Posters - Schedules

Posters Home

Predicting and classifying protein functions is a significant step in advancing our knowledge of biology and leveraging the full potential of genomics and metagenomics sequencing projects.
Significant advances have been made by the application of deep learning models with various architectures (such as CNNs and GCNs) for protein representation learning. However these models, albeit with some success, face limitations in capturing the complex structure of protein domain hierarchies which have a crucial role in determining functions of new, first-seen proteins.
This work proposes a novel Hyperbolic Relation Graph Neural Network (HRGNN) for protein representation learning. HRGNN employs hyperbolic geometry to represent the intrinsic curvature of the protein space, enabling it to effectively capture the complex hierarchical structures of proteins.
The proposed model leverages the inherent graph structure of proteins as it can capture the connectivity patterns of interactions among amino acids to learn informative representations of proteins.
We demonstrate the effectiveness of our approach on three benchmark protein datasets over the EC and GO criteria and show that HRGNN achieves state-of-the-art performance on protein function prediction tasks.
Overall, the promising results of HRGNN provide more accurate predictions of protein functions, ultimately leading to fundamental advances in biology.

Many photosynthesis-related genes in the model alga Chlamydomonas reinhardtii remain largely unknown, and elucidation of these gene functions can further efforts towards optimizing algae for biofuel production, improving the photosynthetic efficiency of agricultural crops, and discovering new genes in a variety of biological processes. Previous work has produced a manually curated set of 253 high-confidence candidate genes from a collection of 328 plasmid-insertion (acetate-requiring, photosynthetically-deficient) mutants. This candidate list contains 55 genes with known photosynthesis-related function while the remaining are conserved in green algae and plants, or have limited to no functional annotations. We approach functional discovery using a tiered approach. Phylogenomic comparison to well-characterized genomes such as Arabidopsis thaliana yielded 88 photosynthesis-related orthologs from which we can infer function. We developed a structure comparison pipeline using Alphafold2 and FATCAT/FoldSeek to identify structurally similar proteins from various databases that may represent distant homologs. We leveraged large multi-omics datasets (transcriptomics, proteomics) using unsupervised machine learning techniques, UMAP, to cluster expression profiles and infer additional protein function. This multi-faceted approach using phylogenomic, multi-omics, machine learning and comparative structural analyses has yielded promising functional predictions of previously unknown photosynthesis-related genes.

As one of the state-of-the-art automated function prediction (AFP) methods, NetGO 2.0 integrates multi-source information to improve the performance. However, it mainly utilizes the proteins with experimentally supported functional annotations without leveraging valuable information from a vast number of unannotated proteins. Recently, protein language models have been proposed to learn informative representations [e.g., Evolutionary Scale Modelling (ESM)-1b embedding] from protein sequences based on self-supervision. We represented each protein by ESM-1b and used logistic regression (LR) to train a new model, LR-ESM, for AFP. The experimental results showed that LR-ESM achieved comparable performance with the best-performing component of NetGO 2.0. Therefore, by incorporating LR-ESM into NetGO 2.0, we developed NetGO 3.0 to improve the performance of AFP extensively. NetGO 3.0 is freely accessible at https://dmiip.sjtu.edu.cn/ng3.0.