Monash University Launches DECODE: A Game-Changer for Multiomics Deconvolution
In a major advancement for bioinformatics, researchers from Monash University's Monash AI Institute have published 'A unified framework for multiomics deconvolution' in the prestigious journal Nature Methods on March 2, 2026. Led by PhD candidate Xiaoyu Wang and corresponding author Professor Jiangning Song from the Department of Biochemistry and Molecular Biology, the paper introduces DECODE, a deep learning-based algorithm that unifies cell type and cell state deconvolution across transcriptomics, proteomics, and metabolomics data. This breakthrough addresses the fragmentation in current tools, enabling more accurate analysis of bulk tissue samples.
The publication highlights Monash's leadership in AI-driven biomedicine, with funding from Australian Research Council grants and Monash's inter-disciplinary projects. As bulk omics data remains cost-effective and widely used, DECODE promises to unlock deeper insights into complex diseases like cancer.
Understanding Multiomics: The Foundation of Modern Biology
Multiomics integration combines data from multiple biological layers to provide a holistic view of cellular processes. Transcriptomics captures gene expression via RNA sequencing, proteomics examines protein levels and modifications, and metabolomics profiles small molecules like lipids and sugars. Each 'omic' layer reveals unique aspects: RNA shows active genes, proteins indicate function, and metabolites reflect activity outcomes.
In heterogeneous tissues, such as tumors with mixed cell types (e.g., cancer cells, immune cells, fibroblasts), bulk sequencing averages signals, obscuring individual contributions. This 'mixture problem' limits discoveries in precision medicine, where understanding tumor microenvironments (TME) is crucial for targeted therapies.
Monash's DECODE framework bridges these layers, allowing researchers to deconvolve mixtures using single-cell references, step-by-step transforming averaged data into cell-resolved profiles.
The Challenge of Bulk Omics Data Deconvolution
Deconvolution computationally 'unmixes' bulk data by estimating cell proportions and states from known single-cell signatures. Traditional methods like CIBERSORT or MuSiC excel in transcriptomics but falter in proteomics or metabolomics due to data sparsity, modality differences, and non-linear relationships.
For example, protein data has lower dynamic range than RNA, and metabolites vary widely across cells. Omics-specific tools ignore cross-layer synergies, leading to errors in applications like immune infiltration estimation in cancer biopsies. Studies show up to 30% inaccuracy in tumor purity estimates, impacting immunotherapy predictions.
DECODE tackles this with a modality-agnostic deep learning approach, trained adversarially to align features across omics, improving robustness.
DECODE Unveiled: Architecture and Innovation
DECODE (Deep learning-based Common deconvolution framework) employs a multi-branch neural network with shared encoders for cell type/state signatures and bulk data. It uses contrastive learning to align embeddings and adversarial training to minimize omics biases.
The workflow (Fig. 1 in the paper): (1) Input single-cell references and bulk data; (2) Extract features via autoencoders; (3) Optimize proportions via non-negative matrix factorization-like loss; (4) Output cell fractions and states.
This unified model outperforms specialized tools, as validated on diverse datasets.
Benchmarking DECODE: Superior Performance Across Omics
Benchmarks in the paper demonstrate DECODE's edge. On GTEx transcriptomics, it achieved Pearson correlations >0.95 for cell proportions vs. baselines like Bisque (0.88). In proteomics (e.g., CPTAC cancer cohorts), mean absolute error dropped 25% compared to PROSDECTE.
For metabolomics, using HMDB references, DECODE handled sparsity better, with RMSE 15% lower than lmDeconvolute. Cross-omics tests on TCGA tumors showed consistent TME reconstructions, revealing immune subsets missed by single-omic methods.
- High accuracy in low-purity samples (<20% target cells)
- Robust to noisy references
- Fast inference (<1 min per sample)
These results position DECODE as a standard tool.
Explore the Nature Methods paperApplications in Cancer Research and Beyond
In oncology, DECODE dissects TME from bulk tumors, quantifying fibroblasts driving resistance or macrophages modulating immunotherapy. Applied to TCGA breast cancer, it identified prognostic immune states linked to survival.
Beyond cancer, it aids neurodegeneration (brain bulk proteomics) and immunology (PBMC metabolomics). Australian researchers can now analyze NHMRC-funded cohorts more precisely, accelerating drug discovery.
Real-world case: Deconvolving liver biopsies reveals metabolic shifts in NAFLD, guiding therapies.
Implications for Precision Medicine in Australia
Australia's precision medicine initiatives, like the MRFF, benefit immensely. DECODE enables re-analysis of archived biobanks (e.g., TCGA-AU), uncovering biomarkers without new sequencing costs. It supports personalized treatments, e.g., predicting CAR-T efficacy from protein deconvolution.
Stakeholders: clinicians gain TME insights; pharma firms optimize trials; policymakers fund AI bioinfo hubs.
| Challenge | DECODE Solution |
|---|---|
| Omics silos | Unified model |
| High error in sparse data | DL robustness |
| Slow computation | Efficient inference |
Monash AI Institute: Powering Bioinformatics Innovation
Monash AI Institute drives ethical AI for societal good, with labs like Song's pioneering ML for biomedicine. Highlights include AI drug screening tools and protein design, funded by ARC/NHMRC.
Professor Song's group integrates graphs, VAEs for genomics. DECODE exemplifies Monash's 360+ pubs/year in AI-bio.
For aspiring researchers, research jobs abound at Monash.
Future Outlook: Scaling DECODE for Spatial Omics
Upcoming: Integrate with spatial transcriptomics (Visium), add epigenomics. Community contributions via potential GitHub (check lab site). Challenges: rare cell detection, pan-cancer atlases.
Australian impact: Boosts 1000 Genomes successor, AI Health Flagship.
Career Paths in AI-Driven Bioinformatics Down Under
Australia needs 5000+ bioinformaticians by 2030. Roles: postdocs analyzing multiomics ($120k+), lecturers teaching DL. Monash offers research assistant tips.
- Skills: Python, TensorFlow, scikit-learn
- Certifications: EMBL-EBI courses
- Jobs: postdoc positions
Explore university jobs.
Photo by Jude Al-Safadi on Unsplash
Conclusion: DECODE Paves the Way for Next-Gen Biomedicine
Monash's DECODE marks a milestone in multiomics deconvolution, empowering Australian researchers in precision medicine. Rate professors via Rate My Professor, seek higher ed jobs, or get career advice. Stay tuned for DECODE updates.
