Cellarity Unveils AI-Powered ToxPredictor Model to Revolutionize Drug Safety Testing

A large-scale human toxicogenomics resource for drug-induced liver injury prediction - Nature Communications

This work, integral to a broader AI/ML drug discovery platform, aims at enhancing predictive power and operational efficiency in drug development. It showcases that the shift from a single-target to a systems-level perspective holds great promise and positions machine learning in toxicogenomics as significant enhancement to existing methods. To advance the field and foster collaborative innovation, we have made our open-source model and validation data publicly available at dilimap.org, providing a powerful tool for de-risking drug candidates and setting the stage for a paradigm shift in safety evaluations. We have created DILImap, a comprehensive RNA-seq library tailored for drug-induced liver injury (DILI) modeling, encompassing 300 compounds tested at four concentrations. As the most extensive toxicogenomics resource to date, DILImap includes a curated selection of DILI-positive and DILI-negative compounds that span a wide range of known DILI mechanisms, including well-documented liver-injuring drugs and idiosyncratic compounds with no characteristic signature (Fig. 1A). All compounds were screened in sandwich-cultured primary human hepatocytes (PHHs), the gold standard and most physiologically relevant in vitro model for liver toxicity, which preserve key hepatic functions such as metabolic activity and bile canaliculi formation. Each compound was tested in triplicate across six concentrations using lactate dehydrogenase (LDH) and Adenosine Triphosphate (ATP) cell viability assays. RNA-seq profiling was performed at four selected doses, spanning the pharmacologically relevant range from therapeutic plasma Cmax to the highest tolerated non-cytotoxic dose just below the IC₁₀ threshold (Supplementary Fig. S1). We selected a 24-hour post-exposure time point, based on the trade-off between signal strength and cellular viability: earlier time points (e.g., 2 h or 8 h) yield weaker transcriptional responses, while longer incubations risk hepatocyte de-differentiation and RNA degradation. This strategy allowed us to capture early transcriptional responses without compromising RNA integrity. Marker analysis confirmed retention of hepatocyte identity at 24 hours. We further ensured data quality by including only wells with sufficient total RNA counts and low mitochondrial RNA content, indicating viable, transcriptionally active cells. This streamlined workflow -- including automated solubility testing, viability screening, and IC₁₀-based dose selection -- enabled us to profile 300 compounds in four months, including 110 drugs tested preclinically for the first time as part of a systematic benchmark (Supplementary Fig. S2; Methods). To support comprehensive benchmarking, we provide detailed annotations for each compound, including clinical DILI labels, DILI mechanisms, molecular information, consensus plasma Cmax from various studies, and DILI classification results from over 20 pre-clinical studies (Supplementary Data S1-S4). The training dataset includes 249 compounds (111 DILI + , 52 DILI-, 17 unlikely DILI, 69 idiosyncratic DILI) for cross-validation. For blind validation, a separate experiment was conducted using an independent set of 51 compounds (33 DILI + , 14 DILI-, and 4 with unknown labels, including real-world clinical failures). This carefully curated dataset provides a robust foundation for predictive modeling and mechanistic insights into DILI. ToxPredictor, a machine learning model trained on our DILImap library, predicts DILI risk from pathway-level transcriptional signatures. These signatures are derived through enrichment analysis (WikiPathways, FDR-adjusted p-values) of genes differentially expressed between compound- vs. DMSO-treated samples using DESeq2, computed for each dose of every compound in DILImap (Fig. 1B; see Methods). For model training, we used only compounds with unambiguous DILI labels, resulting in a high-confidence training set of 111 DILI+ (Withdrawn, Known, Likely) and 52 DILI- (No DILI), while the remaining training data were held out to assess model robustness. To ensure high-confidence DILI labels, we further restricted training to drug concentrations tested at more than 20x of their clinical Cmax to reduce the risk of false-negative labeling for DILI+ compounds that may appear safe at lower doses. For 5-fold cross-validation, we applied stratified, compound-level