The 9th Big Data Forum for Life and Health Sciences & The 21st Asian Bioinformatics Consortium Symposium

In this keynote, I will explore how we are witnessing a revolution in artificial intelligence approaches that is transforming molecular biology and computational biology. I'll discuss how recent advancements in AI, particularly in structural prediction models, are reshaping our understanding and classification of protein families. I will begin by examining the impact of high-accuracy structural models, such as AlphaFold, on our ability to completely and accurately classify protein domains. I'll show how these models have enabled us to refine, define, and classify protein domain families with unprecedented precision. Next, I'll delve into the capabilities of deep learning models like Google Research's ProtENN, which have significantly expanded our ability to identify distant homologues for known protein families. I will also discuss our ongoing collaboration with Google Research, where we're exploring video transformer technologies for annotating InterPro. I believe these advancements collectively represent the most significant progress in protein classification in three decades. I will then cover the emergence of Large Language Models (LLMs) such as ChatGPT and their potential to develop high-throughput tools for annotating proteins, non-coding RNAs, and families. I'll discuss the challenges we face in harnessing these models to write accurate and verifiable annotations at scale, including our strategies to mitigate hallucination issues. Finally, I will explore how these AI-driven approaches are bridging the gap between protein sequence and structure classification, potentially leading to a unified classification system in the near future. I'll conclude with a discussion on the broader implications of these advancements for the field of molecular biology and the exciting possibilities they present for future research and applications.

蛋白质组研究的快速发展产生了海量的实验数据，如何有效地管理和共享蛋白质组学数据集是目前面临的重大挑战之一。海量蛋白质组学数据交换共享的难点在于缺乏标准信息规范，在满足数据集高效快捷共享的同时提供足够的实验和生物学相关信息。人类蛋白质组组织（Human Proteome Organization，HUPO）所属的蛋白质组标准计划（Proteomics Standards Initiative，PSI）是国际上蛋白质组领域最权威的数据标准和信息指南制定和发布的组织。从2002年至今，PSI组织从蛋白质组最低信息准则、数据格式、控制词汇表、数据资源和分析软件五个方面组织研究并交付成果。但目前国内外还没有针对基于质谱的蛋白质组学实验数据集在不同实验室、数据平台之间的交换制定相关标准。此外，蛋白质组学公共数据集的重用正变得越来越流行，而数据集重用在很大程度上仅限于与肽段和蛋白质的鉴定相关的基准研究和应用。公共定量实验的广泛重用受到缺乏样本和实验元数据的限制，导致不能将每个数据集中的数据文件与样本明确地关联起来。本研究制定了基于质谱的蛋白质组学数据集交换标准，对数据集交换过程中涉及的数据集元数据、实验元数据和实验数据文件规范信息进行了详细的描述，将是首个蛋白质组学数据集交换的信息标准，可以作为本领域未来发展的基础，实现蛋白质组实验数据集的高质量发布、交换和共享，促进蛋白质组学实验数据集的重用。

We present the South and East Asian reference Database (SEAD) panel (https://imputationserver.westlake.edu.cn/), which comprises whole genome sequencing data from 11,067 individuals across 17 countries in Asia. The SEAD panel, which excludes singleton variants, consists of 22,134 haplotypes and 88,294,957 variants. The SEAD panel demonstrated higher accuracy compared to 1kGP, TOPMed and ChinaMAP in South Asian population. And as the proportion of South Asian ancestry increased, the proportion of low-frequency and rare well-imputed variants imputed using the SEAD panel progressively increased, whereas those imputed with TOPMed panel significantly decreased. Additionally, when imputing the East Asian population, the SEAD showed comparable concordance in imputation with ChinaMAP panel, while the TOPMed panel was inferior. Finally, we applied the augmented SEAD panel to conduct a discovery and replication genome-wide association study (GWAS) for hip and femoral neck (FN) bone mineral density (BMD) traits within the 5,369 Westlake BioBank for Chinese (WBBC) genotyped samples. The single-variant test suggests that rare variants near SNTG1 gene are associated with hip BMD (rs60103302, MAF=0.0092, P=1.67×10−7). The variant-set analysis also suggests the association of this gene (Pslide_window=9.08×10−9, Pgene_centric=5.27×10−8). The gene and variants achieved a suggestive level for FN BMD. This gene was not reported previously and can only be detected by using Asian reference panel. The preliminary experiment in-vitro demonstrated that the identified rare variant could upregulate the SNTG1 expression, which in turn inhibits the proliferation and differentiation of preosteoblast.

