Intrinsic-Extrinsic Convolution and Pooling for Learning on 3D Protein\n Structures

2020/07/13 by Pedro Hermosilla, Hermosilla, Pedro, Marco Schäfer +16 · 8 citations
Biochemistry, Genetics and Molecular Biology · Computer Science · Mathematics · #Algorithm #Artificial intelligence #Artificial neural network #Bioinformatics and Genomic Networks #Biomolecules (q-bio.BM) #Computer science #Convolution (computer science) #Euclidean distance #FOS: Biological sciences #FOS: Computer and information sciences #Feature learning #Genomics and Chromatin Dynamics #Geodesic #Graph #Invariant (physics) #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning in Bioinformatics #Machine learning #Mathematics #Pooling #Protein Structure and Dynamics #Representation (politics) #Scale (ratio) #Set (abstract data type) #Theoretical computer science #cs.LG #q-bio.BM #stat.ML

paper · pdf · doi:10.48550/arxiv.2007.06252

published in arXiv (Cornell University), 1-16 (Cornell University) · International Conference on Learning Representations (ICLR) 2021

openalex publication_date 2020/07/13 · arxiv created 2021/04/19 · arxiv updated 2021/04/20 · openalex created_date 2021/04/26 · openalex updated_date 2026/07/28

Abstract

Proteins perform a large variety of functions in living organisms, thus\nplaying a key role in biology. As of now, available learning algorithms to\nprocess protein data do not consider several particularities of such data\nand/or do not scale well for large protein conformations. To fill this gap, we\npropose two new learning operations enabling deep 3D analysis of large-scale\nprotein data. First, we introduce a novel convolution operator which considers\nboth, the intrinsic (invariant under protein folding) as well as extrinsic\n(invariant under bonding) structure, by using n-D convolutions defined on\nboth the Euclidean distance, as well as multiple geodesic distances between\natoms in a multi-graph. Second, we enable a multi-scale protein analysis by\nintroducing hierarchical pooling operators, exploiting the fact that proteins\nare a recombination of a finite set of amino acids, which can be pooled using\nshared pooling matrices. Lastly, we evaluate the accuracy of our algorithms on\nseveral large-scale data sets for common protein analysis tasks, where we\noutperform state-of-the-art methods.\n

Citations

• Dictionary of protein secondary structure: Pattern recognition of hydrogen‐bonded and geometrical features
• Basic local alignment search tool
• A tutorial on spectral clustering
• ImageNet classification with deep convolutional neural networks
• Semi-Supervised Classification with Graph Convolutional Networks
• PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space
• Deformable Filter Convolution for Point Cloud Reasoning
• Hierarchical Graph Pooling with Structure Learning
• Improved protein structure prediction using potentials from deep learning
• A Note on Over-Smoothing for Graph Neural Networks

Related

• Deep Learning of Protein Structural Classes: Any Evidence for an 'Urfold'?
• Protein sequence-to-structure learning: Is this the end(-to-end\n revolution)?
• Learning inverse folding from millions of predicted structures
• A deep learning approach to the structural analysis of proteins
• Ophiuchus: Scalable Modeling of Protein Structures through Hierarchical Coarse-graining SO(3)-Equivariant Autoencoders