Research questions and open gaps
A map of what is unsolved, organised so you can find a thesis-sized problem. Each gap names why it is hard and what a contribution could look like. Strong PhD projects usually sit where a biological question, a data resource and a methodological weakness meet.
Cross-cutting gaps
1 · Causality, not correlation
Most multi-omics results are associations. Combining genetic anchors (QTLs, MR), perturbations (CRISPR, drugs) and time series to build causal multi-layer models is a central open problem.
Contribution: a causal integration method validated against Perturb-seq or CRISPR screens.
2 · Ground truth for evaluation
Many fields (spatial domains, GRNs, multi-omics integration, cell communication) lack ground truth, so methods cannot be compared fairly.
Contribution: a benchmark dataset or calibrated simulator, plus a neutral comparison.
3 · Context specificity
Variant and regulatory effects depend on cell type, state, stimulus and development. Current reference maps cover few contexts.
Contribution: dynamic or single-cell QTL mapping; models that transfer effects across contexts.
4 · Diversity and equity
Reference panels, GWAS, QTL resources and training data for predictors over-represent European ancestry, which limits accuracy elsewhere.
Contribution: multi-ancestry fine-mapping, PRS or QTL methods; analysis of under-represented cohorts.
5 · Foundation models, honestly evaluated
Large pre-trained models for DNA, RNA, proteins and cells are proliferating, but evaluations frequently show they do not beat simple baselines on key tasks.
Contribution: rigorous evaluation, failure analysis, or task-specific models that demonstrably add value.
6 · Uncertainty and calibration
Most pipelines output point estimates. Propagating uncertainty from raw signal to biological conclusion (quantification, imputation, latent factors) is rare.
Contribution: Bayesian or conformal methods that give calibrated intervals.
7 · Interpretability
Deep models predict well but do not explain. Linking latent dimensions to pathways and mechanisms remains hard.
Contribution: biologically constrained architectures (pathway-informed, sparse) with validation.
8 · Scale, privacy and federation
Biobanks with millions of participants and atlases with hundreds of millions of cells strain current algorithms; human data often cannot leave its institution.
Contribution: scalable or federated algorithms, sketching, out-of-core methods.
Gaps by layer
| Layer | Open problem | Why it is hard | Possible project angle |
|---|---|---|---|
| Genomics | Interpreting non-coding and structural variants in rare disease | Few labelled examples; effects are context-specific | Integrate long-read SV calls with tissue RNA-seq outliers and regulatory maps to raise diagnostic yield |
| Genomics | Pangenome-aware variant calling and association | Tools and coordinates assume a linear reference | GWAS or QTL mapping on graph-genome genotypes; benchmarking in diverse samples |
| Epigenomics | Enhancer–gene linking and enhancer grammar | Sparse CRISPR ground truth; redundancy between enhancers | Sequence-to-function deep models (Enformer, Borzoi-style) evaluated on CRISPRi data in new cell types |
| Epigenomics | Haplotype-resolved methylation and allele-specific regulation | Needed long reads only recently became affordable | Joint phasing, methylation and expression from long-read multi-omics |
| Transcriptomics | Isoform-level function and quantification | Short reads are ambiguous; long reads are lower depth | Long-read differential isoform usage with uncertainty; isoform-to-proteoform validation with MS |
| Single-cell | Multi-sample, multi-condition statistics | Nested design (cells within donors), compositional effects | Mixed-model or pseudobulk frameworks for spatial and multi-modal data |
| Spatial | Segmentation, domains and cell–cell communication with validation | Lack of ground truth; platform-specific artefacts | Benchmarks combining imaging and sequencing spatial data from serial sections |
| Perturbation | Predicting responses to unseen perturbations and combinations | Combinatorial space; models fail to beat linear baselines | Causal or mechanistic priors for perturbation prediction; better evaluation metrics |
| Proteomics | Proteoforms and PTM crosstalk | Bottom-up digestion loses combinatorial information | Methods linking top-down and bottom-up data; PTM site-level integration with phospho-signalling models |
| Proteomics | Plasma proteomics platform harmonisation | Olink vs SomaScan disagreement; epitope effects | Cross-platform calibration using pQTLs and MS as anchors |
| Metabolomics | Annotating the dark metabolome | Few reference spectra; isomers | ML structure and class prediction from MS/MS; integrating genomic context (enzymes present) to constrain candidates |
| Microbiome | From composition to mechanism | Compositional data, confounders, strain-level function | Integrated metagenome–metabolome models with gnotobiotic or culture validation |
| Integration | Mosaic, temporal and causal integration | Missing layers, different timescales, no ground truth | Dynamic latent factor models; integration with genetic anchors (MR within MOFA-like frameworks) |
| Clinical | Translating multi-omics into decisions | Small cohorts, overfitting, cost, regulatory evidence | Minimal biomarker panels validated across cohorts; cost-effectiveness of adding layers |
How to choose a PhD problem in this field
- Start from a biological question that someone cares about (a disease, a developmental process, a drug response), then ask which layer limits the answer.
- Secure data early: public cohorts (TCGA, GTEx, UK Biobank, CELLxGENE), a collaborator's matched multi-omics, or your own. Many methods projects stall for lack of the right dataset.
- Find the methodological weakness using the benchmark hub: where do existing tools disagree, fail on certain data, or lack evaluation?
- Plan a validation: an orthogonal experiment, an independent cohort, a held-out perturbation, or a simulation calibrated to real data.
- Scope three papers: commonly (a) a benchmark or resource, (b) a method, (c) a biological application of the method.
- Release software as a documented package (Bioconductor, PyPI/scverse ecosystem) with tests; it multiplies the impact of the thesis.
- Do single-cell foundation model embeddings improve cell-type-specific eQTL mapping compared with pseudobulk on annotated clusters?
- Can long-read RNA-seq resolve which GWAS-colocalised sQTLs change protein-coding potential?
- Which share of GWAS loci without eQTL colocalisation are explained by caQTLs in disease-relevant cell types?
- How robust are multi-omics cancer subtypes (MOFA+, SNF, iCluster) across independent cohorts and platforms?
- Can cis-pQTL Mendelian randomisation and metabolite QTLs together identify enzyme targets with on-target metabolic biomarkers?
- Does integrating phosphoproteomics with transcriptomics (e.g. COSMOS) predict drug response better than either layer alone in cell-line panels?
- How much do spatial deconvolution results change with the choice of single-cell reference, and can this uncertainty be modelled?
- What fraction of untargeted metabolomic features in human plasma are microbial in origin, and can metagenomes predict them?
- Can calibrated variant effect predictors from protein language models reduce VUS rates in under-represented ancestries?
- How should a multi-omics study allocate a fixed budget between more samples and more layers?