3IA PhD/Postdoc Seminar #51

Secreted proteins as early, sensitive indicators of chemical bioactivity

Abstract: High-throughput proteomics allows systematic profiling of secreted proteins across multiple compounds and doses, yet remains underused in toxicology compared with transcriptomics or image-based approaches. Here, we report the first comprehensive dose–response analysis of the secretome using Nomic Bio’s platform. HepaRG cells were exposed to 37 compounds across 10 concentrations, and 984 extracellular proteins were quantified. Protein-level dose–response modeling was applied to estimate benchmark doses (BMDs) as points of departure (PODs), which were compared with PODs from Cell Painting and cytotoxicity assays. Responses were compound-specific, highlighting biologically interpretable changes in inflammation, stress, and liver-related pathways. For some compounds, proteomic and Cell Painting PODs were closely aligned, while others showed modality-specific sensitivity. Overall, moderate to strong correlations between proteomic and morphological PODs, along with clustering of protein responses by mechanism, show that secretome profiling provides complementary, early and mechanistically informative insights for chemical safety assessment.

Principles of Analytic Provenance for Intelligence Augmentation : a case study of search interactions

Abstract: Today, vast amounts of information are readily accessible, and it would be absolutely impossible for users to find what they are looking for without the right tools. As we will see, since exploration of information spaces rarely follows predetermined paths, the process involves navigating unfamiliar terrain, often requiring backtracking, revisiting, and refining the approach. As users travel this path, they leave behind search trails: the sequences of steps, decisions, and interactions that document their exploration. These trails capture how users navigate through various sources and ideas, revealing the progression of their reasoning and discovery. As users branch out into subtopics, refine hypotheses, or revisit previous lines of inquiry, the flat, temporal structure of a chat log becomes a bottleneck. New tools based on Large Language Models (LLMs), such as GPT, have significantly transformed the landscape of information seeking. Two fundamental challenges arise: (1) the difficulty of revisiting and reusing relevant information hidden in long conversation threads, and (2) the lack of a structural organization that reveals the conceptual relationships between the ideas explored. As a result, users lose context, duplicate exploration steps, and struggle to form a coherent overview of their search journey. In this presentation, we analyze the principles of analytical provenance (which involves documenting and tracking steps, decisions, and data transformations) and how these principles can be implemented to create intelligent user interfaces that augment user interaction with exploratory searches. These ideas are illustrated through a new interactive interface based on mind maps, implemented using a tool called ChatInVis, which has a browser extension that extends the interface and capabilities of an existing mind map visualization tool to support LLM-based information search interactions.

Open discussion about all the contributions