AI steps into drug safety: predicting liver injury earlier than ever before

Drug-induced liver injury (DILI) continues to be one of the most persistent and costly challenges in drug development. Hepatic safety issues that go undetected in preclinical testing can surface later in patients, affecting both clinical trial outcomes and post-market safety. In several cases, promising therapeutics have advanced through years of development before being withdrawn due to toxicity concerns that standard assays failed to detect. It is a fraught area for industry and one with increasing urgency as the sector shifts towards more human-relevant testing strategies.

Cellarity has now published a landmark paper in Nature Communications describing an AI framework for predicting and characterising DILI using transcriptomic data. Central to this are DILImap and ToxPredictor, a resource and model that integrate large-scale toxicogenomics with machine learning. According to the publication, the system outperformed more than twenty preclinical safety assays and identified clinical liver toxicity cases missed by animal models.

To explore the significance of this development, Drug Target Review spoke with Parul Doshi, Chief Data Officer at Cellarity, about the origins of the project, how the model operates and what it means for future discovery efforts.

A platform grounded in computation and biology

Doshi joined Cellarity in 2021 and leads the company’s data and computational strategy, covering machine learning, data infrastructure, high-performance computing and software engineering. Her work ensures that computational outputs integrate seamlessly with biological insights.

Cellarity is a clinical-stage biotechnology company developing cell state-correcting therapies. The platform combines multi-omics with artificial intelligence to understand disease in terms of cell state, which allows researchers to observe how compounds alter cellular programmes and pathways instead of focusing on a single target.

This approach provided the basis for an AI-driven strategy to address DILI prediction. Traditional preclinical models test individual pathways linked to toxicity, but they offer only narrow snapshots of hepatotoxicity mechanisms. Enzyme readouts, imaging markers and pathology panels highlight specific endpoints, but they do not capture the complex cell state perturbations that can later present clinically.

The motivation

The challenge of DILI was already well recognised within the Cellarity team. Doshi explains that DILI is the leading cause of toxicity-related drug withdrawal, noting that “DILI remains a major barrier to advancing new therapeutics, often derailing promising programmes late in development.” With regulators encouraging more human-relevant methodologies and reduced reliance on animal models, the opportunity to improve predictive accuracy through transcriptomics became compelling. “We believe that an early, transcriptomics-based prediction model can materially reduce this source of clinical and preclinical failure.”

Cellarity began by building DILImap, a vast transcriptomic reference library of 300 compounds. Each was tested at multiple concentrations in primary human hepatocytes, capturing a wide range of DILI mechanisms including mitochondrial failure, oxidative stress and immune activation. This diversity of response profiles proved critical. Instead of testing one failure point at a time, the map captures mechanistic patterns as cells respond across different doses, providing an unbiased window into toxicity pathways. While traditional approaches can build assays for what is already known, this framework reveals emergent signals and pathways that would otherwise remain invisible.

The dataset was then used to train ToxPredictor, a machine learning model that classifies liver injury potential and predicts dose-related risk. According to the Nature Communications study, ToxPredictor reached 88 percent sensitivity at 100 percent specificity in blind evaluation and uniquely identified high-profile clinical failures overlooked by standard preclinical models. Doshi emphasised the strength of this evidence, stating that the model “correctly flagged risks associated with recent high-profile clinical failures that had not been signalled preclinically.”

What makes this different from current models?

Most preclinical safety frameworks still rely on single-endpoint assays, in vitro testing and animal studies. While useful, they do not reflect the complexity of human hepatotoxicity. Doshi notes that large panels of molecular markers have been used previously, yet even these only illuminate a fraction of the cellular network.

Transcriptomics enables a panoramic view of gene regulation changes when a compound engages with a cell. Instead of observing one outcome, researchers see how entire pathways respond. Doshi describes this as a high-resolution lens, revealing early events and latent liabilities that might otherwise remain invisible until late development. “With our platform, we’re using high-dimensional transcriptomics to analyse the complex gene regulations, identifying a broad range of pathway signatures indicative of liver injury.”

Furthermore, the scale of DILImap has been highlighted as the largest toxicogenomics resource created for DILI research. Its size allows machine learning models to generalise across structurally diverse compounds and rare mechanistic triggers, helping reduce false negatives and false positives. Importantly, as the dataset is derived from primary human hepatocytes, it may avoid some of the translational disconnect seen with animal systems.

A new tool for go/no-go confidence

One of the major claims of the Cellarity framework is its potential to inform decision making during early discovery. If transcriptomic risk signals can be detected preclinically, candidates could be deprioritised, modified or dose-adjusted before reaching animal studies or first-in-human trials.

Doshi emphasises that the goal is not merely to flag toxicity, but to explain why it occurs. Mechanistic transparency means medicinal chemists can adjust scaffolds more intelligently and programme leaders can set informed safety margins. “This framework provides an opportunity to bridge the gap between lab findings and human biology,” Doshi explains.

In practice, this could mean more efficient portfolio progression and fewer late-stage failures. Doshi notes that drugs induce different cellular trajectories depending on dose. Mapping dose-dependent inflection points could therefore guide exposure thresholds and inform clinical design. With continual refinement, she envisages ToxPredictor becoming a standard gate in candidate advancement, reducing animal use and aligning with evolving regulatory expectations.

Validation and the roadblocks encountered

Validating any model against real patient outcomes is inherently difficult. Clinical DILI events arise from complex interactions influenced by genetics, comorbidities and environment. Doshi explains that several challenges became clear during evaluation, noting that “Validating any model against real clinical outcomes is complicated, given the range of variables in day-to-day use of medicines across patient populations.” One priority was ensuring that the system classified safe drugs correctly. Excessively conservative systems risk halting promising molecules unnecessarily. This is where transcriptomic profiling offers an advantage over structure-based predictors, which often over-penalise chemical scaffolds due to superficial similarity.

Another challenge lies in bridging in vitro and in vivo translation. Three-dimensional liver models are more physiologically relevant than two-dimensional cultures, yet even they fail to recapitulate all mechanistic diversity. Doshi highlighted this gap, explaining that “One of the biggest bottlenecks for clinical utility was ensuring that our model correctly classifies safe drugs as safe, thus ensuring that high potential drug candidates are not discontinued.”

Opening the framework to the wider community

Cellarity has released DILImap, ToxPredictor and the underlying data as open source to encourage wider evaluation and use. They hope this transparency will support integration into other discovery platforms and stimulate further method development. Doshi notes that several avenues exist for future enhancement, including integrating transcriptomics with proteomics, metabolomics, morphological readouts, pharmacokinetics and chemical structure descriptors.

Expansion to encompass co-perturbations such as inflammatory stimuli or metabolic stress could capture compound behaviour under disease-relevant conditions. Additionally, extending into advanced 3D culture systems may further improve translational confidence.

Reducing animal reliance

The publication arrives at a moment when toxicology is increasingly focused on identifying clinical risk earlier and more accurately. Many compounds that appear acceptable in animal studies ultimately fail in the clinic due to human-specific toxicity. By capturing transcriptomic responses that are not predicted by animal models, AI-driven frameworks like this offer a path to flagging such liabilities sooner – potentially reducing the number of compounds advanced into animal studies that are unlikely to translate to humans.

As Doshi summarises, applying machine learning to toxicogenomics holds great promise for making drug discovery more efficient and ultimately improving patient safety. The framework set out in Nature Communications represents more than a model alone, inviting a more holistic approach to safety evaluation in early development.