Point people/organizers: Jian Ma, Katie Pollard, Christina Leslie, Yun Li

The overarching goal of this Working Group is to foster collaborations in the 4DN Consortium for the development of predictive models to reveal fundamental molecular mechanisms that connect structure and function of the 4D Nucleome. We expect that this Working Group will coordinate consortium-wide activities related to computational model development for mechanistic understanding of nuclear structure and function, maximize the complementarity of different data modalities from various biological systems, assess the impact of nuclear structure in human health and diseases, and facilitate coordinated outreach in computational development to other NIH Consortia and the broad community. The deliverables from this Working Group will include best practices for the 4DN community, new methods, common datasets, and collaborative publications. Initial discussions from the Working Group will include the following topics.

• Predictive model development and benchmark datasets for mechanistic understanding of the nuclear organization

• Learning the roles of sequence, epigenomic signatures, RNAs, and trans-factors on nuclear structure features and their formation mechanisms (e.g., loop extrusion, phase separation)

• Learning the context-dependent connections between 3D genome structure (including its dynamics) and vital genome functions (e.g., transcription, DNA replication, DNA repair, recombination, X-inactivation)

• Interpreting the cell type-specific impact of genetic variants on nuclear structure and function, with additional integration with medical genetics datasets in health and diseases

• Assessing the phenotypic impact of 3D genome organization at various levels (i.e., molecular, cellular, and organism)

• Benchmark loci and datasets for experimental validations

1. Predictive models for multiscale navigable 4D Nucleome maps to reveal nuclear organization mechanisms

1. Integration of multimodal datasets, including genomic assays and imaging methods

1. Integration of single-cell 3D genome measurements for mechanistic understanding of cell-to-cell variability

Integration of perturbation datasets to uncover causal relationships of structure and function

Determining the most informative set of data types for various 4DN maps in multiple scales

Integrative modeling for realistic nuclear structures and dynamics

Best practices for predictive machine learning model development

  Model selection, formulation, and transferring, e.g. neural network architectures

Model evaluation strategies to enhance rigor and reproducibility

Model interpretability, generalizability, visualization, and expressiveness to reveal important biological features

Experimental design from predictive modeling

Optimizing consortium-wide shared resources by using predictive models to prioritize loci or perturbations that will maximize discovery

Predicting missing data sets (e.g., cell types, assays) that, if collected, would enhance prediction accuracy or our ability to derive mechanisms from modeling

Guiding sound experimental design across consortium projects (e.g., for controlling batch effects or maximizing power).