ToxCompl completion of the DrugMatrix toxicogenomics database: an integrated resource for hypothesis generation

Abstract

The DrugMatrix database contains systematically generated toxicogenomics data from short-term in vivo studies for over 600 chemicals. However, most potential endpoints are missing due to a lack of experimental measurements. Therefore, we leveraged matrix factorization and machine learning methods to predict the missing values, which includes gene expression across 8 tissues on 2 expression platforms along with paired clinical chemistry, hematology, and histopathology. We propose a method, ToxCompl, that applies systematic hybrid sampling guided by Bayesian optimization in conjunction with low-rank matrix factorization to predict the missing values. In-depth validation of the ToxCompl predicted data from machine learning, biological, and toxicological perspectives shows that the predicted differential gene expression aligns well with what would be anticipated. This includes examining the connectivity pattern of predicted gene expression responses, characterizing molecular pathway-level responses from sets of differentially expressed genes, evaluating known transcriptional biomarkers of tissue toxicity, and characterizing predicted apical endpoints. For example, we identified kidney toxicants using the transcriptional biomarker Havcr1. All measured and predicted DrugMatrix data (i.e., gene expression, clinical chemistry, hematology, and histopathology) are available to the public (https://rstudio.niehs.nih.gov/toxcompl/). Notably, predicted clinical chemistry of subtle effects and histopathological prediction are 2 areas we will continue to improve. The main advantage of the ToxCompl approach is that it drastically extends the toxicogenomic landscape into many data-poor tissues in the absence of acquiring additional experimental data, thereby allowing researchers to formulate mechanistic hypotheses about effects in tissues that have been underrepresented in the literature.

Data is critical for modeling relationships between effects historically used to assess toxicity and those measured by newer approaches, such as toxicogenomics. DrugMatrix represents a significant effort to create an integrated data resource, characterizing both traditional endpoints, e.g., histopathology/clinical pathology, and whole genome transcriptomics in the same animals. With over 600 chemicals studied in a collection of short-term in vivo rat studies, it is currently the largest systematically generated toxicogenomics database. As such, the data has been extensively reused in many studies (e.g. see Ganter et al. 2006), particularly for deriving signatures of toxicity effects and mechanistic processes.

The DrugMatrix database was generated using 2 different rat microarray platforms to analyze gene expression within the samples: the GE Healthcare CodeLink UniSet Rat I (RU1) (Gene Expression Omnibus (GEO) 2007), a first-generation microarray technology no longer available, and the Affymetrix GeneChip Rat Genome 230 2.0 (RG230) (Gene Expression Omnibus (GEO) 2008), a second-generation technology still in use (Svoboda et al. 2019). In DrugMatrix, the tissue/sample coverage is more comprehensive with the CodeLink platform, measuring data for 8 tissues including liver, kidney, heart, skeletal muscle, spleen, bone marrow, intestine, and brain. In contrast, Affymetrix covers only liver, kidney, heart, and skeletal muscle. Most DrugMatrix microarray studies are paired with standard clinical pathology measurements and targeted histopathology (i.e., pathology performed on the same tissues as the transcriptomic measurements). Due to cost constraints, each study conducted to create the DrugMatrix resource was targeted based on prior knowledge of target organ toxicity. Consequently, many studies that make up the DrugMatrix data set are focused on the most common target organs, liver and kidney, with much more limited assessment of less common target tissues (e.g., spleen and brain). Therefore, when assembled in matrix form organized as studies (chemical-dose level-duration) by endpoints (tissue gene expression, tissue histopathology, clinical pathology), the DrugMatrix data is missing about 88% of potential endpoints. These missing data are primarily gene expression endpoints in tissues not typically overtly affected by chemical exposure, which reduces the power of the data set to inform on less commonly studied target organs and understand toxicity relationships across tissues. Here we apply artificial intelligence (AI) and machine learning methods to infer these missing endpoints which will help address the major challenge of developing new toxicity assessment methods despite the limited availability of experimental data across many different tissue types.

In recent years, other efforts have already demonstrated the inherent connectivity and covariant behavior of biological features across different levels of organization in complex biological systems, and have demonstrated the ability of AI-driven methods to identify and utilize these relationships to enhance the biological data landscape. For example, innovative AI methodologies have been employed to extend and create gene expression data, leveraging the power of covariance inference methods. Accordingly, techniques like L1000 and S1500+, in conjunction with the GENIE application, have shown promising results (Subramanian et al. 2017; Mav et al. 2018, 2020). These methods take subgenomic measurements and extrapolate them to whole-genome expression profiles, echoing the principles used in haplotype tagging single-nucleotide polymorphisms (SNPs) to infer complete genotypes in genome-wide association studies (GWAS) (Daly et al. 2001). This approach allows for a comprehensive understanding of genomic expressions by analyzing a fraction of the genome.

Additionally, generative adversarial network (GAN)-based methods developed at the National Center for Toxicological Research (NCTR), particularly ToxGAN and TransOrGAN (Chen et al. 2022; Li et al. 2023), have represented a significant advancement in inferring patterns of gene expression. These methods distinctively integrate various factors such as chemical structure, affected organ, dosage level, and exposure duration in ToxGAN, or age and inter-tissue correlative patterns of expression in TransOrGAN, to predict whole-genome expression. In addition, the NCTR group has also applied GAN technology in a similar manner to what was done with ToxGAN, creating AnimalGAN, which infers patterns of clinical pathological response to chemical challenges (Chen et al. 2023). These AI-driven approaches demonstrate varying degrees of success and have the potential to closely approximate actual gene expression measurements.