splitting to ensure that all doses and replicates of a given compound were held out together in each fold, mimicking real-world generalization to unseen compounds. From 193 tested configurations across eight model classes, we selected a Random Forest classifier for its strong validation AUC, minimal overfitting, and highest consistency across folds. These properties, combined with its interpretability, motivated its choice over more complex boosting and deep learning models (Supplementary Fig. S4). The final model is an ensemble of 30 Random Forest models (ensemble members) trained on different cross-validation splits, which together enhance generalization and prediction stability. The ensemble size was chosen based on empirical benchmarking that showed stable test AUC and consistency between models (Supplementary Fig. S5). By estimating DILI probabilities across dose levels, ToxPredictor enables calculation of drug safety margins, defined as the ratio between the first predicted DILI dose (i.e., the lowest dose with predicted probability >0.7) and the maximum plasma concentration (Cmax) at therapeutic levels. This provides a transcriptomics-based surrogate of the clinical therapeutic window. A safety margin threshold of 80 provides an actionable classification into high- and low-risk compounds. The probability threshold of 0.7 and margin of safety (MOS) cutoff of 80 were both optimized on held-out training data to reach performance plateaus while minimizing false positives. Our selected MOS threshold of 80, while on the higher end of literature-reported ranges (10-100), reflects the greater sensitivity of transcriptomic assays compared to cytotoxicity or mechanistic readouts. Since transcriptional changes often occur at lower doses -- before overt toxicity -- a higher cutoff is needed to avoid false positives and maintain high specificity in a transcriptome-based model (Supplementary Fig. S5). Among available exposure measures, we used total Cmax instead of free Cmax due to its broader availability across compounds. Both measures showed comparable predictive performance, with total Cmax performing slightly better, possibly due to more robust consensus values derived from a greater number of studies (Supplementary Fig. S6). All model selection, hyperparameter tuning, and threshold optimization were performed exclusively on the training data. For final evaluation, we used a fully independent blind-validation set of 51 compounds (33 DILI + , 14 DILI - , and 4 unknowns), profiled in a separate experiment using separate plates and sequencing runs. This set was withheld from all stages of model development. Compound selection for the validation study was finalized prior to training and intentionally enriched for withdrawals and recent clinical failures. The four unknowns represent compounds currently in clinical use or trials without confirmed DILI liability (Supplementary Table S1). In blind validation, the model achieved 88% sensitivity, correctly identifying 29 of the 33 DILI+ compounds, and 100% specificity, with all 14 DILI- compounds correctly classified as safe (Fig. 1C). These results represent a substantial improvement over our initial proof-of-concept model trained on TG-GATES microarray data, which achieved 62% sensitivity and 92% specificity. Leveraging our DILImap library, ToxPredictor substantially improved both sensitivity of 88% and specificity of 100% on the same validation set (Fig. 2A). In cross-validation of the entire library, the model identified 110 out of 144 DILIs -- surpassing the previous 62 out of 144 with TG-GATES -- and misclassified only 8 out of 66 non-DILIs (Fig. 2B, Suppl. Figure S7). This enhancement is attributed to DILImap's larger dataset with broader mechanistic coverage and the higher resolution of RNA-seq over microarrays, enabling better gene detection and a wider quantitative range for expression level changes compared to microarrays. Notably, post-market withdrawals, missed in both pre-clinical models and clinical trials, were most confidently flagged by our model as high DILI risk. Our DILI safety margin and classification is derived from three parameters: Cmax (baseline concentration), cell viability assays (indicating cell death), and transcriptomics (based on differential pathways). To assess each parameter's contribution, we assessed their ability to classify DILI cases independently. Out of 144 DILIs, 29 were detected solely based on plasma Cmax (>25 μM) and 42 through the LDH cytotoxicity assay (safety margin <80), both at ≥90% specificity. Combining