Multimodal biomarker discovery for patient stratification

Programme question

Was a single candidate biomarker sufficient, or did reliable patient classification depend on evidence across multiple molecular, imaging and clinical signals?

Before validating any single biomarker, the team needed to know whether classification depended on one dominant signal or the combined weight of evidence across modalities. Sonrai structured that investigation as a progressive evaluation of candidate evidence: separating what each data type contributed, whether biomarkers reflected prognosis or treatment response, and whether combining modalities added genuine value over simpler alternatives.

Case study highlights

A rich dataset, but no clear route from candidate signals to a biomarker strategy

The client had generated transcriptomic, epigenomic, proteomic, histopathology and clinical outcome data through an oncology development programme. Each dataset contained potentially informative features, but high-dimensional analyses could produce many associations without showing which ones were robust, complementary or clinically useful.

The team needed to understand the biological basis of patient heterogeneity and decide which candidate biomarkers merited further validation. That required separating three related but different questions:

• Which features were associated with the underlying prognosis?
• Which features were associated with treatment response?
• Did combining modalities improve stratification enough to justify a more complex biomarker?

Sonrai structured the work as a progressive evaluation of candidate evidence rather than a search for the most statistically impressive signature.

From candidate features to a prioritised biomarker strategy

Stage 1 · Establish what each modality contributes

Transcriptomic, methylation, proteomic and pathology datasets were first assessed independently. This established the quality and characteristics of each data source and identified candidate features associated with relevant clinical endpoints.

Analysing each modality on its own provided an essential reference point. Without that baseline, it would not be possible to determine whether multimodal integration added useful information or merely reproduced signals already captured by a simpler measurement.

Stage 2 · Separate prognosis from treatment response

Candidate molecular and imaging features were evaluated using analyses appropriate to the available outcomes and study design.

Survival and time-to-event methods were used to investigate whether biomarkers distinguished patients with different underlying outcomes. Kaplan-Meier analyses offered an interpretable view of cohort separation, while complementary statistical analyses were used to assess the strength and stability of candidate associations.

Features were also examined in relation to treatment response. The interpretation remained deliberately precise: a biomarker associated with response is not necessarily predictive of differential treatment benefit. A predictive claim would require suitable comparator data, interaction analysis and subsequent validation.

Illustrative examples of how similar survival analyses are performed within Sonrai Discovery; not taken directly from this example.

Stage 3 · Test whether modalities are complementary

Once individual signals had been characterised, integrated analyses examined relationships across molecular, imaging and clinical data.

The key question was not whether a multimodal model could be built. It was whether combining data types produced clinically meaningful improvement over the strongest single-modality or individual-biomarker alternatives.

This analysis helped identify:

• features representing the same underlying biological process
• modalities contributing genuinely complementary information
• composite signatures with improved patient-group separation
• complex combinations that added little beyond a simpler candidate biomarker

Illustrative view of candidate biomarker exploration and cohort separation within Sonrai Discovery; not taken directly from this example.

Stage 4 · Prioritise candidates for validation

Candidates were compared using the totality of available evidence. Prioritisation considered:

• association with a clinically relevant outcome
• distinction between prognostic and response-associated evidence
• biological interpretability
• consistency across analyses and patient subsets
• incremental value over simpler biomarker approaches
• feasibility of measurement in the intended clinical setting
• evidence required for analytical and clinical validation

This shifted the output from a long list of interesting features to a smaller set of defensible candidates with a clearer route to further development.

What each data type added

The resulting decision framework

The programme gave the client a clearer basis for deciding what to validate next. The team could:

• identify molecular and imaging features associated with prognosis
• investigate features associated with treatment response without overstating predictive evidence
• determine where modalities captured overlapping or complementary biology
• quantify whether multimodal combinations improved patient stratification
• compare complex signatures with simpler and potentially more deployable alternatives
• prioritise a focused set of candidate biomarkers for further validation
• preserve the analysis and its provenance as new data became available

The work did not declare an exploratory signature clinically validated. It answered the decision that comes before validation: