CD2K: An Accessible Framework for Leveraging Public Omics Data in Biomedical Research Training

1. Executive Summary/Introduction

The Collective Omics Data to Knowledge (CD2K) initiative represents an original approach to biomedical research training, leveraging the vast potential of publicly available omics datasets [1]. This program primarily aims to address a critical need in the field: supporting the development of research capacity, particularly in low- and middle-income countries (LMICs). As a byproduct of these training activities, the program also contributes to generating actionable knowledge from the wealth of existing data [1,2].

At its core, CD2K is built on the principle that the endpoint of the training curriculum should be publication, aligning with the primary currency for advancing research careers [1]. This focus on generating "actionable" or translational knowledge ensures that participants not only develop crucial skills but also produce tangible outputs with real-world impact, serving as a source of motivation for trainees [2].

The CD2K initiative encompasses three core training curricula [1]:

• COD1: Focuses on reductionist interpretation of collective omics data
• COD2: Involves the creation of curated dataset collections
• COD3: Trains participants in the re-analysis of omics data on a global scale

These modules equip researchers with the skills to navigate public data repositories, identify knowledge gaps, apply novel analytical approaches, and ultimately generate publishable findings [1,2]. Importantly, some of these programs, particularly the COD1 workflows, are accessible to participants who do not have extensive data science training or experience. This accessibility demonstrates that publication based on collections of public omics datasets and resulting career advancement are attainable for mainstream bench scientists [1].

As will be detailed further in this article, the success of the CD2K initiative is evidenced by numerous proof-of-principle publications from researchers who have completed one or more of the COD training modules. These successes demonstrate the feasibility of using collective omics data as a springboard for research output and career advancement, even in resource-limited settings [1,2].

Recent advancements in the program include the integration of Large Language Models (LLMs) into the COD1 workflow, further enhancing the efficiency and depth of analysis possible within this framework [2]. This integration represents a significant step forward in leveraging cutting-edge AI technologies for biomedical research and training.

As the CD2K initiative continues to evolve, it holds significant potential to democratize access to omics data and empower researchers, particularly in LMICs, to make meaningful contributions to biomedical knowledge [1,2]. By providing a structured framework for data reuse and emphasizing actionable outcomes, CD2K aligns with broader efforts to promote equity and accelerate discovery in global health research.

2. Historical Development and Overview of COD Training Modules

2.1. COD1: Reductionist Interpretation of Collective Omics Data

The COD1 module focuses on the reductionist interpretation of collective omics data, emphasizing gene-centric investigations. This approach has proven to be accessible to participants without extensive data science training, making it particularly valuable for mainstream bench scientists. The methodology involves selecting genes presenting differences in transcript abundance between study groups, assessing knowledge gaps by examining the overlap between gene-associated literature and study-relevant concepts, and validating observations using independent datasets when possible [1].

Since its inception, the COD1 training program has involved 49 unique individuals from 12 different institutions across 8 countries, demonstrating its global reach and collaborative nature. These workshops have resulted in 10 peer-reviewed publications, each focusing on a different gene and its potential role in various immunological and disease contexts (Table 1).

The genes investigated through this program span a wide range of biological processes and disease states. For instance, ALAS2's role in erythroid cells [2], ACVR1B, NUDT16, ERLIN1, ANXA3, and ACSL1 in sepsis [3,4,5,6,7], CST7 in acute inflammation [8], BANK1 in B cell tolerance [9], TKT in inflammation [10], and ADAM9 in tissue damage [11] have all been explored.

While primarily serving as a training tool, these investigations have yielded potentially significant findings. For example, the study on ANXA3 suggested its potential role in neutrophil function during sepsis, which could have implications for sepsis management [6]. The investigation of ACSL1 pointed to its possible involvement in inflammasome activation in neutrophils during sepsis, with potential prognostic value [7].

It is important to note that while these findings suggest potential translational significance, they should be considered preliminary and require further validation. Nonetheless, they demonstrate the capacity of this training approach to generate hypotheses and identify potential biomarkers or therapeutic targets worthy of further investigation. Taken together this iteration of the COD1 module has successfully equipped participants with skills in navigating public data repositories, identifying knowledge gaps in biomedical literature, developing advanced PubMed query skills, and formulating and testing hypotheses.

The COD1 module has recently undergone significant evolution, with a sharpened focus on the development of targeted assays as an actionable endpoint (Figure 1). This shift aligns with the CD2K/COD program's emphasis on producing tangible, translatable outcomes alongside publications. The integration of Large Language Models (LLMs) into the workflow has markedly enhanced both the efficiency and depth of analysis possible within this framework [2]. This technological advancement has enabled the implementation of sophisticated workflows for gene prioritization and selection, followed by in-depth characterization of selected candidates, exemplified by studies such as the CEACAM6 analysis [52]. The new orientation of COD1 reflects a more holistic approach to biomarker discovery and validation, bridging the gap between high-throughput data analysis and practical clinical applications. To date, three workshops implementing this updated curriculum have been conducted The Jackson Laboratory in the US, Mahidol University in Thailand, and Khon Kaen University in Thailand. These workshops have provided participants with hands-on experience in leveraging cutting-edge AI technologies for biomedical research. This evolution of the COD1 module represents a significant step forward in translating complex genomic data into potential diagnostic and therapeutic tools, further enhancing the impact and relevance of the CD2K initiative in the field of precision medicine.

