HAVOC: Small-scale histomic mapping of cancer biodiversity across large tissue distances using deep neural networks

Abstract

Intratumoral heterogeneity can wreak havoc on current precision medicine strategies because of challenges in sufficient sampling of geographically separated areas of biodiversity distributed across centimeter-scale tumor distances. To address this gap, we developed a deep learning pipeline that leverages histomorphologic fingerprints of tissue to create "Histomic Atlases of Variation Of Cancers" (HAVOC). Using a number of objective molecular readouts, we demonstrate that HAVOC can define regional cancer boundaries with distinct biology. Using larger tumor specimens, we show that HAVOC can map biodiversity even across multiple tissue sections. By guiding profiling of 19 partitions across six high-grade gliomas, HAVOC revealed that distinct differentiation states can often coexist and be regionally distributed within these tumors. Last, to highlight generalizability, we benchmark HAVOC on additional tumor types. Together, we establish HAVOC as a versatile tool to generate small-scale maps of tissue heterogeneity and guide regional deployment of molecular resources to relevant biodiverse niches.

Fig. 1.. Mapping biodiversity in cancer tissue with HAVOC.

(A) Cartoon depicting regional evolution of intratumoral subclones and a map-guided approach to sampling. (B) HAVOC workflow summary. A pretrained CNN is used as a morphology feature extractor for input images. The generated DLFVs of individual image patches (tiles) are then used to carry out image feature-based clustering and map spatial coordinates to distinct histomic fingerprints back to the original WSI. (C to E) Representative mapping of a diffuse glioma. The relative spectrum of morphologic patterns of image tiles can be explored by dimensionality reduction (e.g., t-SNE) and clustering. Both highlight distinct tumor (blue, purple, and cyan) and nontumoral (yellow, green, and red) regions. Scale bar, 6 mm. (F) Strip boxplot of Mask R-CNN–based quantifications of nuclei per HAVOC region. The boxplot shows minimum, first quartile, median, third quartile, and maximum along with outliers (diamonds). Counts represent nuclear instances per 4218 μm². (G) Mapping of individual DLFs overrepresented in these HAVOC-defined tumor regions highlights interpretable morphological patterns that correlate with these defined glioma niches (e.g., tumor nuclei and edema, respectively). Dimensions of image patches shown: 0.27 mm². (H) Pairwise Pearson correlation coefficients (r) of the DLFVs of HAVOC-defined partitions in (E), highlighting inverse correlation with the degree of morphological heterogeneity across this representative case. Sample pairs of mild/moderate/severe correlations are provided. (I) Violin plots showing concordance between HAVOC r values and semiquantitative assessments of regional heterogeneity (e.g., mild, moderate, or severe). DLFV r of lesional versus nonlesional regions included as reference.

Fig. 2.. HAVOC-defined partitions align with regional biodiversity.

(A) Histopathology images of an IDH–wild-type glioblastoma demonstrating a BRAFV600E-mutated hyperproliferative subclone resolved by immunohistochemistry. This molecularly distinct subpopulation was also resolved by HAVOC (yellow partition). Scale bar, 2 mm. (B to E) LC-MS/MS profiling of HAVOC-defined partitions and bulk tissue samples shows distinct proteomic profiles. (B) HAVOC delineated four distinct regions in this glioblastoma section. The two large cellular tumor regions (red/yellow) were excised and analyzed by LC-MS/MS. Scale bars, 4 mm. (C) Hierarchical clustering of the resolved regional proteomes highlights differences in protein expression of HAVOC partitions and bulk samples (n = 3 replicate tissue sections). (D) Normalized ssGSEA scores of 64 GBM-informative signatures (45) across the different HAVOC regions and bulk tissue. (E) Boxplots more specifically highlighting regional differences in the functional status of three key protein-level programs in glioblastoma [MYC_targets (proliferation), KRAS_targets (invasion), and Hypoxia (33)]. Bulk signatures provided for reference. Enrichment scores were inferred with existing XGBoost models, and the dotted line represents average values from the training GBM cohort (33). (F) Spatial proliferation index differences (as assessed via CNN-based quantification of Ki-67) spatially align with the different HAVOC partitions. (G) Feature activation mapping defined morphologic correlates of the profiled HAVOC partitions. Partitions showing proliferative and invasive biology show high cellularity and cytoplasmic/fibrillary patterns, respectively. Dimensions of image patches shown: 0.27 mm². Sample masks from the Mask R-CNN cellularity analysis provided for reference (dimension: 4218 μm²). (H) Strip boxplot of Mask R-CNN–based quantifications of nuclei/HAVOC region. The boxplot shows minimum, first quartile, median, third quartile, and maximum along with outliers (diamonds). Counts represent nuclear instances per 4218 μm². (I) Representative H&E and HAVOC map of an IDH–wild-type glioblastoma section, with geographically separated subclusters (e.g., O1, O2, and O3) showing similarly grouped histomic (HAVOC) signatures. Scale bars, 2 mm.

Fig. 3.. Small-scale mapping of biovariation across an entire tumor specimen using HAVOC.

(A) Twelve sequential H&E sections generated from a recurrent IDH-mutated, 1p19q- codeleted oligodendroglioma measuring 5.4 cm × 4.1 cm × 1.8 cm in dimensions with accompanying HAVOC heterogeneity maps. Scale bar, 4 mm. The assigned slide color scheme is arbitrary. (B) Pairwise Pearson correlation matrix arranging all 84 clusters shown in (A) in a hierarchical fashion. Slide colors correspond to the same colors shown in the former panel. There is significant distribution enrichment of human-annotated morphologic patterns within different subclusters defined by the hierarchical tree. A seven-cluster solution is shown with additional solutions offered in fig. S10. There was no significant enrichment of a specific slide in any of the proposed clusters (chi-square test; please see table S3 for contingency tables and associated P values) (C) Representative images of Ki-67 (MIB1) from the different clustered regions. Image patch border colors correspond to their originating WSI Slide color assigned in (A). Dimensions of image patches shown: 0.27 mm². (D) Histogram of estimated Ki-67 (MIB1) across different histologically defined regions (P values generated using Mann-Whitney U test). Note: Color of histogram bars aligns with the color legend shown in (B).

