Cancer genomics and computational oncology

Research question

Which molecular alterations in a given tumour actually carry weight for a clinical decision, and how confidently can that be established from the data available?

Methods

Next-generation sequencing analysis across panel, exome and genome scales; somatic and germline variant calling; copy-number and structural variant analysis; variant annotation and tiering against curated knowledge bases.

Clinical relevance

Most molecular reports return more findings than are decision-relevant. Separating the consequential from the incidental is the practical problem, and it is a genomics problem before it is a clinical one.

Research question

How does the molecular composition of a tumour change under selection pressure from treatment, and what does that imply for sequencing therapy?

Methods

Clonal and subclonal reconstruction, phylogenetic inference from multi-region and longitudinal sampling, analysis of mutational signatures and selection.

Clinical relevance

A tumour at relapse is often not the tumour that was profiled at diagnosis. Treating disease as a fixed molecular object explains a good deal of unexpected resistance.

Research question

When does circulating tumour DNA add information that changes a decision, and when is it simply another number?

Methods

Analysis of cell-free DNA sequencing data, variant allele fraction dynamics, longitudinal trajectory modelling, and assessment of assay sensitivity and limits of detection.

Clinical relevance

ctDNA is genuinely informative in specific, bounded situations. Being precise about which situations those are is more useful than general enthusiasm.

Research question

What does cell-level resolution reveal about tumour heterogeneity and the microenvironment that bulk profiling averages away?

Methods

Single-cell RNA sequencing analysis, dimensionality reduction and clustering, cell-type annotation, trajectory inference, and integration with bulk and spatial data.

Clinical relevance

Bulk measurements describe a mixture. Where the mixture itself is the biological question — immune infiltration, resistant subpopulations — averaging is the wrong operation.

Research question

How should genomic, transcriptomic, epigenomic and proteomic layers be combined so that the integration adds signal rather than noise?

Methods

Multi-modal data integration, dimensionality reduction across layers, joint modelling, and batch and platform effect handling.

Clinical relevance

A DNA-level finding often only becomes interpretable when read against expression, methylation or protein evidence from the same tumour.

Research question

Which candidate markers survive the move from a discovery cohort to an independent one, and under what analytical conditions?

Methods

Differential expression and variant association analysis, feature selection with explicit regularisation, cross-validation and independent-cohort replication.

Clinical relevance

The gap between a marker that separates groups in a discovery set and one that holds up elsewhere is where most candidate biomarkers are lost.

Research question

Where can machine learning materially assist oncology decision-making, and where does it mainly add unaccountable complexity?

Methods

Supervised and representation learning on molecular and clinical data, calibration and uncertainty quantification, evaluation against clinically meaningful endpoints rather than accuracy alone.

Clinical relevance

The useful question is not whether a model performs well on a benchmark, but whether its output changes a decision and whether a clinician can tell when to distrust it.

Research question

How can unstructured clinical and scientific text be turned into structured, checkable information without losing the qualifications the text carried?

Methods

Named entity recognition over clinical and biomedical text, relation extraction, negation and uncertainty detection, ontology mapping and normalisation.

Clinical relevance

Most of the clinically important record — pathology narrative, clinic letters, the literature itself — is prose. Structured fields capture only part of it.

Research question

How can genuinely relevant trials be identified for a specific molecular and clinical profile, rather than merely retrieving trials that mention the same gene?

Methods

Structured representation of eligibility criteria, molecular and clinical profile matching, geographic and logistical feasibility filtering, ranked retrieval.

Clinical relevance

The failure mode in trial search is rarely finding nothing. It is finding a long, undifferentiated list in which relevance and eligibility have not been assessed.

Research question

What analytical infrastructure does an oncology service need for molecular interpretation to be reproducible rather than ad hoc?

Methods

Reproducible analysis pipelines, versioned reference and knowledge base management, quality control frameworks, and structured reporting.

Clinical relevance

Interpretation that cannot be reproduced six months later is difficult to audit, difficult to improve and difficult to trust.

Research question

How do germline variants affecting drug metabolism alter the risk profile of systemic anti-cancer therapy for an individual patient?

Methods

Pharmacogenomic variant calling and star-allele assignment, phenotype prediction, and annotation against curated pharmacogenomic guidance.

Clinical relevance

Toxicity risk is part of a treatment decision, not a separate conversation. Some of it is predictable from the germline before the first dose.

Research question

What form should molecular decision support take so that it supports clinical reasoning instead of substituting a black box for it?

Methods

Evidence tiering frameworks, structured actionability categorisation, explicit uncertainty representation, and interface design centred on clinician workflow.

Clinical relevance

Decision support that cannot show its reasoning is difficult to challenge — and a system that cannot be challenged is difficult to use safely.

Research question

How should evidence be assembled and weighed for diseases where large randomised trials do not and may never exist?

Methods

Systematic evidence synthesis across small series and case reports, cross-disease molecular reasoning, registry and real-world data analysis.

Clinical relevance

Rare cancers are not simply under-studied common cancers. They require a different information architecture, because the usual evidence hierarchy has gaps in it by construction.