COD2: Creation of Curated Collective Omics Dataset Collections

The COD2 module focuses on the creation of curated dataset collections, emphasizing the organization and accessibility of large-scale omics data (Figure 2). This approach enables researchers to leverage existing public data repositories effectively, promoting data reuse and fostering new discoveries [12].

The methodology involves identifying relevant datasets from public repositories such as the NCBI Gene Expression Omnibus (GEO), curating and annotating these datasets, and making them accessible through user-friendly web applications [13]. These applications, such as the Gene Expression Browser (GXB), allow for interactive query and visualization of integrated large-scale data [14].

Since its inception, the COD2 training program has involved 84 unique individuals from over 20 different institutions across multiple countries, demonstrating its global reach and collaborative nature. These efforts have resulted in 12 publicly available dataset collections, each focusing on a different biological theme or disease context (Table 2).

The dataset collections span a wide range of biological processes and disease states, including systemic lupus erythematosus [15], pregnancy [16], primary immunodeficiencies [17], IgE-mediated atopic diseases [18], viral respiratory tract infections [19], breast cancer [20], in vitro fertilization [21], hematopoietic cells in early life [22], placental development [23], monocyte immunobiology [24], HIV infection [25], and sepsis [26].

While primarily serving as a training tool and resource for the research community, these curated dataset collections have yielded significant benefits. They have facilitated the re-analysis of public data, leading to new insights and hypotheses. For example, the breast cancer dataset collection has been used to investigate immunologic classifications of tumors [20], while the sepsis collection has enabled the identification of specific human blood gene signatures in response to infection [26].

The COD2 module has successfully equipped participants with skills in data curation, annotation, and the development of user-friendly interfaces for data exploration. Recent advancements include the integration of module-level data browsing capabilities [15,16], which ties in with COD3 activities and enhances the depth of analysis possible within this framework.

The impact of the COD2 module extends beyond the immediate training outcomes. By making curated dataset collections publicly available, it has contributed to the broader scientific community's efforts in promoting data reuse and reproducibility in biomedical research. These resources serve as valuable tools for hypothesis generation, validation of findings across multiple studies, and the identification of robust biomarkers or therapeutic targets.

COD3: Re-Analysis of Collective Omics Data on a Global Scale

The COD3 module focuses on the analysis and interpretation of large-scale profiling data, particularly aimed at individuals with data science training or those interested in pursuing a career in this field. However, it is worth noting that data science expertise is not always a prerequisite for participation in COD3 activities.

Central to the COD3 approach is its reliance on well-characterized fixed module repertoires, such as the BloodGen3 module repertoire [27], a fixed set of 382 transcriptional modules encompassing 14,168 transcripts. This repertoire, along with the accompanying R package BloodGen3Module [28], and web applications for module-level data browsing [29,30], form the core resources leveraged in COD3 training activities (Figure 3).

The COD3 training activities have explored multiple avenues, as evidenced by several published use cases (Table 3):

• Meta-analyses at the module level: These studies encompass multiple independent datasets, focusing either on investigating disease pathogenesis or biomarker discovery. For example, a study on psoriasis [31] utilized module-level meta-analysis to investigate the neutrophil-driven inflammatory signature characterizing the blood transcriptome fingerprint of the disease. Similarly, research on respiratory syncytial virus (RSV) infection [32] employed module-level meta-analysis for biomarker discovery, identifying erythroid cell-positive blood transcriptome phenotypes associated with severe RSV infection.
• Data and knowledge-driven biomarker discovery: Several studies have employed this approach, including work on pregnancy [33], COVID-19 [34], and most recently, the use of Large Language Models (LLMs) for developing a generic immune profiling assay [35]. This last study is particularly noteworthy as it demonstrates the potential of integrating artificial intelligence into the biomarker discovery process.
• Data interpretation workshops: These workshops, supported by BloodGen3 applications, involve participants with expertise in medicine and immunology who may not have extensive data science backgrounds. These activities emphasize the accessibility of COD3 workflows, even when focusing on the analysis of large-scale datasets. A manuscript describing this approach is currently in preparation.

The LLM-enabled workflow, detailed in the recent publication by Toufiq et al. [35], demonstrates how generative AI technology can be harnessed for knowledge-driven candidate gene prioritization and selection. By leveraging both data-driven and knowledge-based approaches, this integrated COD3 workflow represents a means to significantly improve candidate gene prioritization, enabling, for instance, the development of targeted profiling assays.

The resources developed as part of COD3, particularly the BloodGen3 module repertoire [27] and the BloodGen3Module R package [28], have facilitated the creation of interactive web applications for exploring blood transcriptomic diversity in various conditions. Notable examples include applications for systemic lupus erythematosus [29] and pregnancy [30], which allow researchers to interactively explore module-level blood transcriptome data.

Overall, the COD3 module has successfully equipped participants with advanced skills in large-scale data analysis, interpretation, and visualization. The integration of novel technologies like LLMs and the development of user-friendly web applications have further enhanced the accessibility and impact of these training activities.