representations perform comparably to established latent-variable methods, but no layout leads across datasets and Random Mapping is strongest among reported GE2I methods on Tabula Sapiens.DeepInsight (t-SNE) beats Random Mapping on GTEx v9, while external methods retain their recommended protocols, so this is not a fully optimized pipeline ranking.Source: §5.1 · pp. 5–6

5.2 How Do Image Properties Relate to Optimization and Prediction?

  • DeepInsight and REFINED have full-transcriptome LFER values of approximately 0.8–1.0 versus about 0.24 for Random Mapping, and high-LFER mappings reduce DKD loss faster over the first 60 steps.Faster early optimization does not establish better final prediction.Source: Figure 2 · p. 6, §5.2 · p. 7, Figure 4 · p. 7
  • Full-transcriptome beta regressions associate higher APPG and CR with lower prediction, and higher LFER and SE with higher prediction; all eight associations remain significant after adjustment.APPG has a negative coefficient even after holding CR and spectral descriptors fixed, so more pixels alone do not guarantee better prediction.Source: §5.2 · p. 7, Table 3 · p. 8

5.3 Can Changing Pixel Assignment Improve Prediction?

  • Increasing DKD image dimensions from 64² to 384² changes F1-Macro from 0.8990 to 0.9013, with lower scores at intermediate sizes and substantial gene overlap remaining.Changing pixel allocation, rather than enlarging the canvas, is the intervention that removes collisions in these experiments.Source: Table 4 · p. 8, §5.3 · p. 8
  • All six greedy reassignment comparisons raised mean F1-Macro with CR = 0 and APPG = 1; DeepInsight (t-SNE) on GTEx v9 rose approximately 0.506 to 0.556 while LFER fell from 0.867 to 0.756.LFER decreased in every comparison.Source: Table 5 · p. 9, §5.3 · p. 9

5.4 How Do Visual Backbones and Training Choices Affect GE2I?

  • Unique assignment improves mean F1-Macro by 3.4%–7.0% across five tested backbones; pretraining, geometric augmentation, and combined augmentation have architecture- or mapping-dependent effects.Intensity perturbations improve mean F1-Macro for all three tested mappings, while DeepInsight (t-SNE) has a higher Grad-CAM marker hit ratio than Random Mapping on DKD, 0.50 versus 0.25.Source: §5.4 · pp. 9–10

6 Conclusion and Limitations

Under controlled training conditions, GE2I trades off spatial smoothness against gene-level fidelity, and preserving expression information improves prediction even when images become less smooth. The experiments focus mainly on cell-type prediction with a shared CNN and training protocol, and pixel-level descriptors do not capture higher-order gene relationships.

6 Conclusion and Limitations

  • Preserving gene-level information improves prediction even when GE2I images become less smooth.The conclusion advises balancing spatial organization with information preservation rather than treating smoothness or early optimization as a proxy for prediction quality.Source: §6 · p. 10
  • The experiments focus mainly on cell-type prediction with a shared CNN and training protocol, so they do not measure best performance after method-specific tuning.Source: §6 · p. 10
  • Pixel-level descriptors do not capture higher-order gene relationships; future work should test adaptive allocations, larger multimodal datasets, and foundation models.Source: §6 · p. 10

AI Use Statement

Generative AI tools supported research ideation, literature discovery, manuscript drafting, and language polishing; the authors reviewed suggestions and verified the work.

AI Use Statement

  • Generative AI supported ideation, literature discovery, drafting, and language polishing, while the authors verified experiments, numerical results, and statistical analyses.Source: AI Use Statement · p. 10

Ethics Statement

The study used previously released single-cell transcriptomic datasets and collected no new human-subject data; its analyses are intended for computational genomics, not clinical use.

Ethics Statement

  • The study used previously released single-cell transcriptomic datasets, collected no new human-subject data, and was intended for computational genomics rather than clinical diagnosis or treatment.Source: Ethics Statement · p. 11

Reproducibility Statement

The paper describes its experimental and analytical procedures in the main text and appendix, reports results across three random seeds where applicable, and plans to release source code upon publication.

Reproducibility Statement

  • Experimental settings, preprocessing, mapping, training, statistical analyses, and pixel-assignment methods are described in the main text and appendix; results use three random seeds where applicable.Source code will be released upon publication.Source: Reproducibility Statement · p. 11

A A Theoretical View of the Smoothness–Fidelity Trade-off

The analysis explains how pixel aggregation can simultaneously smooth expression images and erase gene-level class information. The accompanying experiments examine collision-free assignment, cross-modality prediction, raw-expression baselines, and Grad-CAM interpretability.

