HiC4D

HiC4D predicts future spatiotemporal Hi-C contact matrices to model genome dynamic reorganization.

Key Features:

• ResConvLSTM architecture: Implements a Residual Convolutional Long Short-Term Memory (ResConvLSTM) network that combines residual networks with ConvLSTM.
• Residual connections: Integrates residual connections to mitigate the vanishing gradient problem in deep recurrent convolutional models.
• Spatiotemporal modeling: Uses ConvLSTM components to capture spatial and temporal dependencies within Hi-C contact matrices.
• Benchmarking against multiple models: Evaluated against NaiveNet and four video-prediction models: ConvLSTM, ST-LSTM (spatiotemporal LSTM), SA-LSTM (self-attention LSTM), and SimVP.
• Dataset diversity: Benchmarked on eight spatiotemporal Hi-C datasets, including two from mouse embryogenesis, one from somatic cell nuclear transfer (SCNT) embryos, three from different species' embryogenesis, and two non-embryogenesis datasets.
• Performance outcomes: ResConvLSTM networks consistently outperformed the compared methods across blind-test datasets in predicting future Hi-C contact matrices.
• TAD recovery: All evaluated methods were able to delineate topologically associating domains (TADs) within experimental Hi-C data.

Scientific Applications:

• Genomic architecture dynamics: Forecasting spatiotemporal Hi-C data to study genome reorganization and chromosomal interaction dynamics.
• Developmental biology: Applying predicted Hi-C matrices to investigate embryogenesis across species and somatic cell nuclear transfer (SCNT) embryo development.
• Chromatin domain analysis: Using forecasts to analyze topologically associating domains (TADs) and their changes relevant to development and disease.

Methodology:

HiC4D implements a Residual ConvLSTM (ResConvLSTM) architecture that integrates residual connections with ConvLSTM to capture spatial and temporal dependencies and was benchmarked against NaiveNet, ConvLSTM, ST-LSTM, SA-LSTM, and SimVP on eight spatiotemporal Hi-C datasets.