cs.CVOct 20, 2026

VGL-stGen: Learning a shared multimodal model for lung histology and spatial transcriptomics

Authors: Haenara Shin, Dongjoo Lee, Sungwoo Bae, Kwon Joong Na, Daeseung Lee, Hongyoon Choi

Organizations: Portrai, Inc., Seoul, Republic of Korea

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

VGL-stGen is a multimodal large language model adapted from a pretrained medical vision-language model using paired lung H&E and spatial transcriptomics with single-cell references. Training integrates image-expression and expression-cell-type pairs, and gene-conditioned pseudo-labels transfer transcriptomic cell-type supervision to histology images. One deployed checkpoint supports thirteen evaluated tasks spanning tissue and cell-type classification, gene prediction and conditional expression completion from combinations of histology, cell identity and observed gene expression. Task-specific decoding produces categorical labels, gene lists and structured expression levels. For conditional completion, fixing queried gene identities and predicting their levels autoregressively improves structured prediction. Controlled adaptation shows that restricting supervision to requested expression-level tokens can further improve image-conditioned completion. Image-permutation experiments indicate that correctly paired histology adds information beyond observed expression. External classification relies partly on computational references with shared annotation lineage, and some H&E sources were exposed during cancer training or model development. Within these conditions, VGL-stGen demonstrates how a medical vision-language backbone can learn from spatial transcriptomics and histology to connect tissue morphology and molecular states through shared generative modeling.