our transcriptomics-based model with Cmax and LDH data was most effective, identifying 110 out of 142 DILIs (safety margin <80), underscoring the added value of toxicogenomics in DILI detection beyond mere cell death (Supplementary Fig. S8). ToxPredictor achieved a ROCAUC of 0.82 in cross-validation, compared to 0.66 for viability alone. In blind validation, it achieved a ROCAUC of 0.96, compared to 0.65 for using viability alone (Supplementary Fig. S9). Our model highlights distinct DILI profiles among closely related COX-2 inhibitor non-steroidal anti-inflammatory drugs (NSAIDs) and imparts unique mechanistic insights linking predictions to mechanisms such as hepatocellular injury, oxidative stress, and mitochondrial dysfunction. For instance, Valdecoxib, used for cancer pain, shows no DILI risk (Fig. 3A), while Sulindac, an arthritis treatment with rare but established idiosyncratic DILI cases, and Lumiracoxib, withdrawn due to severe liver failures, are flagged as DILI risks with safety margins below the classification threshold of 80 (Fig. 3B). Crucially, it highlights the pathways implicated in DILI, encompassing direct contributors like oxidative stress leading to cell injury, as well as indirect factors such as disturbances in fatty acid metabolism, which can be particularly relevant to explain idiosyncratic effects (Suppl. Table S2). Sulindac, for example, is linked to disruptions in fatty acid synthesis and cholesterol biosynthesis, aligning with recent studies connecting it to hepatic steatosis. By pinpointing these pathways, the model provides mechanistic insights into idiosyncratic DILI, offering an understanding previously thought unpredictable (Fig. 3C). The model assesses DILI risks in a dose-resolved manner, revealing how dosage impacts liver injury likelihood. A key demonstration is its accurate prediction of Sulindac's DILI risk, despite it being believed to be unpredictable in a dose-resolved manner. These results highlight the model's ability to deliver actionable predictions and enable targeted optimization of drug safety profiles by focusing on critical pathways (Fig. 3D). DILI arises from disruptions in diverse pathways. Our model highlights pathways with high predictive value (AUC ≈ 0.8) strongly associated with DILI risk, including amino acid metabolism (toxic metabolite buildup causing oxidative stress and liver injury), fatty acid biosynthesis (disruptions leading to lipid accumulation and hepatocyte damage), tryptophan metabolism (toxic intermediates driving oxidative stress and inflammation) and ferroptosis (iron-dependent oxidative stress leading to lipid peroxide accumulation). Additionally, pathway activations highly correlated with predicted DILI risk include nuclear receptor signaling (e.g., PXR/CAR/FXR), one-carbon metabolism, and bile acid regulation -- highlighting transcriptional reprogramming and metabolic stress as key contributors to hepatotoxicity (Fig. 4A). When compounds are ranked by predicted DILI probabilities, a clear gradient of pathway activation emerges, revealing distinct enrichment patterns for these biological processes. This correlation reinforces the direct mechanistic relevance and interpretability of the model's predictions and highlights these pathways as potential drivers of DILI (Fig. 4B). To identify genes most significant for DILI, we determined the frequency at which each gene was differentially up- or downregulated across DILI drugs in our library, using an adjusted p-value threshold of 0.05. This analysis focused on the concentrations at which toxic effects were first predicted, aiming to uncover early upstream regulators potentially driving DILI. Most frequently up-regulated genes were associated with drug metabolism, transport, stress response, and lipid metabolism. Novel genes linked to inflammation, autophagy, and mitochondrial dysfunction were also implicated (Fig. 4C). Frequently down-regulated genes include those critical for liver functions such as drug metabolism, transport, lipid metabolism, amino acid metabolism, mitochondrial function, coagulation and inflammatory responses. Altered expression of these genes may serve as early indicators of liver injury and reflect DILI's multifaceted mechanisms (Fig. 4D). Establishing that the DILI pathways and genes identified by our model are specific to liver toxicity rather than general toxicity is inherently challenging. However, the model's precision is evident in its accurate classification of non-DILI compounds with known toxicities in other systems, such as