High-performance deep learning pipeline predicts individuals in mixtures of DNA using sequencing data
(impact factor: 13.994, journal ranking: 2%)
In this study, we proposed a deep learning (DL) model for classifying individuals from mixtures of DNA samples using 27 short tandem repeats and 94 single nucleotide polymorphisms obtained through massively parallel sequencing protocol. The model was trained/tested/validated with sequenced data from 6 individuals and then evaluated using mixtures from forensic DNA samples. The model successfully identified both the major and the minor contributors with 100% accuracy for 90 DNA mixtures, that were manually prepared by mixing sequence reads of 3 individuals at different ratios. Furthermore, the model identified 100% of the major contributors and 50–80% of the minor contributors in 20 two-sample external-mixed-samples at ratios of 1:39 and 1:9, respectively. To further demonstrate the versatility and applicability of the pipeline, we tested it on whole exome sequence data to classify subtypes of 20 breast cancer patients and achieved an area under curve of 0.85. Overall, we present, for the first time, a complete pipeline, including sequencing data processing steps and DL steps, that is applicable across different NGS platforms.We also introduced a sliding window approach, to overcome the sequence length variation problem of sequencing data, and demonstrate that it improves the model performance dramatically.
Prediction of Breast Cancer Recurrence Using a Deep Convolutional Neural Network Without Region-of-Interest Labeling
(impact factor: 5.738, journal ranking: 32%)
We proposed a highly versatile two-step transfer learning pipeline for predicting the gene signature defining the intrinsic breast cancer subtypes using unannotated pathological images. Deciphering breast cancer molecular subtypes by deep learning approaches could provide a convenient and efficient method for the diagnosis of breast cancer patients. It could reduce costs associated with transcriptional profiling and subtyping discrepancy between IHC assays and mRNA expression. Four pretrained models such as VGG16, ResNet50, ResNet101, and Xception were trained with our in-house pathological images from breast cancer patient with recurrent status in the first transfer learning step and TCGA-BRCA dataset for the second transfer learning step. Furthermore, we also trained ResNet101 model with weight from ImageNet for comparison to the aforementioned models. The two-step deep learning models showed promising classification results of the four breast cancer intrinsic subtypes with accuracy ranging from 0.68 (ResNet50) to 0.78 (ResNet101) in both validation and testing sets. Additionally, the overall accuracy of slide-wise prediction showed even higher average accuracy of 0.913 with ResNet101 model. The micro- and macro-average area under the curve (AUC) for these models ranged from 0.88 (ResNet50) to 0.94 (ResNet101), whereas ResNet101_imgnet weighted with ImageNet archived an AUC of 0.92. We also show the deep learning model prediction performance is significantly improved relatively to the common Genefu tool for breast cancer classification. Our study demonstrated the capability of deep learning models to classify breast cancer intrinsic subtypes without the region of interest annotation, which will facilitate the clinical applicability of the proposed models.
Predicting Breast Cancer Gene Expression Signature by Applying Deep Convolutional Neural Networks From Unannotated Pathological Images
(impact factor: 5.738, journal ranking: 32%)
Purpose: The present study aimed to assign a risk score for breast cancer recurrence based on pathological whole slide images (WSIs) using a deep learning model.
Methods: A total of 233 WSIs from 138 breast cancer patients were assigned either a low-risk or a high-risk score based on a 70-gene signature. These images were processed into patches of 512x512 pixels by the PyHIST tool and underwent color normalization using the Macenko method. Afterward, out of focus and pixelated patches were removed using the Laplacian algorithm. Finally, the remaining patches (n=294,562) were split into 3 parts for model training (50%), validation (7%) and testing (43%). We used 6 pretrained models for transfer learning and evaluated their performance using accuracy, precision, recall, F1 score, confusion matrix, and AUC. Additionally, to demonstrate the robustness of the final model and its generalization capacity, the testing set was used for model evaluation. Finally, the GRAD-CAM algorithm was used for model visualization.
Results: Six models, namely VGG16, ResNet50, ResNet101, Inception_ResNet, EfficientB5, and Xception, achieved high performance in the validation set with an overall accuracy of 0.84, 0.85, 0.83, 0.84, 0.87, and 0.91, respectively. We selected Xception for assessment of the testing set, and this model achieved an overall accuracy of 0.87 with a patch-wise approach and 0.90 and 1.00 with a patient-wise approach for high-risk and low-risk groups, respectively.
Conclusions: Our study demonstrated the feasibility and high performance of artificial intelligence models trained without region-of-interest labeling for predicting cancer recurrence based on a 70-gene signature risk score.