Figures & tables

Figure 1: VGL-stGen overview. (a) Complementary lung supervision adapts one medical vision–language model for 13 evaluated tasks. The 6 label, 5 gene-list and 2 expression-level tasks share a decoder and checkpoint. (b) Gene-conditioned cell-type pseudo-labels supervise paired images during training. Conditional expression completion uses 75 observed levels and 25 named queries. Detailed task inputs and training stages are specified in the Methods and supplementary material.
IDInput →\rightarrow outputMetricEvaluation condition
1Genes →\rightarrow cell typeBA / F1mExternal
2Image + genes →\rightarrow cell typeBAH&E overlap
3Image →\rightarrow cell typeBAH&E overlap
4Image →\rightarrow tumourBA452 slides
5Image →\rightarrow tissueF1m39-slide pool
6Image →\rightarrow genesF1gH&E overlap
Table 1: 13 evaluated tasks supported by the deployed checkpoint. BA: balanced accuracy; F1m: macro-F1; F1g: gene-set F1; F1p: pair-F1; Acc: accuracy; AR: autoregressive. H&E overlap denotes partial exposure of evaluation tissue during cancer training. Gene-list completion supplies 15 names and predicts five; expression completion supplies 75 levels and queries 25 named genes. Task 10 supplies true answer length; task 13 supplies ten gene–level pairs. The complete score inventory is supplementary.
CohortEvaluated unitsReference / exposure
Visium9,964 spots / 2 sectionsTransferred cell types; one section supplied H&E for cancer training
Xenium214,246 cells / 2 sectionsTransferred cell types; both sections supplied H&E for cancer training
TCGA-LUAD binary357,692 patches / 452 slides / 398 patientsARA-CNN labels; includes development subset
TCGA tissue classification30,719 patches / 39 slidesDevelopment subset; pooled transductive correction
Masked expression benchmark1,000 rows/taskInternal; 7 image samples / 1 cell-type sample identifier
Table 2: Evaluation units and reference provenance, with spatial datasets grouped by platform. Visium units are spots; Xenium units are cells. Counts refer to processed evaluation items rather than training samples. Source-level provenance and the different exposure of the two Visium sections are retained in the supplementary methods. TCGA-LUAD denotes lung adenocarcinoma. The tissue-classification analysis uses the 39-slide development subset.
ModelT1 BA / F1mT2 BAT3 BAT4 BAT5 F1m
VGL-stGen0.6293 / 0.60230.53940.20120.75240.1200 →\rightarrow 0.2801
MedGemma-1.5-4B0.1836 / 0.15760.20700.11050.59540.0911 →\rightarrow 0.1711
Gemma3-4B0.2254 / 0.17380.18530.12500.66370.1334 →\rightarrow 0.1347
Qwen3-VL-8B0.4662 / 0.31480.45720.09950.51820.1102 →\rightarrow 0.1630
DeepSeek-VL2-small0.1314 / 0.11400.13800.11000.53510.0649 →\rightarrow 0.1321
GLM-4.6V-Flash0.3444 / 0.27750.31600.10080.58080.1411 →\rightarrow 0.1733
Table 3: Medical and general-purpose multimodal comparisons on tasks 1–5. BA is balanced accuracy and F1m is macro-F1. T1–3 use 224,210 items each; T4 uses 357,692 patches from 452 slides; T5 uses 30,719 patches from 39 development slides. T4 uses BF16 and reports slide-bootstrap means; the other tasks use NF4. T5 shows raw →\rightarrow pooled-corrected F1m for every model. Native image processors are retained. Task key: T1, genes →\rightarrow cell type; T2, image + genes →\rightarrow cell type; T3, image →\rightarrow cell type; T4, image →\rightarrow binary tumor status; T5, image →\rightarrow 9-class tissue type.
ModelT6T7T8T9T10T11T12T13
VGL-stGen0.11730.09650.06040.04840.28280.65960.68580.7420
MedGemma-1.5-4B0.00000.00050.00000.00000.00030.57330.55400.0000
Gemma3-4B0.00880.03450.01080.00000.00580.50720.49960.3500
Qwen3-VL-8B0.00000.08640.02100.03800.00320.16340.23720.3500
DeepSeek-VL2-small0.00000.00800.00260.00000.00420.16000.21650.0630
GLM-4.6V-Flash0.00440.03970.01400.02370.00300.36620.36270.4260
Table 4: Multimodal comparisons on gene and expression tasks, with 1,000 rows per task and NF4 weights. T6–10: gene-set F1; T11–12: pair-F1; T13: accuracy. T10 supplies the true answer length; T11–12 supply 75 observed levels and query 25 named genes. T6–10 and T13 use the primary generation protocol, separately from the format-constrained control. Values preserve the original score aggregation and rounding. Task key: T6, image →\rightarrow gene names; T7, cell type →\rightarrow gene names; T8, image + partial gene list →\rightarrow missing gene names; T9, cell type + partial gene list →\rightarrow missing gene names; T10, image →\rightarrow very-high-expression gene names; T11, image + observed levels →\rightarrow queried levels; T12, cell type + observed levels →\rightarrow queried levels; T13, gene-level pairs →\rightarrow cell type.
Figure 2: Task-specific comparisons. (a) Balanced accuracy (BA; circles) and macro-F1 (squares) for the same 20-gene input; connecting lines pair metrics, not confidence intervals. FT denotes fine-tuning; the two Geneformer variants are separate systems. (b) 14 encoder–probe comparators, VGL-stGen and an unranked colour reference on 452 TCGA slides. Intervals resample slides; the shaded band spans the 14 encoder point estimates. Classification references are computationally derived. TCGA includes the 39-slide development subset; point ordering does not establish significance between neighbouring models.
Figure 3: Gene-conditioned cell typing. (a,b) Published and predicted annotations for 7,993 GSE127465 cells. (c) Row-normalised annotation comparison. (d) Marker expression for 12,000 Xenium cells; colours clip at ±\pm 3, the epithelial set uses EPCAM alone, and pDC contains one cell. (e,f) Predicted cell positions and an aligned 400 μ\mu m H&E field.
Figure 4: Histology-derived molecular profiles and whole-slide queries. (a) 24 selected genes in 459 patients, ordered by measured lymphoid score and averaged over approximately seven patients per column. (b,c) Programme Pearson correlations and lymphoid association. (d) Matched-patient ranks. (e) EPCAM, CD3E, ACTA2 and PECAM1 predictions in 526 tiles of TCGA-49-4510. Queries use tile images and gene names without observed levels. Bulk RNA was extracted from a different tissue block than the diagnostic slide.
Appendix figures & tables26 assets

Supplementary material from the paper’s appendix.

