Secure Biomarker Panel Narrowing in Inflammatory Bowel Disease

Overview

Problem

Pharmaceutical companies developing biologics for pediatric inflammatory bowel disease need to validate biomarker panels before committing to $150M+ Phase III trials. Adult validated biomarkers frequently fail in children due to differences in immune maturation and disease biology, but recruiting a pediatric cohort with longitudinal multi omics data takes 18 to 24 months and costs $500K to $2M.

Solution

A secure, reproducible evaluation platform is required to benchmark candidate biomarker models against real pediatric longitudinal data, without the data ever leaving the clinical institution. tracebloc enables pharma companies to submit model architectures and candidate biomarker lists, run thousands of training iterations on protected patient data, and receive ranked performance outputs, all inside a controlled federated environment.

Outcome

[To be completed after evaluation.]

The Underlying Dataset: SCIVIAS Study

SCIVIAS — Seeing Childhood Illness through Multi Omics

SCIVIAS is a monocentric observational study conducted at the Dr. von Hauner Children’s Hospital, LMU Munich, led by Prof. Dr. Dr. Christoph Klein. The study combines retinal imaging (fundus photography, OCT) with multi omics profiling (genome, transcriptome, proteome, metabolome) to identify early diagnostic markers for rare and chronic childhood diseases.

The core premise: children with rare diseases are often diagnosed only when complications arise. SCIVIAS aims to change this by integrating pattern recognition on retinal images with multi layer omics data, using machine learning to detect disease signatures before clinical manifestation. All omics data and retinal images are pseudonymized and processed through ML algorithms, comparing data both within defined disease groups and across phenotypes to uncover pleiotropic factors.

The cohort consists of 2500 patients and covers 13 therapeutic areas including IBD (Crohn’s, ulcerative colitis, celiac disease), cystic fibrosis, Duchenne muscular dystrophy, spinal muscular atrophy, and other rare pediatric conditions.

Ethics approval: LMU Munich, approval no. 17–801. German Clinical Trials Register: DRKS00013306.

Study page: https://www.ccrc-hauner.de/clinical-research/scivias-study

For this tracebloc evaluation, the IBD subset of SCIVIAS is the primary clinical target: Crohn’s patients, ulcerative colitis, biopsy confirmed celiac diseases, and healthy controls, all with longitudinal data from baseline through 2 years post therapy. The platform challenge uses a derived stratification genomics dataset (a few hundred samples, 254 features) that tests the core modeling task on a representative subset of the cohort’s genetic and clinical feature space.

Context and Background

The Operational Challenge

Biologics for IBD (Crohn’s disease, ulcerative colitis) are among the fastest growing segments in gastroenterology, with drugs from Takeda, AbbVie, Pfizer, and Johnson & Johnson competing for pediatric indications. The standard path: validate biomarkers in adult trials, then attempt translation to children. This translation step fails more often than it succeeds. Pediatric immune systems are structurally different, treatment response trajectories diverge, and the molecular signatures that predict remission in adults may be noise in a 7 year old.

The result is a bottleneck. Pharma teams have hundreds of candidate biomarkers from their adult programs and need to know which subset actually predicts treatment response in children before designing their pediatric trial endpoint. Getting this wrong means a failed trial. Getting it right means a faster, cheaper path to approval.

Stakeholders

Pharma biomarker teams: Translational scientists responsible for selecting panel endpoints for Phase III.
Clinical investigators (LMU Munich): Data custodians of the SCIVIAS pediatric cohort, providing longitudinal multi omics under strict data governance.
Regulatory and compliance teams: Ensuring that evaluation workflows satisfy GDPR, EHDS, and institutional data use agreements.

The Data

The SCIVIAS cohort provides five omics layers (genomics, bulk and single cell transcriptomics, proteomics, metabolomics) with longitudinal coverage from baseline through 2 years post therapy. Phenotyping is exceptionally deep: a 900 question HPO based questionnaire captures comorbidities, treatment history, and clinical progression. Celiac cases are confirmed by both serology and biopsy, reducing diagnostic noise.

For the platform evaluation challenge, participants work with a stratification genomics dataset (a few hundred samples, 254 features) derived from the broader SCIVIAS cohort. This dataset contains CFTR mutation profiles (100 variants), dystrophin gene structural variants (60 exon level features), SMN gene variants, modifier genes, and clinical phenotype measurements. It tests the core classification task that underpins the broader biomarker validation use case.

Why Current Workflows Fall Short

Traditional approaches require physical data transfer or on site analyst access, both of which are slow, expensive, and create compliance risk. Sending de identified omics data to external parties still exposes re identification risk given the rarity of pediatric rare disease profiles. Internal validation by clinical teams lacks the ML engineering capacity to rigorously benchmark model architectures. The gap between what pharma needs (structured model comparison at scale) and what academic medical centers can provide (raw data under restrictive DUAs) is where tracebloc operates.

Problem Definition

Core AI Task

Multi class classification: stratify patients by disease category using genomic variants, modifier genes, and clinical phenotype features. The task is a proxy for the broader biomarker validation problem, where the goal is to identify which features carry the strongest predictive signal for patient stratification and, by extension, treatment response prediction.

Evaluation Metric

Mean Squared Error (MSE) on disease classification. Lower is better. MSE penalizes large misclassification errors more heavily than accuracy, making it suitable for a setting where confident wrong predictions (e.g., misclassifying a CF patient as Duchenne) are costlier than uncertain ones.

Technical Constraints

Compute budget: 2 CPUs (8.59 GB RAM), 1 GPU (22.49 GB VRAM), capped at 100 PetaFLOPs total.
Maximum 5 submissions per participant.
All training and inference runs inside the tracebloc secure environment: no data export, no model weight extraction.

Dataset Characteristics Relevant to Modeling

Dimensionality

254 features across a few hundred samples (p/n ratio ~0.8). PCA analysis shows 123 components are needed to capture 90% of variance. This is a high dimensional regime where regularization and feature selection are not optional.

Why Structured Evaluation Matters Here

Without a controlled benchmarking environment, comparing biomarker selection approaches across teams is impossible. Different preprocessing, different splits, different compute environments introduce confounds that make it unclear whether Model A outperforms Model B because of its architecture or because of data handling differences. tracebloc standardizes the evaluation surface: same data, same compute constraints, same metric, same submission pipeline.

tracebloc Solution and Approach

Platform Architecture

tracebloc operates as a secure AI evaluation and vendor selection platform. Participating teams (vendors, internal ML groups, academic collaborators) interact exclusively through a controlled API. They receive metadata summaries and exploratory data analysis outputs to understand the dataset structure, 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 receive a detailed EDA (feature distributions, variance analysis, PCA, correlation structure, outlier profiles) to inform modeling decisions without accessing raw records.
Model Submission: Teams submit model code via the platform. Execution runs on allocated compute (2 CPU, 1 GPU, 100 PF cap). Up to 5 submissions per team.
Automated Benchmarking: Each submission is scored on MSE with results posted to a live leaderboard. Iterative improvement is possible within the submission budget.
Auditability: Every submission, metric, and compute usage is logged. Results are fully reproducible.

Evaluation Dimensions

Predictive Performance

Primary: MSE on disease stratification. Secondary: per class performance to ensure no single disease group is being sacrificed for aggregate score. The leaderboard currently shows LightGBM, XGBoost, and CatBoost as leading architectures, with LightGBM achieving the best MSE of 0.4544.

Operational Performance

Inference speed and compute efficiency within the 100 PF budget. Models that exhaust compute on training leave no headroom for hyperparameter search. Resource aware architecture selection is part of the challenge.

Business and Domain Impact

The downstream question: which model architecture produces the most clinically useful biomarker ranking? A model that achieves strong MSE by relying only on clinical features (age, FEV1) is less valuable than one that identifies genomic variants, because clinical features are already known to clinicians. The models that surface novel, actionable genomic signal are the ones that matter for trial endpoint design.

Omics Biomarker Panel Narrowing in Inflammatory Bowel Disease