The China Kadoorie Biobank (CKB) is one of the world's largest prospective cohort studies. It aims to assess established and emerging risk factors for chronic diseases, understand the interplay between genes, lifestyle, and environmental factors, and identify new biological mechanisms linking risk factors to major chronic diseases for the development of new treatments. Based on data from over 0.5 million participants in the China Kadoorie Biobank (CKB) across 10 regions, our goals are to: 1. Examine the relationship between marital status and health risks among Chinese adults, and assess how marital status is linked to the risk of specific diseases and mortality in the population. 2. Investigate the association between parenthood and health risks among Chinese adults, and evaluate how parenthood relates to the risk of specific diseases and mortality in the population. 3. Study the impact of marital status on mortality rates in the Chinese population from 2021 to 2040.

Plant genomes are enriched with complex and evolutionarily diverse non-coding regulatory sequences that follow unknown cis-regulatory codes. In the last decade, high-throughput sequencing technologies have amassed vast regulomic data across various model plants. Here, we developed ChIP-Hub, a comprehensive database that integrates over 16,000 publicly available datasets from nearly 600 publications on model plant species. Using this extensive dataset, we systematically annotated tissue- and cell-specific regulatory elements, including promoters and enhancers, and constructed corresponding gene regulatory networks. Additionally, comparative genomics approaches were employed to explore the functional conservation of tissue-specific regulatory elements and chromatin states across species. We also developed the SeiPlant model, leveraging dilated convolutions with residual connections to capture both linear and nonlinear sequence features, along with spatial basis function layers to enhance model scalability. This AI framework effectively integrates large-scale plant regulomic data, enabling cross-species prediction of regulatory elements based solely on genomic sequences. This model enables to understand the molecular mechanisms of gene regulation, how mutations in cis-regulatory sequences alter their function, and consequently, how they impact plant traits. This understanding is ultimately crucial for designing synthetic regulatory sequences to enhance crop performance.

As the mainstream high-throughput method to identify protein sequences, tandem mass spectrometry (MS) plays an important role in proteomics research by generating mass spectra (MS1, MS2) and then analyzing the corresponding peptide sequences. However, current popular methods based on database search are limited by the reference protein sequence database and cannot identify the protein sequences outside of the database, making them unsuitable for specific research, such as neoantigen discovery, antibody design, and vaccine development. De novo peptide sequencing is a promising method to address the above shortcomings. In recent years, the rapid development of AI techniques such as large language model has revolutionized many fields, including the proteomics research. Here, we introduce π-PrimeNovo, a non-autoregressive Transformer-based deep learning model designed to perform accurate and efficient de novo peptide sequencing. With the proposed novel architecture, π-PrimeNovo achieves significantly higher accuracy and up to 69x faster sequencing compared to the state-of-the-art methods. This remarkable speed makes it highly suitable for computation-extensive peptide sequencing tasks such as metaproteomic research, where π-PrimeNovo efficiently identifies the microbial species-specific peptides. Moreover, π-PrimeNovo has been demonstrated to have a powerful capability in accurately mining phosphopeptides in a non-enriched phosphoproteomic dataset, showing an alternative solution to detect low-abundance post-translational modifications (PTMs). We suggest that this work not only advances the development of peptide sequencing techniques but also introduces a transformative computational model with wide-range implications for biological research. Key words: computational proteomics, artificial intelligence, de novo peptide sequencing

Every sovereign state intends to impose its own regulations on the access to and sharing of personal genomes. Ultimately, the authority to control shall rest on our comprehension of privacy and fundamental human rights. Since the present definition of privacy was developed over a century ago, we must rethink it in the context of the genomic era. The issue is related to the current dispute about the access and benefit sharing of genetic information in the Convention on Biological Diversity. I will outline key areas of discussion on this subject and propose restorative and reparative measures for possible harmful accidents in the future.

The large variety of data available from biological research is a rich resource that can be used for innovative endeavors. However, we are facing considerable challenges in big data deposition, integration, and translation due to the complexity of biological data and its production at unprecedented exponential rates. To address these problems, in 2020, the Korean government officially announced a national strategy to collect and manage the biological data produced through national R&D fund allocations and provide the collected data to researchers. To this end, the Korea Bioinformation Center (KOBIC) developed a new biological data repository, the Korea BioData Station (K-BDS), for sharing data from individual researchers and research programs to create a data-driven biological study environment. The K-BDS is dedicated to providing free open access to a suite of featured data resources in support of worldwide activities.

High-throughput sequencing technology has been widely employed in medical research and practice. However, how to securely share and effectively utilize these high-throughput gene sequencing data remains a problem that demands to be addressed. Based on our own practice and in collaboration with several domestic units, we have put forward a set of standard solutions, ranging from the definition of terms, basic principles, shared content, scenarios to process management. We cooperated with regulatory authorities, medical institutions, enterprises, and so on, to verify the applicability of this standard solution. The scheme we provide offers an important reference for similar biological data sharing.

