Chemistry Foundation Models (CFMs) that leverage Graph Neural Networks (GNNs) operating on 3D molecular graph structures are becoming indispensable tools for computational chemists and materials scientists. These models facilitate the understanding of matter and the discovery of new molecules and materials. In contrast to GNNs operating on large homogeneous graphs, GNNs used by CFMs process a large number of geometric graphs of varying sizes, requiring different optimization strategies than those developed for large homogeneous GNNs. This paper presents optimizations for two critical phases of CFM training: data distribution and model training, targeting MACE - a state-of-the-art CFM. We address the challenge of load balancing in data distribution by formulating it as a multi-objective bin packing problem. We propose an iterative algorithm that provides a highly effective, fast, and practical solution, ensuring efficient data distribution. For the training phase, we identify symmetric tensor contraction as the key computational kernel in MACE and optimize this kernel to improve the overall performance. Our combined approach of balanced data distribution and kernel optimization significantly enhances the training process of MACE. Experimental results demonstrate a substantial speedup, reducing per-epoch execution time for training from 12 to 2 minutes on 740 GPUs with a 2.6M sample dataset.

MOTIVATION: Hi-C is the most widely used assay for investigating genome-wide 3D organization of chromatin. When working with Hi-C data, it is often useful to calculate the similarity between contact matrices in order to assess experimental reproducibility or to quantify relationships among Hi-C data from related samples. The HiCRep algorithm has been widely adopted for this task, but the existing R implementation suffers from run time limitations on high-resolution Hi-C data or on large single-cell Hi-C datasets. RESULTS: We introduce a Python implementation of HiCRep and demonstrate that it is much faster and consumes much less memory than the existing R implementation. Furthermore, we give examples of HiCRep's ability to accurately distinguish replicates from non-replicates and to reveal cell type structure among collections of Hi-C data. AVAILABILITY AND IMPLEMENTATION: HiCRep.py and its documentation are available with a GPL license at https://github.com/Noble-Lab/hicrep. The software may be installed automatically using the pip package installer. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Motivation: Single-cell Hi-C (scHi-C) data promises to enable scientists to interrogate the 3D architecture of DNA in the nucleus of the cell, studying how this structure varies stochastically or along developmental or cell-cycle axes. However, Hi-C data analysis requires methods that take into account the unique characteristics of this type of data. In this work, we explore whether methods that have been developed previously for the analysis of bulk Hi-C data can be applied to scHi-C data. We apply methods designed for analysis of bulk Hi-C data to scHi-C data in conjunction with unsupervised embedding. Results: We find that one of these methods, HiCRep, when used in conjunction with multidimensional scaling (MDS), strongly outperforms three other methods, including a technique that has been used previously for scHi-C analysis. We also provide evidence that the HiCRep/MDS method is robust to extremely low per-cell sequencing depth, that this robustness is improved even further when high-coverage and low-coverage cells are projected together, and that the method can be used to jointly embed cells from multiple published datasets.