Biologically informed deep neural networks provide quantitative assessment of intratumoral heterogeneity in post treatment glioblastoma

Abstract

Intratumoral heterogeneity poses a significant challenge to the diagnosis and treatment of recurrent glioblastoma. This study addresses the need for non-invasive approaches to map heterogeneous landscape of histopathological alterations throughout the entire lesion for each patient. We developed BioNet, a biologically-informed neural network, to predict regional distributions of two primary tissue-specific gene modules: proliferating tumor (Pro) and reactive/inflammatory cells (Inf). BioNet significantly outperforms existing methods (p < 2e-26). In cross-validation, BioNet achieved AUCs of 0.80 (Pro) and 0.81 (Inf), with accuracies of 80% and 75%, respectively. In blind tests, BioNet achieved AUCs of 0.80 (Pro) and 0.76 (Inf), with accuracies of 81% and 74%. Competing methods had AUCs lower or around 0.6 and accuracies lower or around 70%. BioNet's voxel-level prediction maps reveal intratumoral heterogeneity, potentially improving biopsy targeting and treatment evaluation. This non-invasive approach facilitates regular monitoring and timely therapeutic adjustments, highlighting the role of ML in precision medicine.

Fig. 1. Overview of the application of BioNet in assisting the assessment of treatment responses and informing subsequent therapy decisions.

Our datasets comprise two types of data: sparsely labeled data from biopsy locations, and abundant unlabeled data from all locations throughout the entire brain. The labels for biopsy samples ($y_i$) are obtained through comprehensive transcriptomic and immunohistochemical profiling. The input features ($x_i$) for all samples, both labeled and unlabeled, which are utilized in the training and testing of BioNet, are extracted from multiparametric MRI. Within the tumoral Area of Interest (AOI) of each patient (blue outline), local regions (small squares) were created by sliding windows according to the physical size of surgical biopsies. Statistical and texture features $\left(x_i\right)$ were computed based on multiparametric MRI within the sliding windows at biopsy locations (red, a few) and remaining unlabeled locations (yellow, abundant). Labeled samples, along with selectively chosen unlabeled samples, are employed in the training of BioNet. Once adequately trained, BioNet is capable of generating voxel-level prediction maps for proliferative/recurrent tumors (Pro) and treatment-induced reactive/inflammatory cells (Inf), respectively, within AOI. These prediction maps yield crucial insights into the gene status at the voxel level throughout the entire tumor.

Fig. 2. Overview of the biological relationships.

a Revealing biological relationships from domain knowledge to inspire BioNet design. b Bar chart for the group mean over the average scores of Pro and Inf in the group with high Neu in comparison to that in the group with low Neu. c Scatter plot of Pro and Inf for the group with low Neu. The Pearson correlation coefficient is represented by $r$. d Hierarchical design of BioNet inspired by two biological relationships (1) and (2) between Pro and Inf given the high/low status of Neu.

Fig. 3. Overall architecture of BioNet.

BioNet consists of two networks: BioNet_Neu to predict Neu using MRI; BioNet_ProInf to simultaneously predict Pro and Inf using MRI. a BioNet_Neu is a feedforward neural network pre-trained using a large number of unlabeled samples with noisy Neu labels informed by biological knowledge, and fine-tuned using biopsy samples with data augmentation. It also corporates Monte Carlo dropout to enable uncertainty quantification for the predictions. The role of BioNet_Neu is to stratify unlabeled samples with high predictive certainty, which were then incorporated into the training of BioNet_ProInf. b BioNet_ProInf is a multitask semi-supervised learning model with a custom loss function. The architecture consists of a shared block and task-specific blocks. The loss function combines a prediction loss and a knowledge attention loss that penalizes violation of the knowledge-based relationships on unlabeled samples.

Fig. 4. Performance of BioNet and competing methods evaluated by Classification Accuracy and Area Under the Curves (AUCs).

Results are shown for developmental cohort A under leave-one-patient-out cross-validation (LOPO CV) in a, and for test cohort B in b. c A table summarizes the key performance metrics of BioNet and competing methods. The standard deviation of accuracy for cohort A, calculated using leave-one-patient-out cross-validation (LOPO CV), is shown in brackets.

Fig. 5. Generalizability of BioNet_ProInf and competing methods.

a The patient’s MRI images. b Knowledge concordance (KC) metrics, denoted as ${\mathit{KC}}_{{\mathit{neu}}^{+}}$ and ${\mathit{KC}}_{{\mathit{neu}}^{-}}$, for the predictions on unlabeled samples with Neu high and Neu low, respectively from the tumoral area of interest (AOI) of each patient on developmental cohort A and test cohort B. Prediction maps of c recGBM_Pro and d recGBM_Inf within the tumoral AOI by BioNet and the best-performing ML and DL methods (color maps overlaid on (a)). Three purple boxes denote the locations of three biopsies on this MRI slice. The maps indicate that BioNet attains the lowest absolute errors in predicting both Pro and Inf.

Fig. 6. A flowchart of the biopsy acquisition procedure.

Biopsy acquisition in this study followed the same procedure that has been used in our prior publications^,,,.

Fig. 7. Defining tissue-specific gene modules to connect with key immunohistochemical features.

a Heatmap depicting correlation between normalized gene expression and immunohistochemical labeling indices, with subsequent hierarchical clustering revealed three orthogonal tissue-specific gene modules. Module 1 consists of 3688 genes significantly positively correlated with SOX2/Ki67/H&E; module 2 consists of 1673 genes correlated with NeuN; module 3 consists of 2418 genes correlated with CD68. b Bar plot depicting top significant gene ontologies enriched in each of the three tissue-specific gene modules derived from the IHC-RNAseq correlation analysis. X axis is –log₁₀ (p value) of each ontology. Module 1 is enriched in genes involved in proliferation (Pro), module 2 in neuronal-specific genes (Neu), and module 3 in genes in immune infiltration (Inf). c Heatmap depicting single-sample Gene Set Variation Analysis (GSVA) for each of 84 MRI-localized biopsies for each of the three tissue-specific gene modules. Color gradient represents magnitude/direction of tissue-specific enrichment for each biopsy.