Predictive Patient Stratification in Bleeding Disorders - Multi Omics Classification for Pediatric Hemostaseology 

Overview

Problem

Bleeding disorders in children, including hemophilia A, hemophilia B, von Willebrand disease, and thrombophilia, are clinically heterogeneous. Patients with the same diagnosis can respond very differently to the same therapy. Gene therapy developers and rare disease focused biotechs need to stratify pediatric patients into clinically meaningful subgroups before trial enrollment, but the multi omics data required for this does not exist commercially.

Solution

A secure evaluation platform is required to benchmark patient stratification models on real pediatric multi omics data without the data leaving the clinical institution. tracebloc enables participating teams to submit classification models that integrate genomic, transcriptomic, proteomic, and clinical features, evaluating which architectures and integration strategies produce the most reliable stratification in a controlled, reproducible environment.

Outcome

To be completed after evaluation concludes.

The Underlying Dataset: SCIVIAS Study

For this challenge, the hemostaseology subset of SCIVIAS is the focus: pediatric patients with hemophilia A, hemophilia B, von Willebrand disease, and thrombophilia. A substantial proportion of patients are genetically confirmed, providing high diagnostic certainty. The cohort is longitudinal (baseline plus 2 year post therapy), and all five omics layers are available. Feature names in the challenge dataset are anonymized to protect the underlying clinical data structure.

Context and Background

The Operational Challenge

Pediatric bleeding disorders span a wide clinical spectrum. Two children with hemophilia A can have fundamentally different molecular profiles, different comorbidity patterns, and different responses to factor replacement or gene therapy. Current clinical practice stratifies patients primarily by factor activity levels and bleeding phenotype, but these clinical markers alone are poor predictors of long term treatment response, particularly for gene therapy durability and inhibitor development risk.

The question pharma and biotech companies face: can multi omics data stratify patients into subgroups with distinct treatment trajectories? If so, this directly informs trial design (enrichment strategies, endpoint selection), companion diagnostic development, and eventually clinical decision making. The challenge is that the data required to answer this question, pediatric multi omics with longitudinal treatment follow up across multiple bleeding disorder subtypes, does not exist in any commercially accessible form.

Why Multi Omics Matters for Stratification

Single omics approaches capture one layer of biology. Genomic mutations define the genetic basis of disease but say nothing about how the mutation manifests at the protein or metabolic level. Transcriptomics captures gene activity but misses post translational regulation. Proteomics sees the functional molecules but not their genetic drivers. Metabolomics reflects downstream pathway activity but not upstream causes.

Effective patient stratification in heterogeneous diseases like bleeding disorders requires integrating across these layers. A patient subgroup defined by the convergence of a specific mutation pattern, a transcriptomic signature, and a proteomic profile is biologically more meaningful (and clinically more actionable) than one defined by any single layer alone. Models that can identify these multi layer subgroups produce stratifications that are more likely to replicate in independent cohorts and more likely to predict treatment response.

Features that consistently drive stratification across multiple omics layers also become candidates for downstream biomarker validation and, where biological plausibility supports it, target discovery. The stratification task is the computational foundation on which these higher order questions are built.

The Data

Participants work with a multi omics dataset (480 samples, 282 features) derived from the SCIVIAS hemostaseology cohort. The dataset contains four feature blocks representing genomic mutations, gene expression levels, protein measurements, and clinical phenotype variables. Feature names are anonymized (mutation_0, gene_0, protein_0, clinical_0, etc.) to protect the underlying clinical data structure. The four feature blocks correspond to real molecular and clinical measurements from the original cohort.

The classification target is a binary label. The dataset is complete with no missing values.

Why Current Workflows Fall Short

Patient stratification using multi omics data requires both the data and the analytical infrastructure to be in the same place. Pharma and biotech companies have the ML capacity but lack pediatric multi omics cohorts. Academic medical centers have the cohorts but lack systematic model benchmarking infrastructure. Transferring de identified multi omics data externally creates re identification risk, especially in rare disease populations where patient profiles are inherently distinctive. tracebloc resolves this: models travel to the data, the data stays at the institution, and evaluation happens inside a standardized, auditable environment.

Problem Definition

Core AI Task

Binary classification on integrated multi omics features. The model receives genomic mutation profiles, gene expression levels, protein measurements, and clinical variables, and must predict the binary outcome label. The challenge tests two capabilities simultaneously: raw predictive performance (can you classify accurately?) and multi omics integration (does combining layers outperform single layer models?).

Evaluation Metric

Mean Squared Error (MSE). Lower is better. For binary classification, MSE penalizes confident wrong predictions quadratically, making it more sensitive to catastrophic misclassifications than log loss. This is appropriate for a clinical stratification task where confidently placing a patient in the wrong subgroup has real consequences.

Technical Constraints

• Compute budget: 2 CPUs (8.59 GB RAM), 1 GPU (22.49 GB VRAM), capped at 100 PetaFLOPs total.
• All training and inference inside the tracebloc secure environment: no data export, no model weight extraction.
• Invite only access.

Dataset Characteristics

For a detailed overview, see EDA section.

Dimensionality

282 features across 480 samples. PCA shows 178 components are needed to capture 90% of variance, indicating high effective dimensionality. The feature to sample ratio is manageable but regularization remains important.

Feature Structure

Class Distribution

Approximately balanced: the two classes split roughly 50/50.

Why Structured Evaluation Matters Here

Multi omics integration is methodologically open: early fusion, late fusion, stacking, attention based architectures, and tree ensembles all represent viable strategies with different trade offs. Without a controlled benchmarking environment, it is impossible to determine whether one approach genuinely outperforms another or whether differences reflect data handling, compute allocation, or random variation. tracebloc standardizes the evaluation surface so that performance differences are attributable to modeling choices alone.

tracebloc Solution and Approach

Platform Architecture

tracebloc operates as a secure AI evaluation platform. Participating teams interact through a controlled API. They receive exploratory data analysis outputs to understand the dataset, then submit model code that executes inside the tracebloc infrastructure. Raw patient data never leaves the secure environment. Model weights are not extractable. Only aggregate performance metrics are returned.

Evaluation Workflow

• Exploratory Data Analysis: Participants have access to a detailed EDA covering feature distributions, dimensionality analysis, correlation structure, and class balance to inform modeling decisions without accessing raw records.
• Model Submission: Researchers submit model code via the platform. Execution runs on allocated compute.
• Automated Benchmarking: Each submission is scored on MSE with results posted to a live leaderboard.
• Auditability: Every submission, metric, and compute usage is logged. Results are fully reproducible.

Evaluation Dimensions

Predictive Performance

Primary: MSE on the binary classification target. Models are ranked on the leaderboard by this metric.

Operational Performance

Compute efficiency within the 100 PF budget. Multi omics integration models can be computationally expensive, particularly deep learning approaches. Resource aware architecture selection is part of the challenge.

Scientific and Clinical Value

Beyond raw MSE, the scientifically interesting question is whether models that integrate across all four feature blocks outperform those that rely on a single layer. If multi omics integration consistently improves stratification, this validates the premise that cross layer signal matters for bleeding disorder subtyping, and the features driving that improvement become candidates for downstream biomarker validation and target discovery work.

Outcome and Results

To be completed after evaluation concludes.

Lessons Learned and Future Directions

To be completed after evaluation concludes.