In alignment with these conceptual frameworks, we describe a method that capitalizes on these biological relationships to infer the substantial missing portion of the DrugMatrix data. Here, we present the details of DrugMatrix and introduce the factorization algorithm with initial results, focus on techniques to preserve rare signals while still maintaining a low prediction error, validate the predicted matrix through extensive analysis from a biological and toxicological perspective, discuss several potential alternative approaches for predicting the missing endpoints, and summarize the results of using these alternative approaches. Given that our approach is purely computational and does not require further animal testing, its success heralds new research directions and could have a fundamental impact on future toxicogenomics studies.

Materials and methods

Organization of DrugMatrix

When organized into a matrix format, the DrugMatrix data set contains approximately n₁ = 193,000 rows and n₂ = 3,000 columns. The rows include histopathology (treatment group average severity score, 04; 146 m), clinical chemistry (treatment group average, value varies based on endpoint, 19 m), hematology (treatment group average, value varies based on endpoint, 19 m), and gene expression data from Codelink (RU1) and Affymetrix (RG230) gene expression platforms (log10 ratio to a set of unexposed animal controls and recorded separately for each tissue type as “tissue gene”; e.g., liver Cyp3a4, kidney Cyp3a4, etc).

Each column in the data set represents an individual treatment group (chemical-dose-duration) average across 3 animals for each endpoint. The complete data processing approach for the measured DrugMatrix data is provided in a White Paper from 2007 (Bushel et al. 2007). Out of a total of about 580M possible entries, close to 72M or 12.5% are present. Figure 1A shows the distribution of data present for different expression platforms, RU1 and RG230, and organs, liver, kidney, heart, bone marrow, brain, intestine, spleen, and skeletal muscle. Importantly, selection of which tissues would be studied was largely uniform for each chemical at all durations and dosage levels. Ultimately, there is a lot more liver and kidney data than the other tissue types and comparatively much less brain, intestine, and bone marrow data. In fact, the amount of intestine organ data is 1/200th the amount of liver data.

Fig. 1.

Data distributions. (A) Distribution of data from different tissues and the 2 transcriptomic expression platforms. Abbreviations: Second generation transcriptomics platform Affymetrix GeneChip Rat Genome 230 2.0 (RG230), Liver (LI), First-generation microarray GE Healthcare CodeLink UniSet Rat I (RU1), Kidney (KI), Heart (HE), Bone Marrow (BM), Skeletal Muscle (SM), Spleen (SP), Brain (BR), and Intestine (IN). (B) Data distribution within different gene expression categories in DrugMatrix. Zero indicates no change compared to unexposed controls, positive values indicate over-expression compared to controls, and negative values indicate under-expression.

The original data matrix along with all results of the study can be found here (https://cebs.niehs.nih.gov/cebs/paper/15888).

As noted above, the RU1 and RG230 genomic data are recorded as log10 ratio of sample gene expression data to control data, and the values fall within the range of [−4.20, 3.88]. Within this range, depending on the expression level, the data fall in 5 categories: extremely under-expressed (−2), under expressed (−1), normal (0), over expressed (1), and extremely over expressed (2). These were defined as: extremely under-expressed (−5, −1], under-expressed (−1, −0.3], normal (−0.3, 0.3), over-expressed [0.3, 1), and extremely over-expressed [1, 5), respectively, with regard to their log10 values. Figure 1B shows the distribution of data into these 5 different categories. Category 0 dominates the other categories, with about 92% of all available data samples.

Completion of the matrix

We used a matrix completion algorithm that minimizes the Frobenius norm (Nguyen et al. 2019a). Thus, we first checked that the necessary conditions were met for applying matrix completion to generate plausible missing entries, including that the ‘complete’ matrix must be able to take the original collection of data and compress it into a lower-rank matrix with fewer parameters. See the Supplemental Methods “Low rank assumption evaluation” subsection for our assessment showing that this low-rank statistical assumption is met for our analysis.

Next, we adopted the popular Funk-Singular Value Decomposition (Funk-SVD) approach (Piatetsky 2007) for completion, where the matrix of observed entries (M), with n₁ probes (rows) and n₂ treatments (columns), is factored into the product P Qᵀ, with P and Q both mapped into the same r-dimensional latent space. P is of size n₁ × r, and Q is of size n₂ × r. This formulation is common in corroborative filtering applications such as the Netflix challenge that was mentioned earlier (Goel 2020). P and Q are randomly initialized and then learned from the observed entries in the matrix, using the following objective function:

$$\mathit{min}\mathrm{ }{\sum }_{ }^{ }(M_{u,i}-(b{u}_u+b{i}_i+\sum _{f=1}^r{P}_{u,f}\times Q_{i,f}){)}^2$$

with M_u,i being the observed entry at probe u and treatment i, f being the index over the latent dimensions, and bu_u and bi_i being the bias-user-u and bias-item-i for the probes and treatments, respectively. A solution to this equation can be found with stochastic optimization, for which we used 2 popular optimizers in our study: stochastic gradient descent (Robbins and Monro 1951) and adaptive moment estimation (Adam) (Kingma and Ba 2015).