China is populated by 1.4 billion people of multiple ethnic groups with high cultural and language diversities. The Chinese Pangenome Consortium (CPC) aims to produce high-quality genomic sequences from people representative of the majority ethnic group (Han) and the 55 other defined ethnic groups of China, as well as of multiple unrecognized ethnic groups not yet characterized genetically. At the stage of Phase II (CPC2), apart from the collection of 116 high-quality genome assemblies from 58 core individuals, CPC aims to produce high-quality, phased, chromosome-level haplotype sequences of 500 individuals representing diverse populations in China. We expect the CPC2, as part of global efforts in human genomics, to make a considerable contribution towards the building of high-quality pangenome references and their application to various basic and clinical research questions. In particular, The CPC data have the potential for tracing missing links in our understanding of human genetic evolution and are valuable for explaining the heritability of complex diseases that is not explained by known variants.

Artificial intelligence methods to assist scientific research (AI for Science) have become one of the current important research trends, where scientific data plays an indispensable role in the process. For instance, AlphaFold has predicted protein folding with extremely high accuracy, greatly enhancing the research efficiency in the field of structural biology, and the high efficiency of its method is inseparable from high-quality scientific data accumulated over decades, such as PDB. This report explores the innovation driven by scientific data in AI4Science, based on the speaker's technical foundation in big data and knowledge graph technology, combined with the capabilities demonstrated by large language models, targeting scientific research needs represented by the cross-field of microbial nanomaterials, introducing methods for constructing high-quality scientific datasets and attempts of AI4Science based on high-quality scientific data.

After fertilization, most of genes in the human genome are transcriptionally silent until the 8-cell stage, at which point the zygotic genome begins to be extensively transcribed. This process, known as zygotic genome activation (ZGA), is crucial for the further development of human early embryos. However, how human zygotic genome is activated is still elusive. To explore the underlying mechanisms of human ZGA, we investigated the epigenetic landscapes in human early embryos, including chromatin accessibility and histone modifications. We uncovered that OCT4 is essential for human ZGA, which is the first transcription factor identified to regulate this process. Unexpectedly, we observed widespread non-canonical broad H3K27ac domains in human embryos prior to ZGA. The transition from these non-canonical broad H3K27ac domains to typical narrow H3K27ac peaks is associated with ZGA. Furthermore, histone deacetylases are critical for this transition and ZGA. We also compared the contribution of parental genomes to ZGA using human parthenogenetic and androgenetic haploid embryos. Our results indicate that human ZGA is initiated from the paternal genome. In contrast, ZGA occurs simultaneously from both paternal and maternal genomes in mouse. Mechanismly, a primate-specific transcription factor, ZNF675, with paternally biased chromatin accessibility and gene expression, is essential for human ZGA. Taken together, our studies provide insight into the regulatory mechanisms of human ZGA.

The rapid spread of carbapenemase-producing Klebsiella pneumoniae (cpKP) poses serious threats to public health; however, the underlying genetic basis for its dissemination is still unknown. We conducted a comprehensive genomic epidemiology analysis on 420 cpKP isolates collected from 70 hospitals in 24 provinces/autonomous regions/municipalities of China during 2009–2017 by short-/long-read sequencing. The results showed that most cpKP isolates were categorized into clonal group 258 (CG258), in which ST11 was the dominant clone. Phylogenetic analysis revealed three major clades including the top one of Clade 3 for CG258 cpKP isolates. Additionally, carbapenemase gene analysis indicated that blaKPC was dominant in the cpKP isolates, and most blaKPC genes were located in five major incompatibility (Inc) groups of blaKPC-harboring plasmids. Importantly, three advantageous combinations of host–blaKPC-carrying plasmid (Clade 3.1+3.2–IncFIIpHN7A8, Clade 3.1+3.2–IncFIIpHN7A8:IncR, and Clade 3.3–IncFIIpHN7A8:IncpA1763-KPC) were identified to confer cpKP isolates the advantages in both genotypes (strong correlation/coevolution) and phenotypes (resistance/growth/competition) to facilitate the nationwide spread of ST11/CG258 cpKP. Intriguingly, Bayesian skyline analysis illustrated that the three advantageous combinations might be directly associated with the strong population expansion during 2007–2008 and subsequent maintenance of the population of ST11/CG258 cpKP after 2008. We then examined drug resistance profiles of these cpKP isolates and proposed combination treatment regimens for CG258/non-CG258 cpKP infections. Thus, the findings of our systematical analysis shed light on the molecular epidemiology and genetic basis for the dissemination of ST11/CG258 cpKP in China, and much emphasis should be given to the close monitoring of advantageous cpKP–plasmid combinations.