Valdecoxib (cardiovascular toxicity), Bupropion (neurologic and cardiovascular toxicity), and Warfarin (hematologic toxicity), indicating the model's ability to distinguish liver-specific toxicity from other forms of organ damage. Bruton tyrosine kinase (BTK) inhibitors, despite their promise in oncology and autoimmune diseases, have faced clinical holds due to liver injury. Recent examples include Evobrutinib, BMS-986142, and Orelabrutinib, all of which were withdrawn or put on hold in phase III in 2023 due to DILI cases. We validated ToxPredictor on four clinical failures: Evobrutinib, BMS-986142, Orelabrutinib, and TAK-875 (type 2 diabetes drug), along with two investigational BTK inhibitors (Rilzabrutinib, Remibrutinib) and two FDA-approved JAK inhibitors (Tofacitinib, Upadacitinib) as negative controls. DILI risk probabilities were assessed at four concentrations and margins of safety (MOS) estimated to classify compounds as high (MOS ≤ 2.5), mid-high (MOS ≤ 12.5), medium (MOS ≤ 80), or low risk (MOS > 80). All clinical failures were flagged as high or medium-high risk with low MOS values, particularly TAK-875, Evobrutinib, and BMS-986142, consistent with their phase III withdrawals. The investigational drugs were classified as medium risk (MOS = 14) and low risk (MOS = 101), which have not yet been linked to DILI in clinical studies yet, while the DILI-negative JAK inhibitors were classified as low risk. These results align closely with their clinical safety profiles (Fig. 5A). ToxPredictor provides dose-dependent DILI risk curves derived from empirical DILI likelihoods across various hypothetical Cmax values, enabling safe dosing recommendations (see Methods). For instance, Rilzabrutinib is categorized as low risk at doses below 100 mg q.d., which is lower than its efficacious dose of 400 mg. In contrast, Remibrutinib's efficacious dose of 100 mg falls within the recommended low-risk range of <155 mg q.d. These findings highlight ToxPredictor's value in informing safe dosing strategies, making it a valuable tool for de-risking new drug candidates (Fig. 5B). We benchmark our model along two key axes: predictive performance and scalability. Predictive performance, measured by balanced accuracy, reflects the model's ability to distinguish DILI-positive from DILI-negative compounds. Scalability captures both technical throughput and biological breadth -- the capacity to generalize across diverse chemistries and mechanisms, including previously uncharacterized ones (Fig. 6A). Our model outperforms a wide range of pre-clinical DILI models, including mechanistic assays, cytotoxicity markers, physicochemical properties, bioactivation and BSEP approaches. In a head-to-head comparison across matched compound sets, it identified 46 out of 66 DILI cases versus the 27 out of 66 identified by Xu et al. HCI assay (49/66 vs 27/66); it shows superior performance over Garside et al. HCI assay (37/46 vs 29/46 DILIs), Vorrink et al. cytotoxicity assay using CD spheroids (37/43 vs 30/43 DILIs), Sakatis et al. bioactivation endpoint GSH adduct (47/65 vs 25/65) as well as their combined assay integrating covalent binding and dose (47/65 vs 32/65). When compared to Kohonen et al.'s transcriptomics-based cytotoxicity model, our approach showed improved sensitivity (26/36 vs. 16/36 DILIs). These comparisons, all at 100% specificity evaluated on the same compounds, underscore the added value of our systems-level, mechanism-agnostic readout (Fig. 6B; Supplementary Table S3). Structure-based in silico models such as TxGemma, DILIGeNN and DILIPredictor underperform in vitro-based approaches in our benchmark. To assess real-world generalizability, we evaluated them on 314 independent compounds (45 DILI + , 269 DILI - ) primarily annotated via LiverTox (scores A/B as DILI + , E as DILI - ); TxGemma was also tested on an expanded set (143 DILI + , 536 DILI - ). All showed limited specificity: DILIGeNN (84% sensitivity, 28% specificity), DILIPredictor (80% sensitivity, 29% specificity), and TxGemma-27B (57% sensitivity, 37% specificity). These findings are slightly below the balanced accuracy of 0.59 reported by Seal et al. (2024) for DILIPredictor. On a benchmark subset of unseen compounds overlapping with DILImap (n = 97), TxGemma reached 63% sensitivity (39/62) and 57% specificity (20/35), while our model achieved 76% sensitivity (47/62) and 86% specificity (30/35). Similarly, DILIGeNN showed perfect sensitivity (5/5) at moderate specificity (2/3), while our model reached 100% on both (5/5 and 3/3). DILIPredictor