Prediction Of Relapse Through Artificial Intelligence and multi-omics after allogeneic hematopoietic stem cell transplantation – PORTRAIT

This project relies on the French multicenter CRYOSTEM collection, providing plasma and PBMC samples with associated clinical data from over 360 alloHSCT patients (2012–2021). Rigorous quality control (cell recovery, viability, audits) ensures sample reliability. The cohort includes both myeloid and lymphoid malignancies, with or without relapse, enabling cross-disease analyses.

WP1 – Cohort and clinical data
Selection, QC, and anonymization of 2214 samples. Clinical variables (donor/recipient characteristics, GVHD, relapse, survival) will be integrated into a secured database. Risks are limited thanks to cohort size and backup samples.

WP2 – Immune and metabolomic biomarkers

Immune profiling: mass cytometry with barcoding and multiple controls to minimize batch effects. Comprehensive mapping of T, B, NK, and myeloid subsets, analyzed with unsupervised clustering (FlowSOM).

Metabolomics: untargeted plasma profiling by LC-HRMS within MetaboHub, using extensive reference libraries. Uni- and multivariate models will identify metabolic fingerprints associated with relapse.

WP3 – Modeling and simulation
Development of a stochastic simulator to integrate immune, metabolic, and clinical data, accounting for post-transplant events (GVHD, relapse). Iterative pipeline: dimensionality reduction, hybrid supervised selection, calibration via black-box optimization (Artelys Knitro), and cross-validation to prevent overfitting. Simulator performance will be benchmarked against classical survival models (Fine-Gray) and deep learning approaches.

WP4 – Experimental validation
Candidate biomarkers or therapeutic targets will be validated in:

a mouse alloHSCT model reproducing leukemia relapse to investigate immune/metabolic pathways of GVL,

a humanized xenograft model (NSG mice) to test therapeutic molecules directly on human cells.

Following evaluation with the Cryostem scientific committee, the project design was adapted to include two patient cohorts: a discovery cohort of 394 patients to build the predictive tool, and an independent validation cohort of 256 patients. All PBMC and plasma samples from the discovery cohort, collected at key time points (donor before transplantation, 3 months, GVHD onset, 1 year and 2 years after transplantation), have been retrieved and stored in the laboratory.Preliminary analysis revealed an overall relapse rate of 37.3% and a mortality rate of 41.9%, with initial associations explored between relapse and major clinical variables.

WP2.1 – Immune mapping by mass cytometry
Experimental conditions have been optimized, including the use of a single healthy donor as internal reference, a molecular barcoding system, and a 43-marker antibody panel covering surface and intracellular proteins. Data are analyzed through unsupervised clustering with FlowSOM to minimize classification bias. At 12 months, 426 samples from 130 patients had been processed, corresponding to more than 105 million cells analyzed. Average sample viability was 83.6% with 4.3 million cells per sample. This work has already identified 39 distinct immune populations and 25 functional profiles, generating 975 potential variables for relapse prediction.

WP2.2 – Metabolomics
In parallel, around 1,100 plasma samples have been prepared using robotic platforms to ensure reproducibility. Untargeted LC-HRMS metabolomics was completed during summer 2024 for the full discovery cohort. Since September, data processing has been underway, including signal cleaning, annotation through internal reference libraries, and relative quantification of metabolites. Complete metabolic profiles are expected by the end of 2025.

WP3 – Predictive model
The modeling effort led by Artelys has started with the definition of the analytical strategy. Regular meetings are held to coordinate methodology, pending the integration of the immune and metabolic datasets required to train and calibrate the predictive algorithm.

WP4 – In vivo validation
Finally, the Créteil team has finalized and calibrated the mouse allo-HSCT model using MLL-AF9-GFP leukemia cells. This system reproduces key human scenarios: absence of GVHD, induction of GVHD, post-transplant relapse, and graft-versus-leukemia effect. GVHD severity is measured through clinical and histological scores, while relapse is assessed by quantification of leukemic cells. The model is now ready to evaluate therapeutic targets emerging from the other work packages.

This project will be the first to combine a multi-omics approach from a large multicenter cohort of allo-HSCT patients to generate a relapse prediction model using artificial intelligence. It brings together three academic teams with expertise in translational research, transplantation, multi-omics technologies, and animal validation models, with Artelys’ expertise in mathematical tools and complex system simulation.

The primary goal is to establish a composite clinical-immune-metabolic score to predict relapse risk after allo-HSCT. This score, derived from a wide range of clinical and biological variables, will be integrated into a simulation tool available to clinicians to estimate relapse probability and guide therapeutic decisions. At the same time, the project will identify candidate molecules as novel therapeutic targets, paving the way for early-phase clinical trials.

Clinical applications
A simple, early test based on readily available clinical and biological data will help identify high-risk patients and adapt their management before relapse occurs (chemotherapy, immunotherapy, CAR-T cells). In some cases, it could even contribute to optimizing donor selection. The partners’ strong experience in post-transplant biotherapy trials will accelerate clinical translation.

Scientific and medical impact
By applying multi-omics to a large allo-HSCT cohort, this project will advance our understanding of relapse mechanisms and the graft-versus-leukemia effect. It will capture the complexity of post-transplant immune responses and may uncover novel biological processes relevant not only to hematologic malignancies but also to solid tumors. The medical impact will be substantial: tailoring therapeutic strategies to relapse risk with direct benefits for patients.

The project will produce high-impact publications and patents, notably for the relapse prediction simulator. A consortium agreement will define intellectual property sharing and pathways for industrial valorization. Furthermore, the success of the project will strengthen the scientific and commercial positioning of Artelys Knitro, promoting its adoption in the medical field.

Scientific background
Allogeneic hematopoietic stem cell transplantation (alloHSCT) is the first cellular immunotherapy developed to cure hematologic malignancies. It is based on the anti-tumor allo-immune response (graft versus tumor effect) induced by the donor immune system also transferred during the transplant process. Despite its efficiency, hematologic malignancies relapse accounts for half of deceases and to date, no biomarker allow to predict whose patient will relapse after allogeneic HSCT and to identify these patients early before relapse. Traditional statistical methods used for biomarker identifications are limited, mostly by their parametric nature, and could benefit from advanced machine learning and optimization techniques to select relevant variables and link them to the relapsing process. This unmet medical need is of critical importance to improve prognosis of patients who are currently treated for a hematologic cancer with allo-HSCT and to adapt their treatment before relapse.

Hypothesis
Here, we assume that integration of clinical data with immune and metabolic variables could provide metadata for a mathematical model to predict relapse occurrence.

Aims
To characterize circulating immune subsets and metabolome in the donor and to compare them at 3 months and one year after transplantation in patients with or without relapse
To build a calibrated stochastic simulator for the relapsing process, accounting for post-transplant events and integrating clinical data with immune and metabolic variables.

Methodology
This project will rely on a multicentric cohort of 369 patients who received an alloHSCT. We will use mass cytometry and mass spectrometry to decipher circulating immune subsets and metabolites associated with relapse and other post-transplantation events. We will then create a simulator that model the dynamics of post-transplant events to identify relevant biomarkers using advanced optimization techniques and to generate a tool to predict relapse after alloHSCT. Validation in animal model will finally help to identify relevant new therapeutic targets.

Expected results and impact
This project will use data from an already constituted large cohort of patients to develop a machine learning tool for clinicians to estimate the probability of relapse based on various clinical and immune-metabolic data.