Dissecting the immune responses underlying human diseases is difficult based on traditional experimental approaches. The advent of single-cell spatial sequencing techniques enables digitalization of clinical samples, opening the door of in silico investigation of human immune mechanisms directly based on clinical samples. However, due to phyical and chemical limitations, current single-cell spatial sequencing techniques are imperfect. We develop a series of AI-based algorithms to synthesize single-cell, spatially-resolved, and transcriptome-wide digitalized clinical samples, which enables the generation of immunological hypotheses of human diseases in silico, exemplified by cancers and infectious diseases.

The utilization of single-cell combinatorial indexing sequencing via droplet microfluidics presents an attractive approach for balancing cost, scalability, robustness, and accessibility. Nevertheless, existing methodologies necessitate tailored protocols for individual modalities, which may constrain their potential for automation and hinder clinical applications. Introducing UDA-seq, a universal workflow that integrates a straightforward post-indexing step to enhance throughput and systematically adapt existing droplet-based single-cell multimodal methods. UDA-seq was subjected to benchmarking across various tissue and cell types, successfully enabling several prevalent multimodal tasks, including single-cell co-assay of RNA and VDJ, RNA and chromatin, and RNA and CRISPR perturbation. Notably, UDA-seq facilitated the efficient generation of over 100,000 high-quality single-cell data from three dozen frozen clinical biopsy specimens in a singular-channel experiment of droplet microfluidics. Subsequent analysis substantiated the potency of our approach in identifying rare cell subpopulations associated with clinical phenotypes and probing the vulnerability of cancer cells.

Emerging evidence showed that paternally acquired phenotypes (e.g., metabolic disorders) from environmental stressors can be memorized by sperm beyond DNA sequence, encoded in the form of sperm RNAs and RNA modifications as a 'sperm RNA code'. However, RNA modifications present significant challenges in constructing complementary DNA libraries, thereby impeding the detection of highly modified small non-coding RNAs. To address this limitation, we developed PANDORA-seq, a novel sequencing technology designed to bypass the sequence interferences caused by RNA modifications. Using PANDORA-seq, we characterized the profiles of tRNA-derived small RNAs (tsRNAs) and rRNA-derived small RNAs (rsRNAs), identifying them as potential key mediators in the transmission of paternal acquired information to offspring.

Technological advances in spatial transcriptomics are critical for better understanding the structures and functions of tissues in biological research. The combination of intelligent or statistical algorithms and spatial transcriptomics has emerged to pave the way for deciphering tissue architecture. We have made great efforts to advance intelligent spatial transcriptomics and developed a group of STA- tools. For example, we created a graph attention auto-encoder tool STAGATE to identify spatial domains by learning low-dimensional latent embeddings via integrating spatial information and gene expression profiles. Second, we introduced STAligner for integrating and aligning ST datasets across different conditions, technologies, and developmental stages to enable spatially-aware data integration, simultaneous spatial domain identification, and downstream comparative analysis. Third, we designed STAMarker for identifying spatially domain-specific variable genes with saliency maps in deep learning. Fourth, we developed a spatial location-supervised auto-encoder generator STAGE for generating high-density spatial transcriptomics. Fifth, we developed STASCAN for deciphering fine-resolution cell-distribution maps in spatial transcriptomics.

Gemcitabine is commonly used for pancreatic ductal adenocarcinoma (PDAC), one of the most lethal cancer types. However, the drug resistance is a critical challenge for improving the PDAC chemotherapy. Here, we applied single-cell RNA sequencing (scRNA-seq) on PDAC patient-derived xenograft (PDX) models to study the complex cellular responses related to the gemcitabine resistances. To reconstruct dynamic tumor cell responses from these static scRNA-seq snapshots, we proposed scConGraph, a scalable bi-layer graph model that can efficiently integrate cross-time context information. Based on scConGraph, we observed that stemness and endoplasmic reticulum stress contribute to intrinsic resistance. As for acquired resistance, cancer cells may resist or evade gemcitabine treatment by activating the cell cycle, entering quiescence, or inducing epithelial-mesenchymal transition. Notably, GDF15 exhibited recurrent and significant upregulations among acquired-resistance cell subpopulations. Experimental validation confirmed that inhibiting GDF15 sensitizes tumor cells to gemcitabine, suggesting a potential target for gemcitabine-induced chemoresistance.

