A Pan-Cancer Patient-Derived Xenograft Histology Image Repository with Genomic and Pathologic Annotations Enables Deep Learning Analysis

Abstract

Patient-derived xenografts (PDX) model human intra- and intertumoral heterogeneity in the context of the intact tissue of immunocompromised mice. Histologic imaging via hematoxylin and eosin (H&E) staining is routinely performed on PDX samples, which could be harnessed for computational analysis. Prior studies of large clinical H&E image repositories have shown that deep learning analysis can identify intercellular and morphologic signals correlated with disease phenotype and therapeutic response. In this study, we developed an extensive, pan-cancer repository of >1,000 PDX and paired parental tumor H&E images. These images, curated from the PDX Development and Trial Centers Research Network Consortium, had a range of associated genomic and transcriptomic data, clinical metadata, pathologic assessments of cell composition, and, in several cases, detailed pathologic annotations of neoplastic, stromal, and necrotic regions. The amenability of these images to deep learning was highlighted through three applications: (i) development of a classifier for neoplastic, stromal, and necrotic regions; (ii) development of a predictor of xenograft-transplant lymphoproliferative disorder; and (iii) application of a published predictor of microsatellite instability. Together, this PDX Development and Trial Centers Research Network image repository provides a valuable resource for controlled digital pathology analysis, both for the evaluation of technical issues and for the development of computational image-based methods that make clinical predictions based on PDX treatment studies. Significance: A pan-cancer repository of >1,000 patient-derived xenograft hematoxylin and eosin-stained images will facilitate cancer biology investigations through histopathologic analysis and contributes important model system data that expand existing human histology repositories.

Figure 1.

PDXNet image repository captures heterogeneity across tumor types, passages, and treatment status and pairs human and PDX histology images with pathologic assessment and genomic and transcriptomic data. A, Top, distribution of the number of PDX models with H&E images by cancer diagnosis. Fraction of PDX models of a given diagnosis paired with HRD and MSI status, slide-level pathologic assessment (Path), WGD predictions, and RNA-seq, Middle, TMB, tumor volume, and WES data. Bottom, fraction of PDX models with paired human data of the indicated modality. B, Distribution of PDX images by passage. C, Counts of images for those derived from human progenitor/PDX pairs, restricted to diagnoses with at least four such progenitors. B and C, Bar charts are colored according to the contributing site as indicated in A. D, Distribution of PDX images across diagnoses with at least two treated PDXs or at least two treated progenitors. BLCA, bladder cancer; HNSC, head and neck squamous carcinoma;NRSTS, nonrhabdomyosarcoma soft tissue sarcoma; Rx, treatment.

Figure 2.

Cell segmentation and classification trained on human data correlate with pathologist estimates of cell-type proportions in PDX samples. A and B, Correlation of pathologist (x-axis) and HoVer-Net–based slide-level estimates (y-axis) for neoplastic (left), stromal (middle), and necrotic (right) cell proportions incolorectal cancer (A) andPDAC (B) from WUSTL. Each point corresponds to an individual image and PDX (i.e., no replication). C, Weighted Pearson correlations of pathologist and HoVer-Net–based estimates of neoplastic, stromal, and necrotic cell proportions in colon adenocarcinoma, colorectal cancer, and PDAC in the PDMR and WUSTL datasets. All PDX strains are NSG. Colorectal cancer not otherwise specified, for example, as colon adenocarcinoma or rectal adenocarcinoma. Correlations are computed over images, inversely weighted by the number of H&E images from each patient (see Materials and Methods).

Figure 3.

Regional classifier predicts cancer epithelium (“tumor”), stroma, and necrotic areas based on nuclei classification. A and B, Cell phenotypes predicted by HoVer-Net (A) or HD-Staining (B) within a PDX lung tumor. C, Labeled tiles (orange, necrosis; blue, stroma; black, tumor) represented according to the cellular fraction that is predicted by HoVer-Net to be neoplastic, necrotic, immune, or stromal. D, Top, confusion matrix comparing pathologist-provided tile labels with corresponding prediction from the random forest (RF) HoVer-Net classifier. Results shown for held-out set following fivefold cross-validation, with folds defined according to the patient/progenitor (see Materials and Methods). Bottom, image and tile count used for cross-validation stratified by diagnosis and site. E, Pathologist regional annotations. F, Random forest HoVer-Net–based classifier predictions of regions (within 512 × 512 pixel tiles). necros, necrotic; neopla, neoplastic; nolabe, unlabeled; no-neo, nonneoplastic epithelial; RBC, red blood cell.

Figure 4.

XTLD can be computationally predicted from H&E images using segmented and classified cells or DL features. A and B, H&E images of LUSC (A) and XTLD (B). Each image from a distinct human progenitor. Images at 20× resolution. Scale bars, 50 µm. C, Image and tile count used for cross-validation stratified by diagnosis and site. D and E, HoVer-Net–based classifier performance. D, Distribution of tile-level XTLD probability predicted by random forest trained using HoVer-Net–derived features, according to the PDX image (x-axis), from an LUSC (red) or XTLD (blue) sample in the MDACC (solid line) or PDMR (dashed line) dataset. Results shown for held-out set following fivefold cross-validation, with folds defined according to the patient/progenitor (see Materials and Methods). E, Confusion matrix showing tile-level concordance between HoVer-Net–based classification (rows) and annotation (columns). F and G, Inception-based classifier performance, as in D and E, respectively, but with predictions from a random forest trained using Inception features. H, Distribution of tile-level XTLD probability predicted by random forest trained using Inception features, according to the PDX image (x-axis), from LUSC (red) or XTLD (blue) samples in JAX external validation dataset. Each image is from a distinct human progenitor. Samples are ordered by median XTLD prediction probability over tiles in D, F, and H. All PDXs are NSG mice.

Figure 5.

Inception encodes morphologic features useful in predicting XTLD. A, Tiles with large values of Inception feature 1895. The tile with the largest value is shown for each of four samples (noted above tile) with highest values, all of which are XTLD cases. B, Tiles with small values of Inception feature 1895. The tile with the smallest value is shown for each of four samples (noted above tile) with smallest values, all of which are LUSC cases. All PDXs are NSG mice. Images at 20× resolution. Scale bars, 50 µm.

Figure 6.

MSI can be computationally predicted from PDX samples using a human-trained model. Distribution of tile-level MSI probability predicted using a published model (34) from human progenitor (A) or PDX (B) model H&E images derived from MSI (red) or MSS (blue) samples. All progenitor samples in A have corresponding derived PDX samples in B. Samples ordered by median (mean over models) MSI prediction over tiles.