Appendix

VariantRowsSource
Genes →\rightarrow cell type116,278Single-cell atlas
Cell type →\rightarrow genes115,207Single-cell atlas
Image →\rightarrow genes1,047,500Spatial samples; stain-expanded
Appendix
Table 5: Stepwise prompt variants used in earlier training. Answer targets remain the original class labels or gene lists.
TaskModelLevel macro-F1Non-Absent accuracy
ImageDeployed0.37800.2276
ImageFull-answer0.4269 ±\pm 0.00240.2664 ±\pm 0.0038
ImageLevel-only0.4463 ±\pm 0.00150.2930 ±\pm 0.0021
ImageFull-input MLP0.4586 ±\pm 0.00210.3160 ±\pm 0.0011
Cell typeDeployed0.46250.3611
Cell typeFull-answer0.5445 ±\pm 0.00080.4433 ±\pm 0.0049
Appendix
Table 6: Secondary expression metrics on the same benchmark rows. Mean ±\pm sample SD across 3 individual seeds; the deployed checkpoint is a single fitted model. Non-Absent accuracy restricts targets to the 4 detected-expression levels.
EvaluationSeedFull-answerLevel-onlyDifference
Image pooled420.670960.68032+0.00936
Image pooled70.672520.68164+0.00912
Image pooled20260.672520.67948+0.00696
Xenium420.717170.72523+0.00806
Xenium70.717050.72493+0.00788
Xenium20260.717350.72343+0.00608
Appendix
Table 7: Every additional-training seed and image platform. The pooled image set contains 665 Xenium cells and 335 Visium spots; the cell-type set contains 1,000 rows. Difference is level-only minus full-answer pair-F1. These repeat additional training from one initial checkpoint.
SupervisionSeedIntactPermutedDifference
Full-answer420.670960.66592+0.00504
Full-answer70.672520.66620+0.00632
Full-answer20260.672520.66596+0.00656
Level-only420.680320.67236+0.00796
Level-only70.681640.67452+0.00712
Level-only20260.679480.67128+0.00820
Appendix
Table 8: Image pairing under both supervision conditions. Each seed is evaluated on the same 1,000 rows and 25 queried levels per row with 75 observed levels. The within-sample/platform image permutation is identical across all 6 checkpoints. Positive differences indicate higher pair-F1 with intact images.
Obs.Model / imagePair-F1Level F1mNon-Absent acc.
75VGL-stGen / Intact0.68030.44660.2930
75VGL-stGen / Permuted0.67240.43080.2727
75Image MLP / Intact0.6811 ±\pm 0.00380.4586 ±\pm 0.00210.3160 ±\pm 0.0011
75Image MLP / Permuted0.6700 ±\pm 0.00300.4375 ±\pm 0.00050.2903 ±\pm 0.0026
75Context MLP / —0.6695 ±\pm 0.00920.4524 ±\pm 0.00860.3187 ±\pm 0.0055
50VGL-stGen / Intact0.66810.41730.2593
Appendix
Table 9: Synthetic completion under reduced observed context, with the same 25 queries. Obs. denotes observed gene levels. VGL-stGen uses the development-selected level-only seed 42; MLPs report mean ±\pm sample SD across 3 fixed seeds (width 512, learning rate 0.001). All models retain weights trained with 75 observed levels. The context-only MLP has no image input.
TasksItems per taskOutput limitBatch / shards
1224,210 cells/spots8 tokens16 / 4
2–3224,210 cells/spots16 tokens16 mixed / 4
4357,692 patches / 452 slides16 tokens32 / 8 slide shards
530,719 patches / 39 slides16 tokens32
6–71,000 rows160 tokens8 / 4
8–91,000 rows56 tokens8 / 4
Appendix
Table 10: Fixed task evaluation settings shared across models. Tokens are maximum generated tokens; T11–12 instead make 25 autoregressive level choices per row. T2–3 preserve the original mixed-task batch composition. Counts are per task. All tasks except T4 use NF4; T4 uses BF16. Task key: T1, genes →\rightarrow cell type; T2, image + genes →\rightarrow cell type; T3, image →\rightarrow cell type; T4, image →\rightarrow binary tumor status; T5, image →\rightarrow nine-class tissue type; T6, image →\rightarrow gene names; T7, cell type →\rightarrow gene names; T8, image + partial gene list →\rightarrow missing gene names; T9, cell type + partial gene list →\rightarrow missing gene names; T10, image →\rightarrow very-high-expression gene names; T11, image + observed levels →\rightarrow queried levels; T12, cell type + observed levels →\rightarrow queried levels; T13, gene–level pairs →\rightarrow cell type.
ModelRepositoryImage processing
VGL-stGengoogle/medgemma-1.5-4b-it896 ×\times 896; mean/std 0.5
MedGemma-1.5-4Bgoogle/medgemma-1.5-4b-it896 ×\times 896; mean/std 0.5
Gemma3-4Bgoogle/gemma-3-4b-it896 ×\times 896; mean/std 0.5
Qwen3-VL-8BQwen/Qwen3-VL-8B-InstructDynamic resolution; 16-pixel patches; merge 2
Qwen3-VL-32BQwen/Qwen3-VL-32B-InstructDynamic resolution; 16-pixel patches; merge 2
Gemma4-12Bgoogle/gemma-4-12B-itVariable resolution; 16-pixel patches; max 280 soft tokens
Appendix
Table 11: Model repositories and native image processing. VGL-stGen additionally uses the deployed lung adapter (merged for binary tumor classification, T4). Full revisions, processor settings, library versions and loading exceptions are supplied in the ancillary model-conditions files. Budgets describe configuration, not measured per-image token counts.
Figure 5: Comparison with ten unadapted vision-language models: (a) classification; (b) gene and expression tasks. Printed values are original scores; colors are column-wise z-scores across 11 models, using population standard deviation. T5 raw and pooled-corrected columns are standardized separately. BA: balanced accuracy; F1m/F1g/F1p: macro/gene-set/pair-F1; Acc: accuracy. Task key: T1–3, cell typing from genes, image + genes, or image, respectively; T4, image →\rightarrow binary tumor status; T5, image →\rightarrow nine tissue classes; T6–7, gene names from image or cell type; T8–9, missing genes from partial lists plus image or cell type; T10, image →\rightarrow Very High genes; T11–12, queried levels from observed levels plus image or cell type; T13, gene–level pairs →\rightarrow cell type. For paired task IDs, inputs follow numerical order.
