graph LR
Input_Data_Pipeline["Input Data Pipeline"]
Neural_Network_Model["Neural Network Model"]
Model_Training["Model Training"]
Variant_Inference["Variant Inference"]
Output_Post_processing["Output Post-processing"]
Input_Data_Pipeline -- "provides training data to" --> Model_Training
Input_Data_Pipeline -- "provides inference data to" --> Variant_Inference
Model_Training -- "trains" --> Neural_Network_Model
Variant_Inference -- "uses" --> Neural_Network_Model
Variant_Inference -- "sends raw calls to" --> Output_Post_processing
Output_Post_processing -- "provides refined data to" --> Input_Data_Pipeline
click Input_Data_Pipeline href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main//neusomatic/Input Data Pipeline.md" "Details"
click Neural_Network_Model href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main//neusomatic/Neural Network Model.md" "Details"
click Model_Training href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main//neusomatic/Model Training.md" "Details"
click Variant_Inference href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main//neusomatic/Variant Inference.md" "Details"
click Output_Post_processing href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main//neusomatic/Output Post-processing.md" "Details"
The NeuSomatic architecture is designed for somatic variant calling, primarily focusing on long-read sequencing data. The main flow involves an Input Data Pipeline that prepares and loads genomic data, which then feeds into either Model Training to train the Neural Network Model, or directly into Variant Inference for prediction. The Variant Inference component generates initial variant calls, which are then refined and finalized by the Output Post-processing component. A notable aspect is the feedback loop from Output Post-processing back to the Input Data Pipeline, suggesting iterative refinement or specialized data preparation based on initial variant analysis, particularly for indel realignment.
Manages the entire process of preparing raw input data, generating structured datasets, and loading them efficiently for both model training and inference. This includes initial data preprocessing, dataset creation from various genomic sources, and batching for the neural network.
Related Classes/Methods:
neusomatic.neusomatic.python.preprocess.preprocess(198:339)neusomatic.neusomatic.python.generate_dataset.generate_dataset(1401:1593)neusomatic.neusomatic.python.dataloader.NeuSomaticDataset(125:425)
Defines the core convolutional neural network architecture (NeuSomaticNet) used by NeuSomatic for variant prediction. It encompasses the network's layers, building blocks, and the forward pass computation.
Related Classes/Methods:
Handles the training lifecycle of the NeuSomatic model. This includes initializing the neural network, loading training data, balancing classes, defining loss functions, optimizing the model parameters, and evaluating its performance.
Related Classes/Methods:
Executes the core variant calling process by applying the trained NeuSomatic model to prepared input data. It performs inference and generates initial VCF records based on the model's predictions.
Related Classes/Methods:
Manages the final stages of variant call refinement and output generation. This component orchestrates complex variant resolution (e.g., for short reads or long-read indels), merges VCFs, and adds supplementary information to produce the final, high-quality VCF output.
Related Classes/Methods: