39
17.09.27
On this page
About this use case: Six research institutions each hold roughly 40 patients per treatment arm — individually underpowered, collectively enough to find the combination therapy signal none of them can reach alone. tracebloc federates the training across all six sites so every institution benefits from the combined 240-patient signal without any site sharing a single patient record with another. Explore the data, submit your own model, and see how your approach compares.
No single partner site has enough combination therapy patients for reliable multi omics analysis. Each of six research institutions holds approximately 40 patients per treatment arm. Individually, that sample size puts every site in the underpowered regime — insufficient to distinguish genomic from proteomic from metabolomic contributions to combination response. Pooled, 240 patients across three treatment arms give every consortium member access to the signal none of them could reach alone.
Dr. Yuki Tanaka, Head of Clinical Data Science at the coordinating institution, deploys a tracebloc workspace loaded with 240 multi-omics patient samples — 78 on DrugA monotherapy, 78 on DrugB monotherapy, and 84 on the combination. Partner institutions submit their response prediction models to the workspace. Inside tracebloc's containerised training environment, models train on the full 240-patient sample set — adapting gene expression, protein, and metabolite feature weights to the cross-site multi-omics distribution — without any site's patient data leaving its own infrastructure. tracebloc orchestrates federated training, scores each adapted model against the evaluation cohort, and publishes results to a live leaderboard. This is a federated learning application of multi-site scientific collaboration: no patient data moves between institutions, but every institution benefits from the combined signal.
In this example evaluation, models trained on the combined 240-patient multi-omics dataset show consistent improvement over models trained on any single institution's 40-patient cohort — particularly in distinguishing combination responders from monotherapy-equivalent patients, which requires the statistical power that only cross-site pooling can provide. The tracebloc workspace stays in place as the consortium expands and new patient samples are added at each site.
The consortium was formed three years ago. Six institutions — two academic medical centres, three pharma-affiliated research sites, and one rare disease patient registry — agreed to collaborate on combination therapy response prediction in a paediatric oncology-adjacent indication. The scientific rationale was clear: combination targeted therapy offers better outcomes for a subset of patients, but predicting which patients benefit requires multi omics integration across biological layers that no single dataset has captured well enough to build a reliable classifier.
The agreement stalled immediately on data sharing. Each institution's legal team had a different view of what cross-border patient data transfer required. Two institutions were subject to GDPR. One was operating under US research regulations. Two had ethics approvals that explicitly limited data sharing to de-identified aggregate summaries. The sixth — the patient registry — had explicit consent language that precluded any data transfer to commercial partners.
After eighteen months of legal negotiation, the consortium had produced no shared dataset and no shared model. The institutions had individually trained response prediction models on their own cohorts — each with approximately 40 patients per arm across three treatment groups — and discovered that none of them had enough statistical power to reliably identify the combination-specific signal. At 40 patients per arm, the confidence intervals on combination benefit were too wide to support clinical decision-making. The model that needed the combined data most was the one most stymied by the data governance problem.
The scientific problem is structural. Multi-omics therapy response prediction — integrating gene expression (~100 features), protein levels (~100 features), metabolomics (~90 features), and clinical variables (20 features) across 293 total predictors — is a high-dimensional, small-sample problem. With only 40 patients per arm per site, any model is at high risk of fitting noise rather than genuine biological signal. The p/n ratio at any single site exceeds 1 across the multi-omics feature space. Regularisation can partially address this, but it cannot substitute for patient volume.
The scientific questions the consortium most needs to answer — which omics layer carries the combination-specific signal, how do treatment-by-omics interactions differ across monotherapy and combination arms, and whether a compact biomarker panel can serve as a companion diagnostic — require the 240-patient cross-site view. None of them are answerable from any single site's cohort.
Omics data analysis at the consortium level also raises a second problem: batch effects. Each site collected samples under slightly different protocols, stored them differently, and used different assay platforms for parts of the proteomics panel. Pooling raw data would require harmonisation — a process that introduces its own analytical uncertainty and takes months to complete even with full data access. Federated training, where models are trained on each site's native data distribution and weights are aggregated rather than data being centralised, is actually the statistically preferable approach for this consortium's design.
The evaluation dataset contains 240 multi-omics patient samples across three treatment arms. Full dataset statistics, feature distributions, and treatment arm composition are available in the Exploratory Data Analysis tab.
This dataset is augmented. It was constructed to reflect the statistical structure of a real-world multi-omics combination therapy trial — the treatment arm distribution, the gene expression variance, the metabolomics covariance structure, and the continuous response magnitude distribution — without containing any identifiable patient records or trial data.
| Property | Value |
|---|---|
| Total samples | 240 |
| Treatment arms | 3 — DrugB monotherapy (84), DrugA monotherapy (78), Combination (78) |
| Features | 293 |
| Gene expression features | ~100 (continuous, transcriptomic measurements) |
| Protein level features | ~100 (continuous, proteomic measurements) |
| Metabolite features | ~90 (continuous, metabolomic profiles) |
| Clinical variables | 20 (continuous) |
| Missing values | None |
| Target | Continuous treatment response score |
| Evaluation metric | MSE |
A note on the p/n ratio: With 293 features and 240 samples, this is a high-dimensional, small-sample dataset — p/n ratio exceeds 1. Overfitting is the primary technical risk. Models that perform well here have solved a genuine regularisation problem, not just fitted a training set. This is intentional: it reflects the realistic statistical challenge of multi-omics therapy response prediction in rare disease populations where patient volumes are structurally limited.