TaskComparatorDifference95% interval
1Gemma4-31B+0.1258[+0.1204, +0.1309]
2Qwen3-VL-32B+0.0654[+0.0598, +0.0709]
3Gemma4-31B+0.0360[+0.0332, +0.0386]
4Gemma4-12B+0.0538[+0.0491, +0.0587]
5Qwen3.8-27B+0.0188[+0.0133, +0.0238]
6Gemma4-31B+0.0617[+0.0568, +0.0665]
Appendix
Table 12: Paired VGL-stGen minus highest-scoring comparator differences, with unadjusted 95% bootstrap intervals (1,000 draws). T5 uses corrected F1m. Units: cells/spots (T1–3), slides (T4), patches (T5), complete rows (T6–13). Differences use unrounded underlying scores; T4 uses observed BA rather than the displayed bootstrap mean. Comparator selection is descriptive and not adjusted for selection or multiple comparisons. Intervals condition on fitted models, not independent-patient generalization. Task key: T1, genes →\rightarrow cell type; T2, image + genes →\rightarrow cell type; T3, image →\rightarrow cell type; T4, image →\rightarrow binary tumor status; T5, image →\rightarrow nine-class tissue type; T6, image →\rightarrow gene names; T7, cell type →\rightarrow gene names; T8, image + partial gene list →\rightarrow missing gene names; T9, cell type + partial gene list →\rightarrow missing gene names; T10, image →\rightarrow very-high-expression gene names; T11, image + observed levels →\rightarrow queried levels; T12, cell type + observed levels →\rightarrow queried levels; T13, gene–level pairs →\rightarrow cell type.
ModelT6T7T8T9T10T13
VGL-stGen0.11730.09650.06040.04840.28280.7420
MedGemma-1.5-4B0.00050.01400.00650.01410.00110.3620
Gemma3-4B0.01860.06250.00880.01800.00570.3530
Qwen3-VL-8B0.00240.08500.02100.03800.00320.3500
DeepSeek-VL2-small0.00000.01030.00200.00280.00430.0630
GLM-4.6V-Flash0.00820.05010.01350.02620.00400.4220
Appendix
Table 13: Format-constrained generation on the same 1,000 rows per task, with unchanged NF4 precision, inputs and token limits. T6–10 report gene-set F1 under a gene-array grammar without a gene-panel whitelist; T13 reports accuracy under allowed-label spellings. Fixed token limits can still truncate outputs. Values preserve the original aggregation and rounding. Task key: T6, image →\rightarrow gene names; T7, cell type →\rightarrow gene names; T8, image + partial gene list →\rightarrow missing gene names; T9, cell type + partial gene list →\rightarrow missing gene names; T10, image →\rightarrow very-high-expression gene names; T13, gene–level pairs →\rightarrow cell type.
Figure 6: Supervised answer positions in the separate expression-adaptation study. Blue groups receive full-vocabulary cross-entropy. Full-answer supervision includes all answer tokens and end-of-turn; level-only supervision selects each queried level’s first token. Level rest denotes any remaining level tokens. Initial weights, inputs, update budget and autoregressive evaluation are matched. These checkpoints are separate from the deployed 13-task model.
Model / supervisionImageCell type
Deployed checkpoint0.65960.6858
Full-answer adaptation0.6720 ±\pm 0.00090.7221 ±\pm 0.0007
Level-only adaptation0.6805 ±\pm 0.00110.7201 ±\pm 0.0004
Full-input MLP0.6811 ±\pm 0.00380.7300 ±\pm 0.0008
Appendix
Table 14: Expression pair-F1 on the same 1,000 rows per task. Additional training and MLPs: mean ±\pm SD of 3 individual seeds, not ensembles. Each VGL-stGen adaptation uses 4,096 additional rows and 256 updates. MLPs use original task training splits; the image MLP uses UNI2-h, so encoders and training histories differ.
Figure 7: Controlled expression prediction after additional adaptation. (a) Level-only minus full-answer pair-F1 with intact inputs. (b) Intact-minus-permuted image effects; Difference denotes the level-only effect minus the full-answer effect. Points average three matched seeds. Conditional 95% bootstrap intervals resample image samples or cell-type rows and condition on fitted models. Panel b resamples samples. Image intervals preserve platform proportions. The 25 queries and within-sample/platform permutation are fixed. Sample identifiers are not verified patients.
ReadoutBalanced accuracyMacro-F1Trainable parameters
Discriminative short run0.55540.520932,819,212
Generative short run0.57370.568332,788,480
Appendix
Table 15: Matched adaptation-capacity control on the gene task. These short runs are separate from deployment. Data, rank, learning rate, schedule and update budget are matched; output heads and losses differ.
Figure 8: Region-query cancer mapping. H&E tissue (left) and the corresponding model cancer-score map (right) are shown for (a) TCGA-73-7499-01Z-00-DX1 and (b) TCGA-55-6979-01Z-00-DX1. These application examples illustrate selecting tissue and requesting a cancer-score map. Blue and red indicate lower and higher model scores, respectively. Scores are model outputs, not calibrated clinical probabilities.
