Benjamin Ricaud

LC-HRMS is widely used in forensic toxicology for broad-scope screening. When a newly emerging or rarely encountered compound is tentatively identified, toxicologists must decide whether it may be relevant to a case and, if so, quantify it. However, acquiring reference material for quantification is costly and time-consuming. A rapid semiquantitative estimation method would help prioritize only compounds above the toxic threshold. This study presents a machine-learning (ML) framework that estimates drug concentrations in whole blood using molecular structure information and LC-HRMS signals. Using a data set of 191 drugs spiked into whole blood at multiple concentration levels, we trained and evaluated several ML models. Standard models, including Random Forests, achieved moderate performance. In contrast, a recently reported Graph Neural Network (GNN) leveraging atomic features and global molecular properties consistently produced the highest accuracy. Under cross-validation, the GNN predicted signal-to-concentration ratios for 79% of all molecules, corresponding to concentration estimates between 50% and 200% of the true value. Toxicological thresholds often span multiple orders of magnitude, making this precision acceptable. The GNN model was additionally evaluated on an external benchmark data set of ionization efficiencies (logIE), where it outperformed the current state of the art. Overall, the results demonstrate the feasibility of using graph-based ML to estimate drug concentrations in whole blood without reference material. This is a practical ML tool that can support decision-making in toxicological evaluation, particularly for newly emerging or rarely encountered drugs. The GNN model is open source, and the data set used for training and testing the models are publicly available.

The archiving of digital materials related to core international crimes (core crimes) faces significant challenges, including the need for autonomy, protection of privacy, exclusion of political interference, as well as efficient and ethical management of vast and diverse datasets. This paper explores the transformative potential of Generative Artificial Intelligence (AI) and AI-enabled tools in addressing these challenges, with a focus on promoting digital sovereignty. By leveraging open-source technologies, we demonstrate how sovereign, trustworthy and intelligent archival systems can enhance accessibility, ensure data integrity, and reduce reliance on external actors, thereby empowering nations, particularly those affected by armed conflicts, to manage their critical data flows.Through a series of experiments, we evaluate the application of AI in core crimes archiving, including transcription, translation, metadata enrichment, semantic search, and content generation. These tools enable efficient processing of multimedia data, ensuring that sensitive information is preserved, searchable, and accessible in a secure and ethical manner. However, the findings also highlight key challenges, such as the need for AI models to better understand contextual nuances and preserve emotional elements in processed data.The paper further examines the legal and ethical implications of AI-enabled archiving, emphasizing compliance with international regulations, such as the EU AI Act, and the necessity of interdisciplinary collaboration. Ultimately, this research underscores the role of AI in advancing digital sovereignty and robust AI governance, fostering innovation, and contributing to justice, accountability, and the preservation of collective memory in the context of core international crimes.

LC-HRMS is widely used in forensic toxicology for broad-scope screening. When a newly emerging or rarely encountered compound is tentatively identified, toxicologists must decide whether it may be relevant to the case and, if so, quantify it. Acquiring reference material for quantification is costly and time-consuming. A fast semi-quantitative estimate would help prioritize only compounds above the toxic threshold. This study presents a machine-learning framework that estimates drug concentrations in whole blood using molecular structure information and LC-HRMS signal intensities. Using a dataset of 191 drugs spiked into whole blood at multiple concentration levels, we trained and evaluated several machine-learning models. Standard models, including random forests, achieved moderate performance. In contrast, a recently published graph neural network (GNN) leveraging atomic features and global molecular properties consistently produced the highest accuracy. Under cross-validation, the GNN predicted signal-to-concentration ratio for 79\% of all molecules, corresponding to concentration estimates between 50-200\% of true value. Toxicological thresholds often span multiple orders of magnitude, making this precision acceptable for application. The GNN model was additionally evaluated on an external benchmark dataset of ionization efficiencies (logIE), where it outperformed the current state of the art. Overall, the results demonstrate the feasibility of using graph-based machine learning to estimate drug concentrations in whole blood without reference material for prioritization. This is a practical and implementation-ready machine learning tool that can support decision-making in toxicological evaluation, particularly for newly emerging or rarely encountered drugs. The GNN model is open source and the dataset used for training and testing the models is publicly available.

