HiC4D: forecasting spatiotemporal Hi-C data with residual ConvLSTM

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Hi-C experiments are crucial for studying genomic structures and genome dynamics.
• Spatiotemporal Hi-C data provides insights into genome reorganization over time.
• Computational modeling and forecasting of spatiotemporal Hi-C data are currently lacking.

Purpose of the Study:

• To present HiC4D, a novel computational tool for forecasting spatiotemporal Hi-C data.
• To introduce and evaluate a new deep learning network, residual ConvLSTM (ResConvLSTM), for this task.

Main Methods:

• Developed a novel ResConvLSTM network, combining residual networks and convolutional long short-term memory.
• Benchmarked ResConvLSTM against five other methods, including NaiveNet, ConvLSTM, ST-LSTM, SA-LSTM, and SimVP.
• Utilized eight diverse spatiotemporal Hi-C datasets for blind testing, covering embryogenesis and other biological processes.

Main Results:

• ResConvLSTM consistently outperformed other methods in accurately predicting future Hi-C contact matrices across eight blind-test datasets.
• All benchmarked methods successfully identified topologically associating domain boundaries from experimental Hi-C data.
• HiC4D demonstrated effectiveness as a tool for predicting spatiotemporal Hi-C data.

Conclusions:

• HiC4D, powered by ResConvLSTM, is an effective tool for forecasting spatiotemporal Hi-C data.
• The developed method advances computational approaches for understanding genome dynamics.
• HiC4D is publicly available for research use.