In our initial experiment with training, we randomly partitioned the data into a 90%: 5%: 5% train-validation-test split. During these initial tests, it became clear from the lower mean absolute error (MAE) in every tissue type that matrix completion beats Nearest-Neighbor-Mean (NNM) for regression for all organs, demonstrating the effectiveness of the Funk-SVD completion prediction approach compared to using NNM (P = 4.4E-6, see data for Funk-SVD completion vs NNM results by tissue type within the “Matrix Completion Full Process” subsection of the Supplemental Methods).

Data sampling approach for preserving rare signals

DrugMatrix carries meaning beyond just a low-rank matrix of numbers; its patterns have important biological implications that must be preserved. For example, positive values in the RU1 and RG230 data indicate over-expression while negative values indicate under-expression compared to controls. Figure 1B shows the imbalance in the distribution of data in which most of the available entries in DrugMatrix fall within Category 0 (i.e. no change from controls). Due to the vast majority of training data being in Category 0, matrix completion learns to predict most data entries to be in Category 0 (evidence of this is shown in Supplemental Methods within the “preserving rare signals” subsection). However, in practice, the other Categories −2, −1, 1, and 2 carry information that is more likely to require further investigation. From a toxicology perspective, we want to recover these values as accurately as possible. However, there are very few data samples available in these categories making it challenging to predict the extreme values accurately. Only about 0.03% of available data fall within Categories −2 and 2 each, therefore they require special attention for prediction.

We found that these rare signals are better predicted and preserved by conducting non-uniform sampling for completing DrugMatrix. Both over-sampling and random under-sampling were utilized in a hybrid sampling approach after determining that it is insufficient to use only over-sampling or under-sampling separately (see Supplemental Methods within the “over- and under-sampling” subsection). In hybrid sampling, we under-sampled the majority category while simultaneously over-sampling the minority categories in a manner designed to improve Mean F₁ while minimizing MAE. This is conceptually similar to the Synthetic Minority Over-sampling Technique (SMOTE) (Chawla et al. 2002) in that it rebalances rare categories via over‐ and under‐sampling. However, our approach differs in that the sampling proportions are further fine-tuned using Bayesian optimization to maximize Mean F1 while maintaining low MAE. For more information about Bayesian optimization mathematics, see Snoek et al. (2012) and Gisperg et al. (2025). Note that when under-sampling, a different sample was taken between each epoch during training, while during over-sampling a single over-sampled training data set was used for the entirety of the training period.

We experimented with different sampling distributions, as shown in Supplemental Methods within the “hybrid sampling” subsection. In summary, we applied Bayesian optimization to find distributions for the best performing sampling approach by maximizing the Mean F₁ score, in which F₁ score is the harmonic mean of precision and recall (sensitivity), providing a single metric that balances false positives and false negatives. It is commonly used to evaluate machine learning classification models, especially on imbalanced datasets where accuracy can be misleading. An F₁ value of 1 would be perfect precision and recall and a value of 0 would mean that either precision or recall were 0. In our study, the mean F₁ score is the mean of F₁ across all 5 gene expression level categories. The Supplemental Methods file provides these results within the “Bayesian optimization for finding the optimal distribution” subsection. Ultimately, the best sampling distribution is shown in Fig. 2. It produced by Bayesian optimization had a Mean F₁ of 0.5122 and MAE of 0.0916. The Mean F₁ is better, at about 2.63× the baseline mean, and the MAE is comparable to baseline.

Fig. 2.

The hybrid over- and under-sampling distribution percentage is shown for each category of differential gene expression levels in exposed animals compared to controls, as generated by Bayesian optimization with the maximal Mean F1.

Leveraging category data to predict the continuous DrugMatrix

As the final step of the approach, we were able to then leverage the categorical data to predict continuous values for the DrugMatrix completion. By using the optimized hybrid over- and under-sampling method described above, we were able to over- and under-sample the measured continuous data based on their category in a way that best retains information present in the minority categories. We did this by creating a combined data set that includes both the categorical and continuous values for each data sample. The categorical values were used to sample the distribution when matrix completion was applied to train on the continuous data sample and we used the Bayesian optimization described above as our hybrid sampling strategy for determining our over- and under-sampling settings for generating the predictions. We called this approach ToxCompl (i.e. low-rank factorization for DrugMatrix with sampling frequencies optimized by Bayesian optimization). Figure 3A shows the detailed mechanism of the ToxCompl approach, and Fig. 3B shows a mock data example of how it works. Since the real data contains mostly zeros, this example table uses simulated data with artificially increased numbers of differentially expressed values to illustrate the method within a compact space.

Fig. 3.

(A) ToxCompl components with control and data flow. The original matrix is shown as M. As an example, we show a 12 × 6 matrix, with M_1,2, M_4,7, and M_10,2 containing measured values. All other entries are missing. ToxCompl attempts to recover all these values while preserving the rare signals (colored blue). ToxCompl first disretizes M into M′ with category values. In our example, 2 values are in the 0 category while the remaining one is in category 2. Matrix M′ is an input to the Bayesian optimizer so that it can compute a sampling distribution of M for the “completion” component to learn the 2 low-rank matrices, P_i,: and Q_:,j. The Bayesian optimizer observes the validation performance of the “completion” component relative to the sampling distribution of M, and produces an optimal distribution. In the figure, M_4,7 is sampled twice (in one training epoch) due to the relative rarity of M′_4,7 in M′, while M_1,2 and M_10,2 are only sampled once. The “completion” component computes the sampled M entries by P_i,: ⦿ Q_:,j, here ⦿ denotes dot product (for simplicity and readability, we do not include the biases for each of the rows and columns in the example). (B) Conceptual demonstration of the data organization structure with one column per exposure scenario representing each “chemical–dose–duration” combination, and rows for each marker across all of the different data types (i.e. gene expression, histopathology, hematology, and clinical chemistry). A subset of markers is shown here for gene expression and clinical chemistry (abbreviated to clin_chem); note that marker names shown are illustrative placeholders, not actual marker identifiers. Gene expression category level values are listed within separate rows for the same gene measured within different tissues using a “tissue_gene” format. The transcriptomics color-coding illustrates the over- and under-sampling approach for the gene expression data: light orange highlighting denotes under-sampled entries (the specific locations of which will vary with each model completion run), light green highlighting denotes over-sampled entries, and darker green denotes heavier over-sampling. Gray highlighting denotes predicted values. All data shown here are illustrative placeholders only, with differentially expressed gene category values artificially increased to accommodate space constraints.

