MSACL 2026 Abstract

Deep Morphology Embeddings from H&E Capture Fucosylation Signatures in Triple-Negative Breast Cancer

Keervani Kandala (1,2,3), Danielle A. Scott (4), Richard R. Drake (4), Peggi Angel (4), Aimilia Gastounioti (1,2,5,6), José Marcio Luna (1,2,3,6)
(1) Mallinckrodt Institute of Radiology, Washington University in St. Louis, St. Louis, MO, USA (2) Institute for Informatics, Data Science and Biostatistics, Division of Biological and Biomedical Sciences, Washington University in St. Louis, St. Louis, MO, USA (3) Luna Lab, Precision Radiotheranostics Translation Center, Washington University in St. Louis, St. Louis, MO, USA (4) Department of Pharmacology and Immunology, Medical University of South Carolina, Charleston, SC, USA (5) Breast Image Computing Lab, Computational Imaging Research Center, Washington University in St. Louis, St. Louis, MO, USA (6) Alvin J. Siteman Cancer Center, Washington University in St. Louis, St. Louis, MO, USA

Keervani Kandala, MS (Presenter) Mallinckrodt Institute of Radiology, Washington University in St. Louis

Presenter Bio: I am a pre-doctoral researcher at the Mallinckrodt Institute of Radiology at Washington University in St. Louis, working in the Precision Radiotheranostics Translation Center under Dr. Jose Marcio Luna. My work sits at the intersection of spatial omics, quantitative imaging, and AI-driven computational pathology, with a focus on understanding the tumor microenvironment in prostate and breast cancer.
A central question driving my research is whether non-invasive imaging can serve as a window into molecular tissue biology that currently requires direct tissue interrogation. I approach this through multimodal frameworks that pair imaging modalities with spatial mass spectrometry, and through foundation models applied to routine histopathology to recover molecular signals without task-specific supervision. Running through all of this is a deeper interest in how the engineering choices underlying representation learning, from supervision regime to pretraining scale, determine what these models actually learn from tissue images and where they silently fail. I care about building tools that are not just technically impressive but clinically grounded and interpretable when it matters most.

Relevant Financial Disclosures (within past 24 months, reported on Apr 22, 2026) No relevant financial relationship(s) to disclose.

INTRODUCTION:
Glycosylation is a key post-translational modification implicated in cancer progression, immune evasion, and metastasis. In triple-negative breast cancer (TNBC), a subtype lacking estrogen receptor, progesterone receptor, and HER2 expression and therefore devoid of targeted therapies, aberrant N-glycosylation has emerged as a hallmark of malignancy. Core fucosylation (CF), defined as the addition of α-1,6-linked fucose to the innermost GlcNAc residue of N-glycans mediated by fucosyltransferase 8 (FUT8), and outer arm fucosylation (OAF), defined as the addition of fucose to peripheral antennae of branched glycan structures, are associated with tumor aggressiveness, altered immune recognition, and patient survival outcomes [1]. A recent spatial glycomics study profiling 348 N-glycans across 59 TNBC tumors found that 80% of all detectable N-glycans are fucosylated, with OAF glycans linked to tumor stage progression and 52% of survival-associated glycans carrying OAF modifications, including polylactosamine extensions linked to metastasis [1]. MALDI-MSI enables spatially resolved glycan profiling on formalin-fixed paraffin-embedded tissue sections [2], but requires specialized instrumentation not yet routinely integrated into clinical workflows. H&E staining is ubiquitous in diagnostic pathology and captures rich morphologic information about tumor cellularity, nuclear pleomorphism, stromal composition, and immune infiltration. Self-supervised histopathology foundation models can extract biologically meaningful morphologic features from H&E images without manual annotation [3]. This study tests whether deep morphological features from routine H&E slides can serve as surrogates for tumor fucosylation signatures in TNBC, hypothesizing that glycosylation states are reflected in histologic architecture and coupled to morphology even without spatial coregistration.

