We recently used in situ Hi-C to create kilobase-resolution 3D maps of mammalian genomes. Here, we combine these maps with new Hi-C, microscopy, and genome-editing experiments to study the physical structure of chromatin fibers, domains, and loops. We find that the observed contact domains are inconsistent with the equilibrium state for an ordinary condensed polymer. Combining Hi-C data and novel mathematical theorems, we show that contact domains are also not consistent with a fractal globule. Instead, we use physical simulations to study two models of genome folding. In one, intermonomer attraction during polymer condensation leads to formation of an anisotropic "tension globule." In the other, CCCTC-binding factor (CTCF) and cohesin act together to extrude unknotted loops during interphase. Both models are consistent with the observed contact domains and with the observation that contact domains tend to form inside loops. However, the extrusion model explains a far wider array of observations, such as why loops tend not to overlap and why the CTCF-binding motifs at pairs of loop anchors lie in the convergent orientation. Finally, we perform 13 genome-editing experiments examining the effect of altering CTCF-binding sites on chromatin folding. The convergent rule correctly predicts the affected loops in every case. Moreover, the extrusion model accurately predicts in silico the 3D maps resulting from each experiment using only the location of CTCF-binding sites in the WT. Thus, we show that it is possible to disrupt, restore, and move loops and domains using targeted mutations as small as a single base pair.genome architecture | molecular dynamics | CTCF | chromatin loops | CRISPR

No abstract

We introduce several new black-box reductions that significantly improve the design of adaptive and parameterfree online learning algorithms by simplifying analysis, improving regret guarantees, and sometimes even improving runtime. We reduce parameter-free online learning to online exp-concave optimization, we reduce optimization in a Banach space to one-dimensional optimization, and we reduce optimization over a constrained domain to unconstrained optimization. All of our reductions run as fast as online gradient descent. We use our new techniques to improve upon the previously best regret bounds for parameter-free learning, and do so for arbitrary norms.

We provide an improved analysis of normalized SGD showing that adding momentum provably removes the need for large batch sizes on non-convex objectives. Then, we consider the case of objectives with bounded second derivative and show that in this case a small tweak to the momentum formula allows normalized SGD with momentum to find an -critical point in O(1/ 3.5 ) iterations, matching the best-known rates without accruing any logarithmic factors or dependence on dimension. We also provide an adaptive method that automatically improves convergence rates when the variance in the gradients is small. Finally, we show that our method is effective when employed on popular large scale tasks such as ResNet-50 and BERT pretraining, matching the performance of the disparate methods used to get state-of-the-art results on both tasks.

No abstract