The setting parameters for ToxCompl were: latent rank r = 300, number of epochs was 300, learning rate was 0.001, regularization was implicitly in weight decay, optimizer was Adam default, early stopping criteria was if validation did not improve for 10 epochs, batch size was 128, Bayesian optimization searched for the sampling ratio for each of the 5 categories, and the acquisition function used was expected improvement (EI). The number of optimization iterations was 50, and the random seed used is 42. The surrogate model was the default within the Scikit-Optimize R package (specifically, the gp_minimize function was used with default settings: Gaussian Process surrogate, Matérn kernel with ν  =  2.5). The parameters we chose to optimize were the percentages of samples coming from each category. Regarding the acquisition functions, we considered the best from the 3 (Expected Improvement, Probability of Improvement, and Lower Confidence Bound), and in the end we used Expected Improvement (EI) because it gave the best results for mean F1 values.

The Bayesian optimization analysis showed that the best sampling ratios for categories −2, −1, 0, 1, 2 was 4.32%, 10.68%, 67.10%, 12.90%, and 5.00%, respectively.

It is important to note that for the predicted data, each column corresponds to a single unique chemical–dose–duration condition, already averaged across animal replicates, and that column contains the full multi-tissue gene expression plus histopathology, clinical chemistry, etc profile for that condition. All entries for a given condition are assigned together to training, validation, or test; no condition was split across sets. Consequently, there are no instances where a test entry for a condition is predicted using training entries from the same condition, and an “interpolation within the same sample” scenario does not occur.

Additionally, in most cases the same tissues were measured for any one given chemical. This means that nearly all chemical-tissue combinations for the predicted data did not have any direct measured data for that specific chemical-tissue pairing. Therefore, the predictions are informed by both (1) measurements of the same chemical within other tissues and (2) measurements of other chemicals that have data within both the target tissue and in tissue(s) that were measured for the chemical for which data is being predicted, combined with the low-rank structure learned across the matrix.

Overall, our goal in this work is to impute missing entries within the existing DrugMatrix matrix, which is a fixed set of chemicals, doses, and durations in which every chemical has some measured data. The intended application is not “cold-start” prediction for entirely new drugs that have no measurements at all, but rather completion of the partially observed matrix given that related conditions for the same and similar chemicals are available. A held-out-by-drug cross-validation (leave-one-drug-out) corresponds to a cold-start setting (evaluating performance when a drug has no observed conditions in training) and therefore addresses a different problem than the one we target here. For that reason, and to keep the scope aligned with the intended use case, we did not perform held-out-by-drug cross-validation and do not claim performance in that setting.

Alternative approach evaluations

Several other approaches for predicting the missing DrugMatrix entries were also explored, including training a random forest regressor with the measured data in the DrugMatrix. Each of the inputs contain 6 variable features, Organ, Platform, Marker, Drug, Dosage, and Duration. The performance of random forest was inferior to ToxCompl and therefore was not used for the final methods approach. However, the random forest modeling with 100 tree estimators did reveal the approximate relative importance of the 6 different variable features in DrugMatrix.

Additional alternative approaches that were tested are elaborated on in the Supplemental Methods within the subsections on “deep neural networks” (in which the deep neural networks performed better than random forest but worse than Funk-SVD completion for ToxCompl), “completion with side features” (which showed similar performance to ToxCompl), and “graph neural networks (GNNs)” (which poses extreme computing challenges for this task at this time).

Evaluation of predicted data performance

To assess the extent to which predicted data mirrors what would be anticipated from measured data, several biological interpretation techniques were utilized (see Figure 4 for the workflow overview). In this assessment, emphasis was placed on evaluating predicted data without corresponding measured data for that tissue type. This approach ensures the elimination of potential information leakage from the measured data through cross-validation. The reference chemicals and biomarkers were selected a-priori based on data availability within DrugMatrix for chemicals with well‐established mechanisms which enabled us to understand whether the predicted data is reflective of biological changes that we would expect for the reference chemicals and biomarkers being evaluated.

Fig. 4.

Visual overview of the study methods and toxicological evaluation of the predicted data for completing the DrugMatrix database, including performance of biological and toxicological validation steps across difference tissues, different chemicals, and when compared to an externally generated data set.

