arXiv Science⌕ Search

arXiv subjects

Jonathan Forstater

Publications and source records attributed to Jonathan Forstater.

2 recordsLinked to original sources

Conic Frames, Cosine Measure, and Gaussian Frames

Conic combinations play an important role across mathematics in areas such as optimization, non-negative matrix factorization, and sparse coding. Frames which span their underlying space using only conic combinations, known as positive spanning sets or conic frames, are of special interest in derivative free optimization. In particular, sets with large cosine measure, a measure of how spread out the vectors are, are favored. We focus on the asymptotics of the cosine measure of Gaussian frames and present tail bounds in a variety of regimes.

math.PR↗

Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs

Graph neural networks (GNNs) have emerged as powerful tools for learning protein structures by capturing spatial relationships at the residue level. However, existing GNN-based methods often face challenges in learning multiscale representations and modeling long-range dependencies efficiently. In this work, we propose an efficient multiscale graph-based learning framework tailored to proteins. Our proposed framework contains two crucial components: (1) It constructs a hierarchical graph representation comprising a collection of fine-grained subgraphs, each corresponding to a secondary structure motif (e.g., $α$-helices, $β$-strands, loops), and a single coarse-grained graph that connects these motifs based on their spatial arrangement and relative orientation. (2) It employs two GNNs for feature learning: the first operates within individual secondary motifs to capture local interactions, and the second models higher-level structural relationships across motifs. Our modular framework allows a flexible choice of GNN in each stage. Theoretically, we show that our hierarchical framework preserves the desired maximal expressiveness, ensuring no loss of critical structural information. Empirically, we demonstrate that integrating baseline GNNs into our multiscale framework remarkably improves prediction accuracy and reduces computational cost across various benchmarks.

cs.LG↗