MDTR: a knowledge-guided interpretable representation for quantifying liver toxicity at transcriptomic level

Inyoung Sung, Sangseon Lee, Dongmin Bang, Jungseob Yi, Sunho Lee, Sun Yeou KimFrontiers in Pharmacology 2025 · DOI

Abstract

In experiments with the LINCS dataset, we show that MDTR outperforms existing methods for measuring the distance of transcriptome data when describing for dose-dependent drug perturbations. In addition, MDTR shows interpretability at the level of DILI mechanisms in terms of the distance, i.e., in a metric space. Furthermore, we provided a user-friendly and freely accessible website (http://biohealth.snu.ac.kr/software/MDTR), enabling users to easily measure DILI in drug-induced transcriptome data.

Introduction

Drug-induced liver injury (DILI) is a primary cause of drug attrition during development and a leading factor in acute liver failure. Traditional toxicological assessments rely heavily on animal models, which are resource-intensive and often fail to translate accurately to human biology. High-throughput transcriptomics—measuring gene expression changes in cell cultures—offers a scalable alternative. However, analyzing this data is challenging due to high dimensionality, low sample sizes, and the inherent noise of biological systems. Most existing metrics provide a single toxicity score that lacks mechanistic context, making it difficult for researchers to understand the underlying cause of a drug's adverse effects.

The MDTR Framework

To address these challenges, we developed the Multi-Dimensional Transcriptomic Ruler (MDTR). This framework translates complex gene expression data into an interpretable, five-dimensional representation of hepatotoxicity. MDTR integrates data-driven outlier detection with curated biological knowledge to provide a quantitative assessment of drug effects across specific toxicological mechanisms.

Data-Driven Outlier Detection (Dual-SVDD)

MDTR employs a two-step Support Vector Data Description (SVDD) approach. Recognizing that toxic responses are distinct from standard drug-induced changes, the model first defines a "drug-effect" boundary to separate treated samples from controls. It then identifies "potentially toxic" (PT) samples as outliers falling outside a secondary "toxic-effect" boundary. This approach effectively handles the biological diversity of drug responses by treating toxicity as a deviation from expected behavior.

Knowledge-Guided Representation

The core of MDTR’s interpretability lies in its mapping of gene expression to five established DILI mechanisms:

  1. Oxidative stress
  2. Immunological response
  3. Altered lipid metabolism
  4. Mitochondrial dysfunction
  5. Bile acids accumulation

These mechanisms are mapped to specific biological pathways using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database.

The knowledge-driven view of the MDTR framework, mapping literature-derived DILI mechanisms to KEGG biological pathways.

Quantification in Kernel Space

To capture complex, non-linear interactions between genes, MDTR maps data into a non-linear kernel space using a Radial Basis Function (RBF) kernel. Within this space, the Mahalanobis distance is calculated to measure how far a sample deviates from the "normal" distribution. These distances are weighted and aggregated into five mechanism-level scores, which can be visualized as a toxicity "fingerprint" via radar charts.

Experimental Results

The framework was evaluated using the LINCS L1000 dataset, cross-referenced with clinical databases such as DILIrank, LiverTox, and the Comparative Toxicogenomics Database (CTD).

Dose-Response Tracking

In tests involving 198 drug-dose combinations, MDTR significantly outperformed standard mathematical distances (e.g., Euclidean, Cosine) and other pathway-based methods. It achieved a 70% positive ratio in correctly identifying dose-dependent perturbations, demonstrating its sensitivity to varying levels of drug exposure.

Mechanistic Interpretability

MDTR successfully distinguished between toxic and non-toxic compounds across nearly all mechanisms with statistical significance (p < 0.05). Case studies on drugs like Mitoxantrone revealed "hormesis" effects—where low doses appear less toxic than controls before a toxicity spike—while drugs like Doxorubicin and Rosiglitazone showed distinct spikes in specific mechanistic dimensions, providing clues to their modes of action.

Conclusion and Availability

MDTR provides a robust, interpretable framework for quantifying liver toxicity at the transcriptomic level. By moving beyond black-box scores to a multi-dimensional "ruler," it offers researchers deeper insights into the biological drivers of DILI. Future work will focus on expanding the gene sets and automating pathway mapping.

The MDTR tool is freely accessible as a web service at http://biohealth.snu.ac.kr/software/MDTR.

Summary updated

Abstract

In experiments with the LINCS dataset, we show that MDTR outperforms existing methods for measuring the distance of transcriptome data when describing for dose-dependent drug perturbations. In addition, MDTR shows interpretability at the level of DILI mechanisms in terms of the distance, i.e., in a metric space. Furthermore, we provided a user-friendly and freely accessible website (http://biohealth.snu.ac.kr/software/MDTR), enabling users to easily measure DILI in drug-induced transcriptome data.
BibTeX
@inproceedings{sung2025mdtr,
  author = {Inyoung Sung and Sangseon Lee and Dongmin Bang and Jungseob Yi and Sunho Lee and Sun Yeou Kim},
  title = {{MDTR: a knowledge-guided interpretable representation for quantifying liver toxicity at transcriptomic level}},
  booktitle = {Frontiers in Pharmacology},
  year = {2025},
  doi = {10.3389/fphar.2024.1398370},
}
Figure