A A Theoretical View of the Smoothness–Fidelity Trade-off

  • Mean aggregation reduces noise-induced spatial variation but removes within-pixel class-discriminative gene contrasts, explaining the smoothness–fidelity trade-off.On a regular grid, lower local variation connects to low-frequency structure; when grouped genes share the same class-dependent signal, averaging can smooth without losing that contrast.Source: A A Theoretical View of the… · pp. 14–15

B Datasets and Evaluation Benchmarks

  • The benchmarks use standardized OpenProblems preprocessing and include disease and cross-tissue atlases, a highly imbalanced Tabula Sapiens cohort, and CITE-Seq protein prediction.Cell-type annotation reports accuracy and F1-Macro, prioritizing F1-Macro because it weights cell types equally; protein regression reports RMSE, Pearson, and Spearman correlations.Source: B Datasets and Evaluation Benchmarks · p. 15, §B.1 · p. 15, §B.2 · p. 16

C Detailed Experimental Settings

  • GE2I mappings use method-specific preprocessing and algorithms; the shared ResNet-18 training setup uses AdamW, warmup, cosine annealing, and early stopping.DeepInsight variants use mean aggregation for collisions, while IGTD enforces one-to-one assignment and Random Mapping supplies a one-to-one baseline.Source: §C.1 · p. 16, §C.4 · p. 18

C.2 Collision-Aware Unique Assignment

  • With HW ≥ G, the greedy assignment guarantees unique pixels, but depends on gene order and tie-breaking and does not guarantee a global optimum or preserve neighborhoods.The procedure minimizes displacement from each rounded target conditional on earlier assignments.Source: §C.2 · p. 18

D Cross-Modality Gene-to-Protein Prediction

E.1 Full-Genome MLP Ablation

  • The raw-expression MLP is competitive on DKD, ICA, and TS, while the best structured GE2I methods have higher F1-Macro on GTEx v9 and HypoMap.On GTEx v9, high MLP accuracy does not imply equally strong performance across cell types; the comparison supports a dataset-dependent role for projection.Source: §E.1 · p. 19, Table 9 · p. 19

E.3 Interpretability Trade-off

  • DeepInsight (t-SNE) has a Grad-CAM marker hit ratio of 0.50 versus 0.25 for Random Mapping, with more canonical markers recovered in salient regions for evaluated cell types.The authors treat marker alignment as evidence for gene-level inspection, not causal validation of identified genes.Source: §E.3 · pp. 19, 21, Table 13 · p. 20

F Comprehensive Metric Breakdown

Full-transcriptome manifold layouts can create more gene collisions than their highly variable gene counterparts, while collision-free methods can retain smoothness. These findings distinguish collision avoidance from texture preservation.

F Comprehensive Metric Breakdown

  • For DeepInsight (t-SNE) on HypoMap, collision rate rises from 0.198 with HVGs to 0.641 with full-genome inputs.Across full-transcriptome manifold layouts, collision rate increases relative to HVG counterparts, potentially reducing the number of expression values represented separately.Source: F Comprehensive Metric Breakdown · p. 21, Table 14 · p. 21
  • IGTD and REFINED achieve CR = 0 and LFER near 1.0, combining unique gene assignments with low-frequency spectral concentration.This separates collision avoidance from texture preservation: both can occur together in these layouts.Source: F Comprehensive Metric Breakdown · p. 21, Table 14 · p. 21

G Statistical Analysis Diagnostics

Descriptor–performance relationships vary between Full and HVG inputs. Full-transcriptome associations also have robustness qualifications, and run coverage is uneven.

G Statistical Analysis Diagnostics

  • Inference clusters repeated seeds by dataset–method, uses CR1 standard errors and t-based p-values, and adjusts 16 descriptor tests across outcomes and input scopes with Benjamini–Hochberg q-values.The adjusted threshold is q < 0.05.Source: G Statistical Analysis Diagnostics · p. 23
  • Of 270 potential runs, 243 contain both outcomes and 240 contain all four descriptors; Full coverage is uneven, and imputed values are treated as fixed in standard errors.The Full and HVG models use 105 and 135 runs, respectively; Full coverage includes three Genomap DKD runs, 12 REFINED runs, and no complete IGTD records.Source: G Statistical Analysis Diagnostics · p. 23
  • In HVG inputs, no descriptor reaches q < 0.05 for F1-Macro; APPG is positive and CR negative for accuracy, while LFER and SE miss the adjusted threshold.The regression table reports the HVG setting.Source: G Statistical Analysis Diagnostics · p. 23, Table 15 · p. 23
  • In Full inputs, APPG and CR associations weaken under dataset-specific precision, while spectral associations attenuate with method fixed effects; spectral positives persist under two key exclusions.The exclusions use only directly measured descriptors or omit Genomap.Source: G Statistical Analysis Diagnostics · p. 24

Related papers

Loading 2610.17699v1…