Abstract:CC-BY-NC-ND 4.0 International license not peer-reviewed) is the author/funder. It is made available under a The copyright holder for this preprint (which was. Abstract Functional relationship networks, which reveal the collaborative roles between genes, have significantly accelerated our understanding of gene functions and phenotypic relevance. However, establishing such networks for alternatively spliced isoforms remains a difficult, unaddressed problem due to the lack of systematic functional annotations at the isoform level, which renders most supervised learning methods difficult to be applied to isoforms. Here we describe a novel multiple instance learning-based probabilistic approach that integrates large-scale, heterogeneous genomic datasets, including RNA-seq, exon array, protein docking and pseudo-amino acid composition, for modeling a global functional relationship network at the isoform level in the mouse. Using this approach, we formulate a gene pair as a set of isoform pairs of potentially different properties. Through simulation and cross-validation studies, we showed the superior accuracy of our algorithm in revealing the isoform-level functional relationships. The local networks reveal functional diversity of the isoforms of the same gene, as demonstrated by both large-scale analyses and experimental and literature evidence for the disparate functions revealed for the isoforms of Ptbp1 and Anxa6 by our network. Our work can assist the understanding of the diversity of functions achieved by alternative splicing of a limited set of genes in mammalian genomes, and may shift the current gene-centered network prediction paradigm to the isoform level. Author summary Proteins carry out their functions through interacting with each other. Such interactions can be achieved through direct physical interactions, genetic interactions, or co-regulation. To summarize these interactions, researches have established functional relationship networks, in which each gene is represented as a node and the connections between the nodes represent how likely two genes work in the same biological process. Currently, these networks are established at the gene level only, while each gene, in mammalian systems, can be alternatively spliced into multiple isoforms that may have drastically different interaction partners. This information can be mined through integrating data that provide isoform-level information, such as RNA-seq and protein docking scores predicted from amino acid sequences. In this study, we developed a novel algorithm to integrate such data for predicting isoform-level functional relationship networks, which allows us to investigate the collaborative roles between genes at a high resolution.

Multi-modal Transfer Learning between Biological Foundation Models

Bridging biomolecular modalities for knowledge transfer in bio-language models

Improving ncRNA family prediction using multi-modal contrastive learning of sequence and structure

BSM: Small but Powerful Biological Sequence Model for Genes and Proteins

Deep Learning Prediction of Ribosome Profiling with Translatomer Reveals Translational Regulation and Interprets Disease Variants

Integration of protein and coding sequences enables mutual augmentation of the language model

Multimodal learning of noncoding variant effects using genome sequence and chromatin structure

ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning

Multi-LSTM: A Multiscale Long Short-Term Memory-Based Framework for Predicting Time-Series Transcriptomic Gene Expression

Transfer learning for cross-context prediction of protein expression from 5'UTR sequence

Orthrus: Towards Evolutionary and Functional RNA Foundation Models

ProtRNA: A Protein-derived RNA Language Model by Cross-Modality Transfer Learning

Species-agnostic transfer learning for cross-species transcriptomics data integration without gene orthology

A Benchmark Dataset for Multimodal Prediction of Enzymatic Function Coupling DNA Sequences and Natural Language

Modeling the functional relationship network at the isoform level through heterogeneous data integration

A long context RNA foundation model for predicting transcriptome architecture

CrossMP: Enabling Cross-Modality Translation between Single-Cell RNA-Seq and Single-Cell ATAC-Seq through Web-Based Portal

ProtTrans: Towards Cracking the Language of Life's Code Through Self-Supervised Deep Learning and High Performance Computing

CD-GPT: A Biological Foundation Model Bridging the Gap between Molecular Sequences Through Central Dogma

Geneverse: A collection of Open-source Multimodal Large Language Models for Genomic and Proteomic Research

MVKTrans: Multi-View Knowledge Transfer for Robust Multiomics Classification