reached complete sensitivity (23/23) but at the expense of poor specificity (1/7), while ToxPredictor maintained high sensitivity (20/23) at markedly higher specificity (5/7). Low specificity is a key limitation of structure-based models, which lack biological context and tend to over-call toxicity. This results in false positives for commonly prescribed drugs with no risk of hepatotoxicity, such as biotin (flagged DILI+ by DILIPredictor), vitamin D (flagged DILI+ by DILIGeNN), and pemetrexed (flagged DILI+ by all three models). Moreover, they provide only binary outputs, without dose or mechanistic insight. In contrast, transcriptomics enables dose-resolved predictions, mechanistic interpretability, and safety margin estimation -- critical for evaluating toxic liabilities and guiding follow-up experiments (Supplementary Fig. S10; Supplementary Data S4). 3D liver systems offer important physiological context. High-content imaging in 3D models, such as those by Walker et al. and Ewart et al., achieves similar performance on small, curated panels (Walker: 23/27 vs. 23/27; Ewart: 11/14 vs. 13/14). However, their limited scalability constrains their utility to a broader range of DILI mechanisms. They may perform well on narrow, curated panels, but struggle with unknown mechanisms or mechanisms not captured by the low-dimensional endpoint, as shown in the following comparison. To explore the unique capabilities of 2D transcriptomics vs 3D cytotoxicity assays, we conducted direct compound-level comparisons with larger 3D screening studies: Vorrink et al. and Fäs et al. In Vorrink et al. 3D cytotoxicity uniquely detected 3 compounds (Fialuridine, Methotrexate, Trazodone), all linked to cytotoxic effects that result in acute cell death. Conversely, our model uniquely flagged 10 compounds -- including fluconazole, phenytoin, and zileuton -- associated with immune activation, metabolic stress, or enzyme modulation, which are not readily captured by viability endpoints. A similar pattern emerged in the Fäs et al. study: 3D cytotoxicity exclusively identified 4 compounds (e.g., Haloperidol, Fialuridine) whose toxicities depend on structural or metabolic context. Our model uniquely identified 5 compounds (e.g., Cimetidine, Fluconazole, Ximelagatran) marked by subtle transcriptional responses rather than overt cell death. These comparisons highlight a key limitation of fixed single-endpoint models: while effective in narrow contexts, they struggle with broader chemical and mechanistic diversity. Our transcriptomic approach, by contrast, offers systems-level resolution that generalizes across DILI pathways -- not only detecting known cytotoxic responses but also uncovering less immediate, non-lethal mechanisms often missed by traditional assays (Supplementary Data S4). As a result of its unbiased modeling, our approach shows improved detection of idiosyncratic compounds -- a class of toxicities that often escape detection in targeted or phenotypically narrow assays. These compounds, many of which are associated with extremely rare clinical incidence (<12 case reports), present a significant challenge for preclinical screening. Our model identified 29 out of 65 of those cases (44%) the highest detection rate among all evaluated models, while maintaining a specificity of 88% (Supplementary Fig. S11A). Next, we analyzed how combining toxicogenomics with orthogonal assays further enhances detection. The three most effective combinations include pairing our model with Walker et al.'s 3D-based HCI assay to improve DILI detection from 23/27 to 26/27 cases, pairing with Persson et al.'s 2D-based HCI assay to improve DILI detection from 28/37 to 30/37 cases, and pairing with Sakatis et al. GSH depletion assay to increase detection from 47/65 to 53/65. Such strategic combinations could raise balanced accuracy to as high as 98% (Supplementary Fig. S11B). Based on these insights, we propose a tiered de-risking funnel strategy that begins with straightforward endpoints, such as PK data (e.g., Cmax <25 μM) and cytotoxicity assays, to flag overt hepatotoxicity. For candidates showing no early toxicity signals, toxicogenomics provides the most comprehensive and unbiased assessment of DILI risk -- capturing both known and novel mechanisms. For a select few advanced candidates, with sufficient resources, applying toxicogenomics in advanced 3D liver models may offer the most accurate prediction of in vivo responses. This strategy ensures a resource-efficient and mechanistically broad DILI risk assessment in drug development.

