Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2051
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-20T22:03:48.115936Z
status: published
lens: g_space
sector: science
headline: vToxiNet: An interpretable deep learning framework for hepatotoxicity prediction and mechanism exploration
dek: Hepatotoxicity remains a leading cause of drug attrition and post-marketing withdrawal, resulting from diverse and complex toxicity mechanisms. Traditional in vitro models can only capture a limited subset of toxicity pathways, and animal studies face translational and ethical limitations. Regulatory agencies have therefore promoted new approach methodologies, including human-relevant assays, omics technologies, and computational models to improve predictive toxicology and support evidence-based decision-making.…
gain_title: Embedding pathway hierarchy into a neural network lets the model predict liver toxicity and attribute predictions to genes and pathways, supporting mechanism-informed chemical prioritization.
problem_title: (none)
trace_subject: (none)
gain_reading: Embedding pathway hierarchy into a neural network lets the model predict liver toxicity and attribute predictions to genes and pathways, supporting mechanism-informed chemical prioritization.
gain_evidence: interpretation of vToxiNet enables gene and pathway-level attribution, supporting mechanism-informed hazard characterization and chemical prioritization
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers built the virtual toxicity network, a deep learning model that combines chemical descriptors, high-throughput assay responses, transcriptomic signatures and the Reactome pathway hierarchy. The pathway hierarchy is built into the network's structure, forming a virtual adverse outcome pathway network that predicts whether chemicals cause liver toxicity.

Hepatotoxicity is a leading cause of drug attrition and post-marketing withdrawal, and many existing machine learning models use only chemical structure or act as black boxes. A model that also points to genes and pathways could support hazard assessment, though the text reports only promising performance and gives no figures or real-world deployment.
limitation: The abstract describes performance only as promising and presents the work as support for further development, with no quantitative results or evidence of regulatory or clinical use.
tag: Evidence-backed gain
key_points: Hepatotoxicity is described as a leading cause of drug attrition and post-marketing withdrawal. | Most existing machine learning models for it use chemical structure alone or work as black boxes. | vToxiNet combines chemical descriptors, assay responses, transcriptomic signatures and Reactome pathway hierarchy. | Results were checked with cross-validation and external validation and called promising. | Interpretation gives gene- and pathway-level attribution for hazard characterization and chemical prioritization.
rundown: The paper frames the work against limits of current toxicity testing: in vitro models capture only a subset of toxicity pathways, and animal studies face translational and ethical limitations. Regulators have promoted new approach methodologies, including human-relevant assays, omics technologies and computational models.

vToxiNet embeds systems toxicology knowledge directly in its architecture rather than adding explanations afterward. By using multimodal data and biological hierarchy as architectural constraints, it aims to combine prediction with mechanistic interpretation.

The authors present it as a framework supporting further development of mechanism-informed chemical hazard assessment, not as a finished regulatory tool.
sources:
- peer_reviewed | Journal of Hazardous Materials | https://doi.org/10.1016/j.jhazmat.2026.143880 | 2026-10-12
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

b926fdd75660b03d545696034f84f7c4757277da6230c1639aaee078189e6a4b

previous

0000000000000000000000000000000000000000000000000000000000000000