Graph neural networks (GNNs) have emerged as a powerful tool in predicting molecular properties based on structural data. While GNNs excel at identifying local patterns within molecules, their ability to capture global properties remains limited due to inherent structural challenges, such as oversmoothing and their expressivity. We build a simple GNN-based model that integrates chemistry knowledge that GNNs may have difficulties to learn. We show that this combination greatly enhances the accuracy compared with the pure GNN approach. It is on part due to the state of the art (SOTA) of much larger models, including large foundation models, and it even outperforms them in some cases on several benchmarks. With a simple approach, this study highlights some limitations of GNNs and the crucial benefit of giving GNN models easy access to global information about the graph in the context of applications to chemistry. We focus on regression tasks at the molecular level, on small-molecule data sets. We also investigated the possible localization of molecular substructures important for the GNN prediction using the SMILES encoding. We designed a GNN predicting molecule properties at the node level, allowing us to identify important nodes for the prediction. Additionally, the model's architecture allows for efficient training with relatively modest computational resources, making it practical for widespread application.

The exploitation of offshore wind resources is a crucial step towards a clean energy future. It requires an advanced approach for high-resolution wind resource evaluations. We explored the suitability of the Sentinel-1 Level-2 OCN ocean wind field (OWI) product for offshore wind resource assessments. The SAR data were compared to in situ observations and three reanalysis products: the global reanalysis ERA5 and two regional reanalyses CARRA and NORA3. This case study matches 238 scenes from 2022 for the Goliat station, an oil platform located 85 km northwest of Hammerfest in the Barents Sea, where a new offshore wind park has been proposed. The analysis showed that despite their unique limitations in spatial and temporal resolutions, all data sources have similar statistical properties (RMSE, correlation coefficient, and standard deviation). The Weibull parameters characterizing the wind speed distributions showed strong similarities between the Sentinel-1 and all reanalysis data. The Weibull parameters of the in situ measurements showed an underestimation of wind speed compared to all other sources. Comparing the full reanalysis datasets with the subsets matching the SAR scenes, only slight changes in Weibull parameters were found, indicating that, despite its low temporal resolution, the Sentinel-1 Level 2 OWI product can compete with the more commonly used reanalysis products in the estimation of offshore wind resources. Its high spatial resolution, which is unmatched by other methods, renders it especially valuable in offshore areas close to complex coastlines and in resolving weather events at a smaller scale.

Studying real-world networks such as social networks or web networks is a challenge. These networks often combine a complex, highly connected structure together with a large size. We propose a new approach for large scale networks that is able to automatically sample user-defined relevant parts of a network. Starting from a few selected places in the network and a reduced set of expansion rules, the method adopts a filtered breadth-first search approach, that expands through edges and nodes matching these properties. Moreover, the expansion is performed over a random subset of neighbors at each step to mitigate further the overwhelming number of connections that may exist in large graphs. This carries the image of a “spiky” expansion. We show that this approach generalize previous exploration sampling methods, such as Snowball or Forest Fire and extend them. We demonstrate its ability to capture groups of nodes with high interactions while discarding weakly connected nodes that are often numerous in social networks and may hide important structures.

An emerging way to deal with high-dimensional non-euclidean data is to assume that the underlying structure can be captured by a graph. Recently, ideas have begun to emerge related to the analysis of time-varying graph signals. This work aims to elevate the notion of joint harmonic analysis to a full-fledged framework denoted as Time-Vertex Signal Processing, that links together the time-domain signal processing techniques with the new tools of graph signal processing. This entails three main contributions: (a) We provide a formal motivation for harmonic time-vertex analysis as an analysis tool for the state evolution of simple Partial Differential Equations on graphs. (b) We improve the accuracy of joint filtering operators by up-to two orders of magnitude. (c) Using our joint filters, we construct time-vertex dictionaries analyzing the different scales and the local time-frequency content of a signal. The utility of our tools is illustrated in numerous applications and datasets, such as dynamic mesh denoising and classification, still-video inpainting, and source localization in seismic events. Our results suggest that joint analysis of time-vertex signals can bring benefits to regression and learning.