Deep learning identifies morphological patterns of homologous recombination deficiency in luminal breast cancers from whole slide images

Abstract

Homologous recombination DNA-repair deficiency (HRD) is becoming a well-recognized marker of platinum salt and polyADP-ribose polymerase inhibitor chemotherapies in ovarian and breast cancers. While large-scale screening for HRD using genomic markers is logistically and economically challenging, stained tissue slides are routinely acquired in clinical practice. With the objectives of providing a robust deep-learning method for HRD prediction from tissue slides and identifying related morphological phenotypes, we first show that digital pathology workflows are sensitive to potential biases in the training set, then we propose a method to overcome the influence of these biases, and we develop an interpretation method capable of identifying complex phenotypes. Application to our carefully curated in-house dataset allows us to predict HRD with high accuracy (area under the receiver-operator characteristics curve 0.86) and to identify morphological phenotypes related to HRD. In particular, the presence of laminated fibrosis and clear tumor cells associated with HRD open new hypotheses regarding its phenotypic impact.

Keywords: bias; breast cancer; computational pathology; deep learning; homologous recombination deficiency; interpretability; molecular subtype; prediction; self-supervised learning; whole slide images.

Graphical abstract

Figure 1

From WSI to prediction Four major components are used in this end-to-end pipeline. First, the WSIs (x) are tiled, the tissue parts are automatically selected, and the resulting tiles are embedded into a low-dimensional space (block 1). The embedded tiles are then scored through the attention module (2). An aggregation module outputs the slide-level vector representative (3) that is finally fed to a decision module (4), which outputs the final prediction. When training, the binary cross-entropy loss between the ground truth $y$ and the prediction $\hat{y}$ is computed and back-propagated to update the parameters of the modules. Both the decision module and the attention module are multilayer perceptrons, the encoder is a ResNet18, and the aggregation module consists of a weighted sum of the tiles, the weights being the attention scores.

Figure 2

Bias corrections and prediction performances (A and B) Estimation of the bias score of two technical confounders ($c_1,c_2$) and one biological confounder ($c_3$) for the Curie dataset (A) and the bias score of the confounder $c_3$ for TCGA dataset (B) for different correction strategies. A Mann-Whitney-Wilcoxon test, two-sided with Bonferroni correction, is performed for each pair of correction strategies. As detailed in STAR Methods, for each correction strategy a series of 30 unbiased subtest sets are sampled on which the model’s bias is evaluated. Error bars indicate standard deviations over the subtest sets. The significance test is performed on this distribution of 30 estimations. The bias score of a model is the average of this distribution. ns, not significant (p > 0.05); ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 1 × 10⁻³, ∗∗∗∗p < 1 × 10⁻⁴. (C) Receiver-operating characteristic curves. The name of each model indicates the origin of its training set. Indices indicate the correction applied through strategic sampling (${\text{Curie}}_{c_1}$ has been debiased with respect to $c_1$). Curie_luminals corresponds to the model trained on a subset containing only luminal tumors.

Figure 3

Decision-based visualization (A) Mechanism of the decision-based visualization. 1: each tile in the whole dataset is scored by the attention module. 2: per slides, the 300 best scoring tiles are selected as candidate tiles. 3: the selected tiles are presented to the decision module, and the logit of the probability of each of these tiles being HRD or HRP (yellow or green) is kept. 4: finally, the K tiles with maximal probability for either HRD/HRP are selected. (B) Morphological map of the HR status in the luminal BC cohort. Each dot is the uniform manifold approximation and projection (UMAP) of a tile extracted by the decision-based visualization method. Crosses (circles) are tiles with high HRD (HRP) logit. Each cluster has been linked to a morphological phenotype by two expert pathologists. We identified six different morphological phenotypes associated with the HRD and two associated with the HRP. The exhibited tiles have been randomly sampled among each cluster. 228 slides contributed to the HRP clusters and 232 to the HRD cluster. In total, 249 among 251 slides contributed to the whole figure. The same protocol has been applied to the public datasets TCGA breast invasive carcinoma (BRCA), TCGA BRCA-TNBC, and TCGA ovarian cancer (see Figures S3, S4, and S5, respectively). Scale bars, 100 μm. (C) Pathological interpretation of the clusters presented in (B).

Figure 4

Illustration of two phenotypic HRD-ness trajectories (A) UMAP projection of the HR status-specific representation of the meaningful tiles relative to the HRD. HRD-ness is the score given to each tile by the HRD output neuron. Two tile trajectories have been extracted (blue and magenta) starting from the same low HRD-ness region, each leading to a different high HRD-ness region. (B and C) Tiles sampled along each of the trajectories. These are ordered from low HRD-ness to high HRD-ness and read from left to right and from top to bottom. Scale bars, 100 μm. (B) Magenta trajectory, toward densely cellular tumors or inflammatory cells. (C) Blue trajectory, toward fibroinflammatory tumor changes and hemorrhagic suffusions.