Connectivity Analysis: THE initial method we employed for biological characterization is connectivity analysis, which is a metric for determining the degree of similarity between lists of differentially expressed genes (Shah et al. 2022). The stronger the biological connectivity the greater the similarity in the up- and down-regulated genes between 2 lists of differentially expressed genes. Connectivity is often used to match patterns produced by reference chemicals with those of unknown mechanism of action in order to infer the mechanistic effects of the unknown agent. In this method, predicted data from 5 reference chemicals underwent correlative analysis with a broad range of signatures from the DrugMatrix database. The goal was to determine if the predicted expression aligns globally with treatments from reference chemicals that have similar mechanisms of action associated with specific patterns of transcriptomic response. To evaluate the accuracy of the predicted gene expression, gene lists derived from treatments with the reference chemicals were selected from the RG230 data that was solely predicted (i.e., there was not any corresponding measured gene expression data for the tissue type being evaluated). This analysis is limited to liver gene expression because there are well defined mechanisms of action in this tissue (Wang et al. 2014). The top 500 differentially expressed genes as determined by absolute value of fold change for each of these treatments was loaded into the Illumina Correlation Engine and connectivity analysis was performed to compare the 3-d Affymetrix DrugMatrix liver data against independent data sets within Illumina, including TG-GATES which is a rat exposure dataset that was generated in Japan (Igarashi et al. 2015; https://doi.org/10.18908/lsdba.nbdc00954-01-000). The statistical metric used to quantify connectivity in the Illumina Correlation Engine is the Running Fischer test (Kupershmidt et al. 2010).

Pathway Analysis of Fenofibrate: In our second evaluation, we investigated if there is biologically coherent pathway activation across tissues evident in the predicted gene expression from the PPARα reference chemical, fenofibrate. Pathway analysis is a means of characterizing molecular pathway-level responses from sets of differentially expressed genes (Nguyen et al. 2019b). Enriched pathways provide higher-order insights into the biological processes altered by a treatment. An example of this is the impact of fibrate drugs on genes within various fatty acid metabolism pathways (Oshida et al. 2015). In rats, due to the somewhat ubiquitous expression of Pparα across most tissues evaluated in DrugMatrix, it is anticipated that most tissues would respond to fibrate exposure, primarily through the up-regulation of fatty acid metabolism (Escher et al. 2001). To evaluate the predictive data for its capacity to identify pathways of known effect, we used the RU1 predicted data (i.e., treatments for which there were no existing measured data) from fenofibrate treatments and checked for up-regulation in pathways related to fatty acid metabolism from Kyoto Encyclopedia of Genes and Genomes Pathways (KEGG, https://www.genome.jp/kegg/pathway.html, Ogata et al. 1999; Kanehisa et al. 2016), Bioplanet (https://tripod.nih.gov/bioplanet/, Huang et al. 2019), Gene Ontology Pathways (GO, https://www.geneontology.org/, The Gene Ontology Consortium 2015), and Reactome Pathways (https://reactome.org/, Milacic et al. 2024). For this analysis, we limited our enrichment to genes that were up-regulated by more than 1.5-fold.

Pathway Analysis of Top 10 Predicted vs Measured Pathways: For the third assessment, we executed paired comparisons of KEGG pathway enrichment between closely matched predicted and measured data across various tissues using the Enrichr tool on genes with expression levels that were at least 1.5 fold up- or down-regulated compared to control sample expression: https://maayanlab.cloud/Enrichr/ (Chen et al. 2013; Kuleshov et al. 2016; Xie et al. 2021), which is a tool feature that has been integrated directly into the ToxCompl website. The objective was to see if the top enriched pathways were consistent between measured and predicted data that were for similar exposure conditions (i.e. slightly different lengths of exposure time or different doses for the same chemical) that are expected to have very similar pathway activation profiles. The brain and intestine, which were not subjected to a measured vs. predicted pathway analysis in this section, were omitted because high-confidence treatment pairings (i.e., high certainty of similar effects) could not be identified.

The chemical treatments evaluated were: (1) Ifosfamide for bone marrow data because it is a genotoxic chemotherapy drug used for cancer treatment. It adversely impacts bone marrow by suppressing blood cell production, which can lead to anemia, decreased white blood cell counts, and reduced platelet counts (Davies et al. 1989; Lee et al. 2011). (2) Doxorubicin for heart tissue because it induces a highly reproducible, dose-dependent model of chronic cardiomyopathy that closely mirrors the biochemical and histopathological progression of anthracycline-induced heart failure observed in clinical settings (Mladěnka et al. 2018). (3) Clofibrate for liver tissue because it is a hypolipidemic drug that elicits various effects on the livers of rats. These effects include changes in enzyme expression, hepatocyte morphology, lipid composition, and increased peroxisome activity (Lau et al. 2010). (4) Cytarabine for spleen tissue because it is a chemotherapeutic agent that is known to have adverse effects within the spleen (Watanabe et al. 1993; Kathare et al. 2025). (5) Enalapril for kidney tissue because it is an angiotensin-converting enzyme (ACE) inhibitor that is used to control blood pressure and proteinuria in the hypertensive patients. High doses can reduce renal blood flow, which leads to kidney damage (Navis et al. 1996). (6) Lovastatin for skeletal muscle because it is an HMG-CoA reductase inhibitor used to treat hypercholesterolemia (Hoeg and Brewer 1987). At high doses in rats, it causes kidney damage due to muscle cell lysis (rhabdomyolysis).

Transcriptional Biomarkers of Tissue Toxicity: In the fourth evaluation, we analyzed the plausibility and biological validity of toxicity-related transcriptional gene biomarker induction observed in the predicted data across several tissues. An inference approach was used by omitting training data from the model and subsequently predicting that data. The validity of the findings was assessed based on existing literature or by reviewing histopathological changes observed in DrugMatrix. All findings from this analysis including references to histopathological change can be found in the Shiny application described in the next section about the web application for ToxCompl that hosts the data from this study.

