G2PDeep platform aims to improve multi-omics predictions for precision medicine

A Marshall University – University of Missouri team has reported a web-based deep-learning platform that combines six common ‘omics’ data types to predict phenotypes and highlight candidate biomarkers, with potential uses that range from risk stratification in complex disease to trait prediction in plant breeding

A collaboration between researchers from Marshall University, Huntington, West Virginia, and the University of Missouri, Columbia, Missouri, both in the US, has developed a web-based platform that uses artificial intelligence to integrate multiple biological datasets in order to more accurately predict complex outcomes. The system – called G2PDeep-v2 – has aimed to make multi-omics analysis more accessible by placing model building, automated optimisation and downstream interpretation within a single browser-based workflow.

Dr. Trupti Joshi was named as the corresponding author for the paper which framed the core problem in familiar terms for translational research: modern studies can measure gene expression, microRNA expression, protein abundance, DNA methylation and genomic variation within the same cohort but teams often lack practical ways to combine these layers without substantial computational support.

G2PDeep-v2 has set out to integrate six data types that recur across biomedical and agribiotech studies:

• gene expression
• microRNA expression
• protein expression
• DNA methylation
• single nucleotide polymorphisms
• copy number variations.

In principle, each layer captures a different aspect of biology. Gene expression and protein expression report which molecular programmes cells have switched on, DNA methylation offers one window into regulation, and sequence variants and copy number changes can point to inherited risk or acquired genomic disruption. The hope is that a model that ‘sees’ several layers at once can outperform models that rely on just one.

In the paper, the authors described G2PDeep-v2 as a deep-learning framework for phenotype prediction and biomarker discovery across organisms, including both humans and plants. Here ‘phenotype’ means the measurable trait or outcome of interest which could range from disease status, survival or treatment response in people to yield, resilience or other agronomic traits in crops. The same basic task underpins both domains: use molecular measurements to predict what happens next and to identify which features carry the most signal.

One practical barrier for many laboratories has nothing to do with biology and everything to do with logistics. Deep-learning models typically require decisions about architecture and training settings, plus access to enough computing power to iterate. G2PDeep-v2 has attempted to reduce that friction by offering an interactive interface and automated hyperparameter tuning. This is a method to search systematically for training settings that improve performance without requiring users to perform manual trial and error. The authors also described links to high-performance computing resources which can matter once datasets move from dozens of samples to hundreds or thousands.

The platform has also emphasised interpretation. Beyond prediction, the authors said users can visualise outputs and then run Gene Set Enrichment Analysis on significant markers, with the aim to connect a list of candidate features to plausible pathways and mechanisms. The intention is to stop the workflow at ‘useful biological story’ rather than at ‘high accuracy number’.

“This collaboration demonstrates the power of academic partnerships in accelerating breakthroughs that can transform patient care,” said Dr. Trupti Joshi, professor and senior associate dean for informatics and population analytics at the Joan C. Edwards School of Medicine at Marshall University.

“G2PDeep equips researchers with an accessible, sophisticated tool to better understand and address the health challenges facing our communities,” she said.

“Our collaboration with Dr Joshi’s lab has been both long-term and productive,” said Dong Xu, a bioinformatics professor in the Department of Electrical Engineering and Computer Science at the University of Missouri.

“We focus to develop bioinformatics platforms, not just algorithms, that can be widely adopted by the research community. G2PDeep demonstrates how advanced AI methods can be translated into practical tools with broad scientific impact,” he said.

The team and its institutional partners have highlighted prospective applications that span cancer, addiction, ageing, obesity and kidney disease, alongside other conditions that have affected populations in Appalachia and the Midwest. That breadth should not surprise anyone who has watched the steady rise of multi-omics studies. Once a platform can accept common molecular inputs, the limiting factor tends to become data quality, cohort design and clinical annotation rather than the nominal disease area. Still, the platform’s practical value will depend on how well it copes with the unglamorous realities of real-world datasets, including missing values, batch effects and uneven cohort sizes.

G2PDeep-v2 is available online, and the authors have positioned it for use by researchers, clinicians and public health professionals who want to analyse multi-omics datasets without building an in-house machine-learning stack from scratch.

Whether it becomes a staple tool will rest on the same old questions that separate durable infrastructure from flashy prototypes: transparent documentation, reproducible outputs, robust privacy and security choices, and evidence that performance holds up outside the datasets that shaped development.

For further reading please visit: 10.3390/biom15121673