Fig. 4.. HAVOC mapping across related WSI pairs aligns with overall molecular correlations.

(A and B) Paired H&E WSI sections from two independent individuals with IDH–wild-type high-grade gliomas. (C and D) Respective HAVOC partitions across respective paired slides. (E) Multislide DLFV correlation matrices across both slides of patient II define a distinct focal hyperdense outlier region (sample IIb blue cluster). These multislide map distances are in agreement with global patterns of proteomic variations derived from each of the major HAVOC-defined partitions. (F) Another example of HAVOC multislide regional partitions that identified reciprocally aligned regions across slide pairs. These HAVOC-defined interslide niche similarities are also in agreement with the intratumoral proteomic variations of this specimen. (G and H) Representative image patches from HAVOC partitions provided to highlight that the magnitude of the qualitative morphological differences is in agreement with the relative HAVOC and molecular differences. Dimensions of image patches shown: 0.27 mm².

Fig. 5.. HAVOC reveals spatially organized patterns of molecular heterogeneity in high-grade gliomas.

(A) Unsupervised analysis of 19 HAVOC-defined tumor regions across 6 high-grade gliomas by performing UMAP dimensional reduction of the ssGSEA scores of 64 proteogenomically concordant gene sets previously described in glioblastoma (33). (B) Inverse relationship between the ssGSEA scores of the embryonic stem cell state (Ben Porath ES_1) and the astrocytic differentiation gene set (Lein Astrocyte Markers Gene Set). P value generated by linear fit model. (C) Regional differences in state of differentiation (Astro-ES axis) across each of the spatially resolved and profiled slides. The level of statistical significance of the differences between the regions of each tissue slide was assessed by ANOVA; P values are indicated as follows: ***P < 0.001, **P < 0.01, and *P < 0.05; n.s., not significant. (D and E) Distribution of the Astro-ES axis in astrocytic tumors (non–1p19q-codeleted) from TCGA-GBM and TCGA-LGG cohorts stratified by WHO grade (D) and IDH status (E) (P values generated by ANOVA). (F) Varying distributions along the Astro-ES axis at the single-cell level in patient-derived GBM cells (n = 3) and glioma stem cells (n = 1); dataset previously published by Richards et al. (41).

Fig. 6.. HAVOC defines genetically distinct metastatic clones.

(A) Schematic and H&E-stained section of a mouse liver modeling polyclonal Kras^G12D/+Trp53^−/− lung cancer metastases from Zhao et al. (6). Spatial profiling in the boxed region allowed benchmarking of HAVOC in this model. Scale bar, 5 mm. (B) Slide–DNA-seq and slide–RNA-seq of focused region provided ground truth of normal liver, tumor clone “A” and “B.” Black lines in the right panel indicate tumor boundaries. Scale bar, 2 mm. (C) HAVOC mapping of entire H&E sections with 11 partitions to ensure saturation of the different histomorphological patterns. HAVOC segmented all five tumors into stable groupings, identified in pink, teal, and purple. (D) HAVOC partitions (k = 4; tile width: 128 pixels) of the focused region mapped by slide–DNA-seq (original study) spatially divided tumor into two homogeneous subclones (green versus blue). Surrounding liver tissue was also separated into regions of peritumoral liver and immune infiltrates (red versus yellow) that match single-cell RNA-seq projection from the original study. Scale bar, 1 mm. Note: Color IDs of regions in (C) and (D) are unrelated due to the independent unsupervised nature of these analyses. (E) FAMs of selected differentially activated DLFs for each tumor region (a and b) and DLFs enriched in the peritumoral region highlighting liver regions with and without inflammatory cells (c and d). Note: DLF219 was enriched in peritumoral regions with inflammation and so is therefore specifically localized to the respective lower portion of the slide. Dimensions of image patches shown: 0.27 mm².

Fig. 7.. HAVOC generalizes to untrained tissue types.

(A) (a) H&E-stained section of a dermatological “collision” tumor composed of distinct regions of squamous cell and neuroendocrine carcinoma. (b) HAVOC partitions resolved the distinct tumor types that matched the immunohistochemical ground truths. (c and d) CK34BE12 and synaptophysin highlight respective regions of squamous and neuroendocrine carcinoma. Scale bar, 2 mm. (B) Representative high-power magnification micrographs of HAVOC-defined distinct tumoral subclones highlighting the squamous (green partition) and neuroendocrine (blue partition) components. Scale bar, 200 μm. (C) (a) H&E-stained lung resection and (b) HAVOC map of a neoplasm showing divergent adenosquamous differentiation. Tumor subcomponents are highlighted with distinct (c) TTF1 (adenocarcinoma) and (d) p40 (squamous) immunohistochemical staining. HAVOC partition map showing subregions that align with the distinct tumoral patterns defined by the squamous/adenocarcinoma markers. Scale bars, 6 mm. (D) Hierarchical clustering of HAVOC-defined regions and representative high-power images of (a) H&E, (d) p40, and (c) TTF1-stained images reveals two distinct tumor subpatterns. The divergent p40⁺ squamous subregion has the lowest DLFV r value among these tumor regions. The colors under immunohistochemistry images correspond to the color codes of the HAVOC map (b). Scale bars, 200 μm.