Predicted Apical Endpoint Characterization: Lastly, we conducted a review of the predicted apical data, performing a plausibility evaluation of predicted findings in comparison to closely related treatments that had measured clinical chemistry and histopathology results.

Web application for ToxCompl access and analysis

The full predicted DrugMatrix enables further analysis such as clustering drug treatments by their profiles. We performed this clustering and constructed a web application, ToxCompl, using the R statistical software and Posit’s Shiny framework for R to provide interactive visualizations and enrichment analysis capabilities for these clusters (Wickham et al. 2019; Allaire 2020; Sali and Attali 2020; Attali 2021; Almende et al. 2022; Couture-Beil 2022; Dowle and Srinivasan 2022; R Special Interest Group on Databases (R-SIG-DB) et al. 2022; Thieurmel and Perrier 2022; Wickham et al. 2022; Atkins et al. 2023; Chang et al. 2023; Jawaid 2023; Kipp and Percival 2023; Perrier et al. 2023; R Core Team 2023; Xie et al. 2023). We first created separate clusters for each valid combination of tissue and microarray. This does not include every possible combination of tissue and microarray because not all tissues were tested with both the RG230 and RU1 microarrays. For each combination, we constructed a matrix consisting of the log₁₀ ratio of treatment gene expression level vs. control gene expression level for each gene for each treatment. A treatment is defined by each combination of chemical, dose, duration, and predicted/experimental status. We then scaled the data in each matrix by calculating the direction-adjusted rank for each gene across each row by first assigning the top 100 genes with the largest absolute value of log₁₀ ratios in a row a score from 100 (largest) to 1 (smallest) and every other gene 0. We then multiplied each gene’s ranking by the directionality of its original log₁₀ ratio (positive or negative) to get the direction adjusted rank. If 2 or more genes had the same log₁₀ ratio, they were given the same ranking to have a maximum of 100 ranked genes; that is, if 2 genes are assigned rank 2, the next highest-ranked gene is assigned rank 4.

After scaling the data, we performed dimension reduction using the Uniform Manifold Approximation and Projection (UMAP) package for Python to reduce the matrix to 2 coordinates per treatment for easy 2-dimensional visualization (McInnes et al. 2018).

Results

Feature importance

In this complementary evaluation, we trained a random forest regressor with the entries in the DrugMatrix. Our experiments achieved training MAEs between 0.08 and 0.11, and test MAEs around 0.16 when we vary the number of tree estimators and the max depth of the trees. As stated in the methods, this shows that the performance of the random forest approach was inferior to ToxCompl, and therefore it was not used for the final methods. However, the random forest modeling did reveal the approximate relative importance of the 6 different features in DrugMatrix for learning. With 100 tree estimators, the feature importance of different key variables shows that Drug (0.32), Organ (0.26), and Dosage (0.16) are much more important (i.e. contribute a greater influence to predictions) than Duration (0.10), Platform, (0.09) and Marker (0.07), although none of them are negligible.

Connectivity analysis

The first predicted treatment evaluated for connectivity was beta-naphthoflavone at a dose of 80 mg/kg for 28 d. Beta-naphthoflavone is an AHR agonist that strongly up-regulates Cyp1a1, Cyp1a2, and Cyp1b1 genes (Okey et al. 1994). This prediction of the treatment indicates that these genes would be up-regulated following this treatment (see ToxCompl application). Table 1 below presents the top treatments with the greatest connectivity. Among the top 5 most connected treatments, only leflunomide is also an AHR activator.

Table 1.

Connectivity analysis used the predicted top 500 most differentially expressed genes (DEGs) to show the connectivity between the predicted and measured data for these 5 chemicals when compared to an externally-produced data set, TG-GATES.

