Biology subjects

Scheibe, P.

Publications and source records attributed to Scheibe, P..

3 recordsLinked to original sources

bioRxiv2025

AI-Driven Science Communication: Leveraging LLMs and Knowledge Graphs for Seamless Knowledge Exchange

PurposeScientific knowledge is increasingly captured in structured formats, such as knowledge graphs, yet it remains largely inaccessible to non-technical users. We present EcoToxFred, a prototype conversational AI agent that enables intuitive, natural language access to curated environmental toxicology data. Designed to support users without programming expertise, EcoToxFred facilitates the exploration of complex datasets, such as chemical exposures and species-specific hazard information in European surface waters. MethodsEcoToxFred integrates a large language model (LLM) with a Neo4j graph database via a retrieval-augmented generation (RAG) architecture. The system employs a decision-making agent to interpret user queries, invoke appropriate tools, and translate natural language input into formal graph queries. Outputs are validated and returned in multiple formats, like text, tables, and interactive maps, and are grounded in structured, curated monitoring and hazard data. ResultsThe agent bridges the gap between human intent and formal data retrieval, enabling researchers, policy advisors, and stakeholders to pose complex, multi-step queries without prior training in query languages. By grounding LLM outputs in structured data, we demonstrate the systems ability to respond to diverse question types and deliver transparent, accurate, and context-aware results. EcoToxFred successfully answers broad and highly specific queries, bridging natural language input with formal data retrieval. ConclusionEcoToxFred represents a scalable and transferable framework for human-AI interaction in domain-specific contexts, combining natural language interfaces with structured data. By lowering access barriers to scientific knowledge, the system supports evidence-based decision-making and fosters responsible, human-centered AI use in environmental science and beyond.

pharmacology and toxicology↗

bioRxiv2022

Error quantification in multi-parameter mapping facilitates robust estimation and enhanced group level sensitivity

Multi-Parameter Mapping (MPM) is a comprehensive quantitative neuroimaging protocol that enables estimation of four physical parameters (longitudinal and effective transverse relaxation rates R1 and [Formula], proton density PD, and magnetization transfer saturation MTsat) that are sensitive to microstructural tissue properties such as iron and myelin content. Their capability to reveal microstructural brain differences, however, is tightly bound to controlling random noise and artefacts (e.g. caused by head motion) in the signal. Here, we introduced a method to estimate the local error of PD, R1 and MTsat maps that captures both noise and artefacts on a routine basis without requiring additional data. To investigate the methods sensitivity to random noise, we calculated the model-based signal-to-noise ratio (mSNR) and showed in measurements and simulations that it correlated linearly with an experimental raw-image-based SNR map. We found that the mSNR varied with MPM protocols, magnetic field strength (3T vs. 7T) and MPM parameters: it halved from PD to R1 and decreased from PD to MTsat by a factor of 3-4. Exploring the artefact-sensitivity of the error maps, we generated robust MPM parameters using two successive acquisitions of each contrast and the acquisition-specific errors to down-weight erroneous regions. The resulting robust MPM parameters showed reduced variability at the group level as compared to their single-repeat or averaged counterparts. The error and mSNR maps may better inform power-calculations by accounting for local data quality variations across measurements. Code to compute the mSNR maps and robustly combined MPM maps is available in the open-source hMRI toolbox.

neuroscience↗

bioRxiv2021

AI for predicting chemical-effect associations at the universe level - deepFPlearn

Many chemicals are out there in our environment, and all living species are exposed. However, numerous chemicals pose risks, such as developing severe diseases, if they occur at the wrong time in the wrong place. For the majority of the chemicals, these risks are not known. Chemical risk assessment and subsequent regulation of use require efficient and systematic strategies. Lab-based methods - even if high throughput - are too slow to keep up with the pace of chemical innovation. Existing computational approaches are designed for specific chemical classes or sub-problems but not usable on a large scale. Further, the application range of these approaches is limited by the low amount of available labeled training data. We present the ready-to-use and stand-alone program deepFPlearn that predicts the association between chemical structures and effects on the gene/pathway level using a combined deep learning approach. deepFPlearn uses a deep autoencoder for feature reduction before training a deep feedforward neural network to predict the target association. We received good prediction qualities and showed that our feature compression preserves relevant chemical structural information. Using a vast chemical inventory (unlabeled data) as input for the autoencoder did not reduce our prediction quality but allowed capturing a much more comprehensive range of chemical structures. We predict meaningful - experimentally verified-associations of chemicals and effects on unseen data. deepFPlearn classifies hundreds of thousands of chemicals in seconds. We provide deepFPlearn as an open-source and flexible tool that can be easily retrained and customized to different application settings at https://github.com/yigbt/deepFPlearn. Supplementary informationSupplementary data are available at bioRxiv online. Contactjana.schor@ufz.de

bioinformatics↗