Class distribution note: Treatment arms are approximately balanced — DrugB (35%), DrugA (32.5%), Combination (32.5%). No arm dominates the sample, which supports treatment-arm-stratified evaluation and avoids the confounding that unbalanced arm assignment would introduce.
Each partner institution submitted their response prediction model to the tracebloc workspace. The evaluation ran in two phases.
Phase 1 — Single-site baseline. Each institution's model was benchmarked as trained on their own ~40-patient cohort, with no access to cross-site data. This establishes the honest baseline: what each institution can achieve working alone, and how much signal is left on the table by not having access to the other five sites' data.
Phase 2 — Federated fine-tuning. Partner institutions were given access to the training environment inside the tracebloc workspace, which aggregated patient samples from all six sites into a 240-patient combined training set. Each institution transferred their model into the workspace and ran training on the full 240-patient sample set. This fine-tuned the model weights to the combined multi-omics distribution — adapting from a model calibrated on one site's 40 patients to one that has seen the full therapeutic and molecular diversity of the consortium. After training, the adapted model was evaluated against the held-out evaluation cohort. Patient data from each institution never left that institution's infrastructure. No institution had visibility into another's training data or raw samples; only model outputs were shared through the tracebloc orchestration layer.
→ View the full model leaderboard — complete institutional rankings, MSE improvement from single-site to federated, and omics-layer attribution across all submissions.
| Institution | Single-Site MSE | After Federated Training | Combination-Arm MSE | Primary Omics Signal |
|---|---|---|---|---|
| Site A | 0.061 | 0.038 | 0.044 | Gene expression |
| Site B | 0.074 | 0.041 | 0.051 | Proteomics |
| Site C ✅ | 0.058 | 0.031 | 0.036 | Metabolomics + Gene |
| Site D | 0.079 | 0.045 | 0.053 | Proteomics |
| Site E | 0.066 | 0.039 | 0.047 | Gene expression |
| Site F | 0.083 | 0.048 | 0.059 | Clinical variables |
What the numbers reveal:
Every institution improved through federated training — the smallest gain was Site A, which went from 0.061 to 0.038 MSE, a 38% improvement. The largest gain was Site F, which went from 0.083 to 0.048 — a 42% improvement. Both numbers make the same point: the signal that each institution was missing was sitting in the other five sites' data, and federated training retrieved it without any site sharing patient records.
Site C achieves the strongest post-federated MSE at 0.031, and its combination-arm MSE of 0.036 is the sharpest result in the evaluation for the arm that matters most clinically. Site C's architecture explicitly models treatment-by-omics interactions — it was built to identify cases where the combination produces effects not predictable from either monotherapy alone. Its metabolomics-plus-gene-expression signal profile aligns with the biological hypothesis: downstream metabolic consequences of dual-pathway inhibition are visible in metabolomics before they appear in protein levels.
Site F shows the largest improvement through federated training precisely because clinical variables alone provide the weakest single-site baseline. Adding the molecular signal from four more institutions reduces Site F's MSE by 42% — the clearest quantification in this evaluation of how much information was being left unreachable.
Illustrative assumptions: 6 consortium institutions / 120 patients newly enrolled per year across all sites (20 per site) / €28,000 cost per patient enrolled in a protocol that would be terminated if the combination benefit prediction model lacks sufficient statistical power / companion diagnostic development programme value: €4.2M if model reaches clinical-grade performance on 240+ patients
| Scenario | Model MSE | Companion Dx Readiness | Patient Programme Cost (Year 1) | Consortium Value |
|---|---|---|---|---|
| Single-site (best) | 0.058 | Not credible at 40 pts/arm | €560,000 per site | Low — no publication-grade power |
| Federated (tracebloc) ✅ | 0.031 | Clinically credible | €560,000 total shared | High — companion Dx programme viable |
The federated approach reduces per-institution cost by sharing the patient enrolment load across six sites while giving every site access to the combined 240-patient signal. The companion diagnostic programme — which requires clinical-grade prediction performance to be viable — moves from aspirational to achievable at 0.031 MSE on the combination arm.
The consortium adopts Site C's architecture as the shared model for companion diagnostic development, based on its combination-arm MSE performance and its explicit treatment-interaction modelling. A federated training protocol is formalised: as each site enrols new patients, the tracebloc workspace is updated and all institutions re-run the federated training cycle. The model improves continuously without any site ever sharing patient records with another.
The tracebloc workspace stays active after the initial evaluation. Each new cohort cycle — quarterly, as new patients are enrolled across all six sites — re-runs the federated fine-tuning and updates the leaderboard. Institutions can submit updated model versions as their bioinformatics teams iterate. The scientific benefit of collaboration is captured without the legal and ethical costs of centralisation.
Explore this use case further:
Related use cases: See how the same federated approach applies to genomic biomarker stratification, prognostic transcriptomics in neuromuscular disease, and pharmacodynamic proteomics validation. For a broader view of federated learning applications across pharma and healthcare, see our federated learning applications guide.
Deploy your workspace or schedule a call.
Disclaimer: The dataset used in this use case is augmented — designed to reflect the statistical structure of a real-world multi-omics combination therapy trial, including treatment arm distribution, gene expression variance, metabolomic covariance, and continuous response magnitude distribution, without containing any identifiable patient records or trial data. The persona, institutional labels, performance figures, business impact assumptions, and consortium scenario are illustrative and based on patterns observed across multi-centre rare disease and oncology research environments. They do not represent any specific institution, company, clinical trial, or consortium arrangement.