METHODS:
We designed a three-phase pipeline linking H&E whole-slide images to fucosylation signatures at the case level. In Phase 1, we performed automated tissue detection at OpenSlide pyramid level 5 using saturation and brightness thresholds to isolate viable tumor regions, with pen annotations masked via hue-based and channel-dominance criteria. Tissue masks were cleaned via morphological erosion and dilation. Non-overlapping 224 x 224 pixel tiles at 20x effective magnification were sampled from regions with ≥50% tissue coverage, extracting 512 tiles per slide. Tile-level morphology embeddings were computed using Phikon-v2, a vision transformer pre-trained on approximately 460 million pathology tiles from over 55,000 whole-slide images across more than 30 cancer sites, aggregated to slide-level representations via median pooling, and H&E morphology principal components (PCs) were extracted via PCA. In Phase 2, slide-level CF and OAF scores were derived from MALDI-MSI glycan profiling: 123 CF N-glycans isolated via endoglycosidase F3 cleavage and 153 OAF N-glycans released by subsequent PNGase F digestion, yielding a CF fraction-high score (CF score) and OAF principal component score (OAF score). In Phase 3, associations between H&E morphology PCs and fucosylation proxies were assessed using Spearman correlation with Benjamini-Hochberg false discovery rate correction at q ≤ 0.10.

RESULTS:
Of 42 TNBC cases, 41 were retained after excluding one slide with insufficient tumor content. PCA of H&E morphology embeddings revealed that PC1 explained 24.62% of variance, with the first ten components collectively accounting for 74.24%, indicating a compact low-dimensional morphologic structure. After Benjamini-Hochberg correction, PC1 remained significant at q ≤ 0.10, with strong monotonic associations with both fucosylation proxies: OAF score (Spearman ρ = −0.787, p = 1.50 × 10⁻⁷) and CF score (Spearman ρ = −0.739, p = 2.06 × 10⁻⁶). The negative directionality indicates that slides scoring higher on PC1 tend to exhibit lower overall fucosylation. No other PCs survived multiple comparison correction, indicating the morphology-fucosylation relationship is concentrated along a single dominant axis.

CONCLUSION:
Deep morphology embeddings from routine H&E slides were significantly associated with CF and OAF in TNBC, demonstrating that tumor glycan remodeling may be inferred from computational histologic analysis. Given that OAF glycans are linked to tumor stage and survival and 80% of the TNBC N-glycome is fucosylated, approximating these signatures from H&E morphology carries substantial clinical implications. The association between PC1 and fucosylation likely reflects differences in tumor cellularity, stromal composition, immune infiltration, and nuclear architecture, all known to covary with glycan remodeling. Future work will pursue spatially resolved validation correlating tile-level embeddings with localized glycan intensity maps from MALDI-MSI, and validation in larger multi-institutional TNBC datasets to assess generalizability and prognostic utility. Once validated, this approach could enable scalable glycobiologic phenotyping from routine pathology slides for TNBC risk stratification, patient selection for fucosylation-targeted therapies, and point-of-care diagnostic applications.

REFERENCES:
1. Scott DA, Seal S, Spruill L, Dunne J, Colditz G, Mehta AS, Drake RR, Ford ME, Angel PM. Fucosylated N-Glycan Landscape of Triple-Negative Breast Cancer. Mol Cancer Res. 2025 Dec 2;23(12):984-996. doi: 10.1158/1541-7786.MCR-25-0483. PMID: 40965337; PMCID: PMC12670077.
2. Drake RR, West CA, Mehta AS, Angel PM. MALDI Mass Spectrometry Imaging of N-Linked Glycans in Tissues. Adv Exp Med Biol. 2018;1104:59-76. doi: 10.1007/978-981-13-2158-0_4. PMID: 30484244.
3. Filiot, A., Jacob, P., Mac Kain, A. and Saillard, C., 2024. Phikon-v2, a large and public feature extractor for biomarker prediction. arXiv preprint arXiv:2409.09173.