Pré-Publication, Document De Travail Année : 2025

Large Language Models-guided Agentic Discovery of Imaging-informed Pharmacodynamic Models for Predicting Hematological Toxicity from Radioligand Therapy

Résumé

Peptide receptor radionuclide therapy (PRRT) with [¹⁷⁷Lu]Lu-DOTATATE has transformed treatment of neuroendocrine tumors, yet persistent thrombocytopenia affects up to 10% of patients and increases risk of secondary malignancies. Current predictive approaches for hematological toxicity remain limited, hindering personalized dosing strategies. Here, we implement and evaluate an imaging-based kinetic pharmacodynamic (IKPD) modeling framework that integrates quantitative imaging biomarkers with mechanistic models of platelet dynamics to predict treatment-related toxicity. We relied on HDTwinGen, a hybrid digital twin generator that autonomously uses large language models to create computational models combining mechanistic differential equations with neural components. First, we developed it further to make it suitable for clinical pharmacology applications (Euler-based integration with positivity enforcement and temporal alignment, multi-compartment support, BIC-driven model selection with per-patient parameter estimation, visualization and automatic reporting tools, and secure local inference for GDPR compliance). Using synthetic datasets, the autonomous system successfully recovered canonical pharmacokinetic-pharmacodynamic structures, including the Friberg neutropenia model, directly from sparse clinical data without prior specification. When applied to our institutional cohort of 21 PRRT patients, HDTwinGen autonomously generated mechanistically plausible models incorporating imaging-derived covariates (osteomedullary invasion score, total metabolic tumor volume, spleen length) extracted via attention U-Net from pretreatment [⁶⁸Ga]Ga-DOTATOC PET/CT scans. While expert-driven nonlinear mixed-effects models in Monolix achieved superior statistical precision (BIC ≈ 397.5 versus ≈ 2,139), the autonomous approach demonstrated feasibility for rapid structure discovery in domains with limited pharmacometric expertise. Local deployment using open-source language models preserved data privacy while maintaining mechanistic interpretability. This framework represents a step toward democratizing personalized dosing in radioligand therapy, potentially enabling pre-treatment risk stratification and adaptive protocols as PRRT expands across oncologic indications. The integration of quantitative imaging with autonomous model discovery offers a scalable approach to precision medicine in radiopharmaceutical therapy.

Fichier non déposé

Dates et versions

hal-05241541 , version 1 (05-09-2025)

Identifiants

  • HAL Id : hal-05241541 , version 1

Citer

Romain Ferrara, Victor Gertner, Adil Moussali, Solène Malmon, Louise Bouchard, et al.. Large Language Models-guided Agentic Discovery of Imaging-informed Pharmacodynamic Models for Predicting Hematological Toxicity from Radioligand Therapy. 2025. ⟨hal-05241541⟩
317 Consultations
0 Téléchargements

Partager

  • More