Reference Chemical for which Gene Expression Data was Predicted (subset to the top 500 DEGs)	Reference Chemical Gene Targets in Liver	TG-GATES Top 10 Most Related Biosets	Common Genes (#)	Common Genes (%)	P-value	Do Main Gene Targets Match, Y/N
Beta-naphthoflavone 80 mg/kg, 28 d	AHR Agonist, increasing transcription of Cyp1a1, Cyp1a2, and Cyp1b1	Liver of rats + NIMESULIDE at 162 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	121	24%	7.4E-39	No
		Liver of rats + LEFLUNOMIDE at 60 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	94	19%	1.1E-34	Yes
		Liver of rats + CARVEDILOL at 2,000 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	121	24%	3E-33	No
		Liver of rats + LOMEFLOXACIN at 2,000 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	101	20%	2E-32	No
		Liver of rats + AMIODARONE at 147 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	95	19%	1E-31	No
Primidone 750 mg/kg, 5 d	CAR/PXR Agonist, increasing transcription of Cyp2b1, Cyp2b2, and Cyp3a9	Liver of rats + ANASTROZOLE at 400 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	175	35%	7.5E-34	Yes
		Liver of rats + CLONAZEPAM at 2,500 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	115	23%	1.5E-25	Yes
		Liver of rats + CARBAMAZEPINE at 490 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	109	22%	2.1E-24	Yes
		Liver of rats + METHAPYRILENE at 100 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	149	30%	2.8E-23	Yes
		Liver of rats + AMINOGLUTETHIMIDE at 350 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	116	23%	3.3E-23	Yes
Clofibrate 500 mg/kg, 3 d	PPARα Agonist, increasing transcription of Acot1, Cyp4a1, and Vnn1	Liver of rats + NAFENOPIN at 338 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	198	40%	1.6E-77	Yes
		Liver of rats + BEZAFIBRATE at 617 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	172	34%	6.6E-75	Yes
		Liver of rats + FENOFIBRATE at 215 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	182	36%	7.1E-71	Yes
		Liver of rats + PIRINIXIC ACID at 364 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	187	37%	2.1E-67	Yes
		Liver of rats + CLOFIBRIC ACID at 448 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	154	31%	1.4E-64	Yes
Beta-estradiol 3-benzoate 25 mg/kg, 5 d	ESR1/ESR2Agonist, increasing transcription of Rbp7, Cited4, and Lifr	Liver of rats + DIETHYLSTILBESTROL at 280 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	163	33%	9E-37	Yes
		Liver of rats + NORETHINDRONE ACETATE at 125 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	140	28%	6.8E-36	Yes
		Liver of rats + DEXAMETHASONE at 1 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	156	31%	4.7E-33	No
		Liver of rats + DIETHYLSTILBESTROL at 2.8 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	132	26%	8.6E-33	Yes
		Liver of rats + ETHINYLESTRADIOL at 10 mg-kg-d in corn oil by oral gavage 3 d _vs_ vehicle	134	27%	1.2E-32	Yes
N-nitrosodiethylamine 34 mg/kg, 5 d	Genotoxic via P53 Signaling, increasing transcription of Ccng1, Phlda3, Aen, Mdm2, and Cdkn1a	Liver of rats + N-NITROSODIETHYLAMINE at 34 mg-kg-d in saline by oral gavage 3 d _vs_ vehicle	100	20%	2.6E-31	Yes
		Liver of rats + STAVUDINE at 1,400 mg-kg-d in water by oral gavage 3 d _vs_ vehicle	142	28%	5.4E-29	Yes
		Liver of rats + AFLATOXIN B1 at .3 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	84	17%	6.8E-23	Yes
		Liver of rats + 2-ACETYLAMINOFLUORENE at 30 mg-kg-d in CMC by oral gavage 3 d _vs_ vehicle	100	20%	3.6E-21	Yes
		Liver of rats + MITOMYCIN C at 1.7 mg-kg-d in saline by intraveneous 3 d _vs_ vehicle	67	13%	1.9E-19	Yes

Connectivity is strongest when the liver gene target for the reference chemical exposure matched the gene target of the chemical within the bioset that had overlapping genes. When the main gene targets did not match, this suggests that the bioset chemical had a different mechanism of action (other MOA) than the DrugMatrix predicted data. All correlations were in the positive direction.

A review of the connecting genes showed that the connectivity of the 4 non-AHR activating treatments was primarily through down-regulated genes. This suggests 2 possibilities: 1) beta-napthoflavone may have a secondary mechanism, separate from AHR activation, influencing gene expression in both beta-naphthoflavone and the 4 connective non-AHR treatments. 2) The predicted expression for beta-naphthoflavone might be inaccurate, at least for the down-regulated genes. This implies that there may be only partial accuracy in the predicted expression for beta-naphthoflavone.

The next predicted gene list that was evaluated for connectivity was from primidone. Primidone is metabolized in vivo to phenobarbital (Butler and Waddell 1956), which activates the nuclear receptors CAR and PXR. This activation leads to an increased expression of Cyp2b1, Cyp2b2, and Cyp3a9 genes in the liver (Men and Wang 2023). The predicted gene expression response from a treatment of Primidone at a dose of 750 mg/kg for 5 d indicated an increased expression of all 3 of these cytochrome P450 genes (see ToxCompl application). Connectivity analysis using the top 500 predicted genes by fold change revealed the strongest connectivity with the treatments listed in Table 1. All these compounds are documented activators of CAR/PXR (Karpale et al. 2022) or show strong induction of Cyp2b1, Cyp2b2, and Cyp3a9 genes consistent with activation of these receptors (observed in Illumina Correlation Engine). Therefore, in the case of primidone, the predicted data show very strong connectivity.

Next, clofibrate is a well-known activator of the nuclear receptor PPARα, which is used therapeutically to treat hyperlipidemia (Desvergne and Wahli 1999). PPARα activators lead to an increased expression of many genes in the rat liver associated with fatty acid metabolism (Tamura et al. 2006). Some of the most pronounced transcriptionally responsive genes to PPARα activation include Acot1, Cyp4a1, and Vnn1. As anticipated, Acot1, Cyp4a1, and Vnn1 were predicted to be up-regulated in the clofibrate-500 mg/kg-3-d treatment (see ToxCompl application). The top 5 connected treatments are all well-known PPARα activators (Tamura et al. 2006) (see Table 1).

Beta-estradiol 3-benzoate (the pro-drug for beta-estradiol) was also evaluated, which is a pharmacological estrogen known for producing a distinct gene expression pattern in the male rat liver. Genes commonly up-regulated in the liver by such estrogenic treatments include Rbp7, Cited4, and Lifr (observed in the Illumina Correlation Engine (Kupershmidt et al. 2010)). The predicted expression response from the treatment with beta-estradiol 3-benzoate at a dose of 25 mg/kg for 5 d showed up-regulation of Rbp7, Cited4, and Lifr, suggesting the model can accurately predict biologically appropriate responses to this estrogenic treatment (see Table 1 and the ToxCompl application). An exception was the connectivity observed with dexamethasone, which is structurally related to estrogens but is a glucocorticoid receptor agonist. However, overall the results of the connectivity analysis for beta-estradiol 3-benzoate indicate that the predicted gene expression is consistent with the effects of an estrogen in the liver (Kupershmidt et al. 2010; Wang et al. 2014).