Characterizing the compositional and phenotypic characteristics of tumor-infiltrating B cells (TIBs) is important for advancing our understanding of their role in cancer development. Here, we establish a comprehensive resource of human B cells by integrating single-cell RNA sequencing data of B cells from 649 patients across 19 major cancer types. We demonstrate substantial heterogeneity in their total abundance and subtype composition and observe immunoglobulin G (IgG)-skewness of antibody-secreting cell isotypes. Moreover, we identify stress-response memory B cells and tumor-associated atypical B cells (TAABs), two tumor-enriched subpopulations with prognostic potential, shared in a pan-cancer manner. In particular, TAABs, characterized by a high clonal expansion level and proliferative capacity as well as by close interactions with activated CD4 T cells in tumors, are predictive of immunotherapy response. Our integrative resource depicts distinct clinically relevant TIB subsets, laying a foundation for further exploration of functional commonality and diversity of B cells in cancer.

Integrating single-cell datasets produced by multiple omics technologies is essential for defining cellular heterogeneity. Mosaic integration, in which different datasets share only some of the measured modalities, poses major challenges, particularly regarding modality alignment and batch effect removal. Here, we present a deep probabilistic framework for the mosaic integration and knowledge transfer (MIDAS) of single-cell multimodal data. MIDAS simultaneously achieves dimensionality reduction, imputation and batch correction of mosaic data by using self-supervised modality alignment and information-theoretic latent disentanglement. We demonstrate its superiority to 19 other methods and reliability by evaluating its performance in trimodal and mosaic integration tasks. We also constructed a single-cell trimodal atlas of human peripheral blood mononuclear cells and tailored transfer learning and reciprocal reference mapping schemes to enable flexible and accurate knowledge transfer from the atlas to new data. Applications in mosaic integration, pseudotime analysis and cross-tissue knowledge transfer on bone marrow mosaic datasets demonstrate the versatility and superiority of MIDAS. With the rapid increasing of single cell multimodal data, single cell multimodal atlas requires frequent update to provide comprehensive knowledge. The update of the atlases conventionally requires reintegration of all data and is computationally intensive, which makes the frequent update infeasible. To address these challenges, we present Multimodal Integration with Continual Learning (MIRACLE), a novel online learning framework for adaptive and efficient integration of single-cell multimodal data. MIRACLE uses dynamic architecture and data rehearsal strategies to support continual learning, integrating diverse data while minimizing information loss over time. Our evaluations show that MIRACLE achieves accurate online integration with reduced computational requirements, effectively updates and expands atlases with new cross-tissue and cross-modal data, and precisely identifies novel cell types and transfers labels across datasets.

Compared to single omics, multi-omics provides a more comprehensive perspective and multi-dimensional analysis. By integrating various data such as genomics, transcriptomics, proteomics, metabolomics, and imaging omics, we can overcome the limitations of single omics, deeply reveal complex interactions and hierarchical regulatory mechanisms in biological systems, and thus gain a more comprehensive understanding of biological processes. This report will focus on key technologies and applications of multimodal spatial omics integration analysis. Multi-omics data often comes with challenges such as strong heterogeneity, high dimensionality, batch effects, and incomplete data, bringing new computational challenges. To address these, we have developed multimodal generative models to learn unified representations and cross-modal generation for heterogeneous multi-omics data. We use unsupervised learning and correlation analysis methods to discover intrinsic connections between omics, identify key biomarkers and signaling pathway networks associated with clinical phenotypes and treatment responses, providing in-depth and accurate panoramic analysis for personalized cancer diagnosis and treatment.

In this talk, I will present a sequence-based computational model developed by our research group that utilizes combinatorial transcription factor (TF) genomic occupancy to predict tissue-specific enhancers. Trained on diverse datasets, including ENCODE and the VISTA enhancer browser data, the model predicted 25,000 forebrain-specific cis-regulatory modules (CRMs) in the human genome. Validation through biochemical features, disease-associated SNPs, and in vivo zebrafish analysis confirmed its effectiveness. This model enhances the prediction of enhancers that lack well-characterized chromatin features, providing a valuable complement to experimental approaches in tissue-specific enhancer discovery.

Artificial intelligence has been widely used in infectious disease surveillance and early warning. Artificial intelligence technology can rapidly collect, analyze and process large-scale data from multiple sources, identify complex patterns in the data based on efficient algorithms, accurately simulate and predict the development trend of epidemics, and assist in making optimal decisions. This talk briefly introduces the application of AI technology in infectious disease surveillance and early warning system, describes its advantages, and discusses the challenges faced in the future, aiming to provide a reliable reference for more effective monitoring and control of infectious disease epidemics.