Cellarity's AI model predicts drug-induced liver injury

CellarityNov 13 2025 Cellarity, a clinical-stage biotechnology company developing Cell State-Correcting therapies through integrated multi-omics and AI modeling, today announced the publication of a seminal manuscript in Nature Communications, which describes a novel framework for the prediction and characterization of drug-induced liver injury (DILI), along with open-source posting of the model and validation data. DILI is one of the most significant safety challenges in developing therapeutics today, as hepatic safety events undetected in preclinical testing can occur in patients leading to clinical trial failures and sometimes even market withdrawals. In fact, animal models fail to identify as many as half of investigational drugs linked to DILI. To address this challenge, Cellarity designed an integrated AI model called ToxPredictor, which evaluates toxicogenomics to predict dose-related DILI risks. The core of this framework is a transcriptomics library in primary human hepatocytes called DILImap, which illustrates the transcriptional signature of 300 compounds linked to DILI at multiple concentrations. This DILImap features the largest known toxicogenomics dataset available for DILI modeling, a significant advancement as regulators aim to reduce reliance on animal models in drug testing. The publication in Nature Communications describes the validation of the framework, which demonstrated 88% sensitivity at 100% specificity in blind evaluation, outperforming more than 20 industry-standard preclinical safety models and identifying numerous phase 3 clinical safety failures that had been undetected in animal studies. We see Cellarity's ToxPredictor as a fundamental step forward in predictive toxicology, as our model provides deep insights that enable a more comprehensive understanding of liver toxicity mechanisms. Applying machine learning to toxicogenomics holds great promise for more efficient drug discovery and development, significant cost savings, and, most importantly, improved patient safety." Parul Doshi, Cellarity's Chief Data Officer In addition to predicting safety risks, the platform provides improved clarity on hepatotoxic pathways to enable decisions on compound safety margins. Unlike single-endpoint readouts-even from 3D models-transcriptomics offers a higher resolution lens on the complex molecular pathways and relationships to detect diverse DILI mechanisms that cannot be captured by conventional assays. By leveraging the full transcriptomic landscape, the model is capable of capturing a wide range of DILI-related mechanisms, such as mitochondrial dysfunction, oxidative stress, immune activation, and metabolic changes. In head-to-head comparisons, the model uniquely identified numerous non-cytotoxic risks missed by 3D assays. Open source data release Cellarity has made this model and validation data publicly available, providing a powerful collaboration tool for de-risking drug candidates and setting the stage for a paradigm shift in safety evaluations. The resources are available at https://dilimap.org/review-dUFZulWv8k7bERJ3FQs438. Cellarity Journal reference: Bergen, V., et al. (2025). A large-scale human toxicogenomics resource for drug-induced liver injury prediction. Nature Communications. doi: 10.1038/s41467-025-65690-3. https://www.nature.com/articles/s41467-025-65690-3