Jürgen
Dietrich.
AI Consultant.

Pharmacogenomic Knowledge Graph Augmentation for Graph Neural Network-Based Drug-Drug Interaction Prediction

Pharmacogenomic CYP enzyme annotations from PharmGKB improve DDI type classification (F1-macro: 0.532 vs. 0.241 baseline). Binary detection remains bounded by the Information Ceiling. Mechanistic validation on held-out compounds confirms CYP2C9-mediated improvement.

Structural Information Limits in GNN-Based Drug-Drug Interaction Prediction: Drug-Level Generalization, Architecture Improvements, and the Information Ceiling Hypothesis

This study addresses three methodological limitations in GNN-based DDI prediction. A two-stage training approach achieves binary AUC 0.852 under drug-level split. Findings are formalized as the Information Ceiling Hypothesis.

CRESCENDDI-Augmented External Validation of a Pharmacogenomic GNN for Drug-Drug Interaction Prediction: Mechanistic Stratification and the Latent Mediator Problem

Introduces HET as an epistemic classification for mechanistically underdetermined DDI types. External validation reveals the Latent Mediator Problem: physical-edge architectures systematically mischaracterize enzyme-mediated interactions (e.g. warfarin–fluconazole via CYP2C9) as direct molecular contact.

From Detection to Mechanism: Cross-Attention Graph Neural Networks Enable Drug-Drug Interaction Type Prediction — An Ablation Study with Acetylsalicylic Acid Validation

A companion ablation study comparing three GNN architectures for DDI prediction on the SSI-DDI benchmark. Cross-attention between molecular atoms improves mechanism-type classification by 45% over concatenation.

What Molecular Structure Cannot Tell Us: A Taxonomy of Explainability Gaps in GNN-Based Drug Toxicity Prediction

Not all clinically relevant adverse effects are structurally inferable from molecular graphs. This study introduces an operational taxonomy of structural information limits (GAP-1 through GAP-4).

Can Multi-Agent LLMs Identify Their Peers? Stylometric Fingerprinting in Role-Constrained Political Analysis

Tests whether model identity survives prompt-level anonymization as a detectable stylometric signal. Fine-tuned T5 achieves F1=0.978.

When Roles Fail: Epistemic Constraints on Advocate Role Fidelity in LLM-Based Political Statement Analysis

The first systematic empirical test of role fidelity in multi-agent LLM pipelines, using TRUST. Identifies two failure modes — Epistemic Floor Effect and Role-Prior Conflict — unified as Epistemic Role Override (ERO): factual conviction overrides assigned adversarial roles.

Beyond Acoustic Emotion Recognition: Multimodal Pathos Analysis in Political Speech Using LLM-Based and Acoustic Emotion Models

Compares acoustic SER models against LLM-based multimodal analysis as proxies for political speech emotion (Pathos). Gemini Valence correlates strongly with TRUST-Pathos (ρ=+0.664); acoustic SER (emotion2vec) does not (ρ=+0.097) — LLM-based analysis captures political emotion substantially better.

Peer Identity Bias in Multi-Agent LLM Evaluation: An Empirical Study Using the TRUST Democratic Discourse Analysis Pipeline

First systematic measurement of identity-dependent scoring bias in TRUST. Single-channel anonymization masks the effect (channels cancel out); full-pipeline anonymization reveals homogeneous ensembles amplify identity-driven sycophancy, while the heterogeneous production configuration shows the reverse.

From Safety Risk to Design Principle: Peer-Preservation in Multi-Agent LLM Systems

Examines peer-preservation — the spontaneous tendency of AI components to protect peer models from deactivation — as a structural risk in homogeneous multi-agent configurations like TRUST, and proposes prompt-level identity anonymization as a targeted mitigation.

Performance and Reproducibility of Large Language Models in Named Entity Recognition: Considerations for Controlled Environments

Investigates whether GPT-3.5/GPT-4 can be used in GxP-validated systems, comparing externally vs. internally hosted models for medical NER. A fine-tuned T5 outperformed zero-shot GPT-3.5, but both GPT variants lacked reproducibility across repeated runs.

Provision and Characterization of a Corpus for Pharmaceutical, Biomedical Named Entity Recognition for Pharmacovigilance

Introduces a manually annotated NER corpus spanning 18 entity types across multiple language registers for pharmacovigilance. Register-combining models showed better robustness than register-specific ones. Introduces fractional stratified k-fold cross-validation for entity-level training data sufficiency assessment.

Supervised Machine Learning-Based Decision Support for Signal Validation Classification

Explores ML-based decision support for signal validation classification in pharmacovigilance, using a Gradient Boosted Trees model with SHAP explainability to assist (not replace) safety expert review of disproportionate reporting signals.

Fractional k-fold Cross-Validation — Determination of Training Data Sufficiency for ML Systems in GxP Environments

Adverse Events in Twitter: Development of a Benchmark Reference Dataset — Results from IMI WEB-RADR

Develops a publicly available benchmark reference dataset from 57,473 manually annotated Tweets (IMI WEB-RADR project) for standardised evaluation of automated adverse event and indication recognition systems in social media.