Figure 9: Input–output map of the 13 deployed tasks. Numbers match the task inventory; the asterisk marks task 2 without a direct training objective. Filled cells use external datasets; outlined cells use internal evaluation data. External refers to data origin: 3 spatial sources supplied cancer-training H&E, task 4 includes TCGA development slides, and task 5 uses the development pool. Classification references are computationally derived. Gene-list completion supplies 15 names and predicts five; expression completion supplies 75 levels and queries 25 named genes. Task 10 supplies true answer length; task 13 supplies ten gene–level pairs. All tasks share the same checkpoint.
Figure 10: Training stages and evaluated checkpoints. Gene and cell-type adaptation precedes gene-mediated image supervision, joint cancer training with replay, and averaging. Stepwise prompts retain the original targets. The vision tower remains frozen. The pre-cancer checkpoint follows image supervision; deployment follows averaging. The averaging window was selected using the 39-slide TCGA development subset.
Figure 11: (a) Gene classifiers and VGL-stGen checkpoints on 224,210 items. The pre-cancer point uses its original evaluation (macro-F1 0.6104). FT denotes fine-tuning. Macro-F1 was not scored for the frozen-representation MedGemma probe in this panel; that probe is separate from the direct multimodal-model benchmark. (b) Paired macro-F1 differences with section-bootstrap intervals. The pre-cancer score in the paired comparison is 0.6102, versus 0.6023 after joint training and averaging. The MLP comparison is provisional because of discrepancies between refits.
Figure 12: Generative and discriminative readouts for gene-based cell typing under matched adaptation settings: rank, learning rate, schedule, batch, 6,000 updates, seed, prompts and data (40,000 evaluation items). Output heads and losses differ. Parameter counts are reported in the detailed Methods. Points are single-run estimates, separate from the deployed checkpoint.
Figure 13: Transferred versus target-cohort probe scores for (a) gene, (b) cancer and (c) morphology classification. Both axes report balanced accuracy; the diagonal denotes equal scores. Target-cohort fits use cell-level folds for genes and morphology and patient-level folds for cancer. The numbered model configurations are 1, pre-cancer VGL-stGen; 2, deployed VGL-stGen; 3, a task-specific linear probe on the same frozen vision tower. Panel (b) shows 2 and 3; panel (c) shows all three. All morphology comparisons use the same 54,058 evaluation items. VGL-stGen imaging coordinates use its own-tower probe reference, not a fitted VGL-stGen upper bound. Classification references are silver; spatial H&E exposure is partial and TCGA includes development slides.
Figure 14: Gene-budget analysis for logistic regression: target-cohort cross-validation and transferred balanced accuracy as the input list is shortened. Shading denotes fold standard deviation. Percentages refer to the target-cohort score’s margin above chance at 20 genes.
Figure 15: TCGA-LUAD tumor prediction: VGL-stGen and 14 encoder–probe systems, including the ImageNet ResNet50 reference. The color-histogram reference is unranked. Scores measure agreement with ARA-CNN-derived tumor labels; intervals resample slides. The shaded band spans encoder point estimates, not confidence bounds. The 452-slide set includes the 39-slide development subset; rank is descriptive.
Figure 16: Cancer scores on the earlier 39-slide subset and the expanded 452-slide evaluation. 13 of 14 encoders and VGL-stGen have lower point estimates on the expanded set. Slide-bootstrap intervals narrow with the larger sample; the two sets overlap and are not independent replications.
Figure 17: Cancer-label sensitivity on the same 452 slides. Each encoder probe is refitted under the wide mapping (TUMOR, NECROSIS and MIXED positive) and the TUMOR-only mapping. Grey and blue indicate the respective mappings; orange marks the colour reference under TUMOR-only. Targets, prevalence and probe fits change together.
Figure 18: Training and external gene frequencies across 1,049 shared symbols. Blue marks 1,014 genes in neither top-20 list; orange marks their 35-gene union. The dashed line denotes equal frequency. Spearman rho is 0.63. The five genes in both top-20 lists are ACTA2, EPCAM, ICAM1, MET and S100A9. Frequencies count cells or spots rather than task rows. Zero frequencies are placed at half the smallest positive observed frequency on the log axes.
IDTaskScoreMetricCondition
1Genes →\rightarrow cell type0.6293 / 0.6023BA / F1mExternal
2Image + genes →\rightarrow cell type0.5394BAH&E overlap
3Image →\rightarrow cell type0.2012BAH&E overlap
4Image →\rightarrow tumour0.7532BA452 slides
5Image →\rightarrow tissue0.1200 →\rightarrow 0.2801F1m39-slide pool
6Image →\rightarrow genes0.1173F1gH&E overlap
Appendix
Table 16: Deployed-checkpoint task inventory. BA: balanced accuracy; F1m: macro-F1; F1g: gene-set F1; F1p: pair-F1; Acc: accuracy; AR: autoregressive. H&E overlap denotes three of four spatial sources used in cancer training; SCC is absent from that manifest. Task 4 includes development slides; task 5 pools NF4-adapter scores from 30,719 patches / 39 slides for transductive correction, selected after labelled evaluation. Task 10 supplies the true answer length. Tasks 10/13 use baselines estimated from evaluation data; task 13 uses ten sampled gene–level pairs. Gene-list completion supplies 15 names and predicts five. Additional-training results are separate.
More in this topic

Gene Expression Prediction from Histopathology Images

Gene expression prediction from histopathology images links H&E tissue morphology to molecular profiles, enabling virtual spatial transcriptomics without sequencing every sample.

See all 45 papers

Explore similar work

CardsList

  1. Adapting a Pre-trained Single-Cell Foundation Model to Spatial Gene Expression Generation from Histology Images

    Mar 20, 2026Donghai Fang, Yongheng Li, Zhen Wang +2Gene Expression Prediction from Histopathology ImagesSpatial Transcriptomics

    Sun Yat-sen University, China · Yunnan University, China · Chongqing University, China

  2. Path2ST: Hierarchical Cell-Tissue Grounded Cross-Modal Translation for Spatial Transcriptomics

    Aug 11, 2026Ruochen Liu, Wei LouSpatial Gene Expression PredictionGene Expression Prediction from Histopathology Images

    Faculty of Science and Engineering University of Liverpool Liverpool, United Kingdom · College of Mathematical Medicine Zhejiang Normal University Jinhua, China

  3. A Multimodal Foundation Model of Spatial Transcriptomics and Histology for Biological Discovery and Clinical Prediction

    Apr 4, 2026Jinxi Xiang, Siyu Hou, Yuchen Li +17Spatial TranscriptomicsGene Expression Prediction from Histopathology Images

    Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA, USA · Department of Statistics and Data Science, Yale University, New Haven, CT, USA · Department of Medicine (Oncology), Stanford University School of Medicine, Stanford, CA, USA +4