Lastly, N-nitrosodiethylamine (NDEA) is a member of the highly genotoxic and carcinogenic class of compounds commonly referred to as nitrosamines (Li et al. 2021a). NDEA induces an increase in a specific set of liver genes indicative of a genotoxic effect (Eichner et al. 2014). Among these genes are Ccng1, Phlda3, Aen, Mdm2, and Cdkn1a, all integral to the P53 signaling network (Molecular Signatures Database (MSigDB) 2015). The predicted Affymetrix data for this treatment in liver showed an up-regulation of all 5 aforementioned genes (see ToxCompl application), therefore aligning with the expected response in the P53 pathway following a high dose exposure to NDEA. Additionally, the connectivity analysis, presented in Table 1, using the predicted data revealed the highest scoring connectivity with treatments known for their well-documented genotoxic properties (Guimaraes et al. 2010; Kirkland et al. 2016).

Overall, the findings from the connectivity analysis suggest that the predicted expression, broadly speaking, aligned with what would be anticipated. In most instances, this prediction offers a connectivity level comparable to that of observed data.

Pathway analysis—fenofibrate across all tissues

After the above connectivity analysis in liver tissue only, we investigated if there is biologically coherent pathway activation across different tissues by assessing the predicted gene expression for exposure to the PPARα reference chemical, fenofibrate. The results are shown in Table 2. The 4 chosen PPAR-related pathway gene sets performed best in kidney tissue (upregulation of all 4 gene sets), followed by heart (3 gene sets upregulated), liver (2 gene sets), and skeletal muscle (2 gene sets), then brain (1 gene set), and bone marrow (1 gene set). Intestine and spleen lacked up-regulation of any of the 4 PPARα related gene sets that were evaluated.

Table 2.

Adjusted P-values for predicted fenofibrate gene expression across the 8 different tissue types in DrugMatrix for PPARα associated pathway enrichment gene sets from 4 different sources (KEGG, Bioplanet, Reactome, and GO BP).

Source: Gene set	Adjusted p-value for each Tissue (number of measured data points)
	Liver (34.6M)	Kidney (19.1M)	Heart (11.8M)	Bone marrow (2.8M)	Skeletal Muscle (1.6M)	Spleen (1.5M)	Brain (0.55M)	Intestine (0.2M)
Reactome: Metabolism of lipids	1.2E-07*	3.0E-06*	0.03*	0.02*	3.0E-06*	0.24	1.8E-07*	0.07
GO BP: Fatty acid metabolic process	0.02*	2.0E-03*	7.0E-03*	0.45	0.05	0.39	0.52	0.79
KEGG: Fatty acid degradation	0.17	3.0E-04*	0.04*	0.57	0.04*	0.52	0.13	0.75
Bioplanet: Fatty acid metabolism	0.19	4.0E-04*	0.05	0.60	0.07	0.53	0.24	0.76

The predictions were for fenofibrate-430 mg/kg-5 d exposure in all tissue types except for liver and skeletal muscle, for which fenofibrate-430 mg/kg-3 d data predictions were used. The columns are arranged in order of tissues with the most data (millions of data points, M) to the fewest measured data points. Adjusted P-values less than 0.05 are bolded and marked with an asterisk* to indicate statistical significance.

The predicted data performed much better for tissues with more prolific measured data (particularly liver, kidney, and heart) compared to tissues with far more sparse measured data (including spleen and intestine) suggests that, currently, the spleen and intestine predictions may be less reliable compared to the other tissue types. Therefore, those tissues would substantially benefit from adding more measured data to the overall database (using the preserved frozen tissue samples from the original exposure studies), which would enable higher confidence predictions in the future.

Pathway analysis—top 10 pathways measured vs. predicted across different tissues

To assess the validity of the KEGG pathways identified from different chemical treatment predicted data, we evaluated data for 6 chemical treatments that was predicted for 6 different tissue types. The full data about the top 10 enriched KEGG pathways for gene expression from measured vs predicted data are presented in the Supplemental Results, and a summary of the percentage of overlap between the measured and predicted data as well as how the rankings of the KEGG pathways lined up is presented in Table 3. Clofibrate predicted data for liver tissue had the highest overlap (70% of the top 10 KEGG pathways), and cytaribine predicted spleen data had the lowest overlap (30%), followed by lovastatin-skeletal muscle at 40% overlap, and with all of the other chemical-tissue predicted data (ifosphamide-bone marrow, doxorubicin-heart, enalapril-kidney) having 50% to 60% overlap in the top 10 KEGG pathways for measured vs predicted data.

Table 3.

Percentage of overlapping pathways across the top 10 Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways for measured versus predicted data spanning both gene expression platforms, 6 different reference chemicals, and 6 organ types for genes that were at least 1.5 fold up- or down-regulated compared to control sample expression.

Organ Being Evaluated	Gene Expression Platform	Chemical Treatment-Dose-Duration	Top 10 KEGG Pathways
		(Measured Data| Predicted Data)	Measured Ranking (Predicted Ranking)	Percentage that were Replicated
Liver	RG230	Clofibrate-500 mg/kg-7 d | Clofibrate-500 mg/kg-3 d	1(2), 2(9), 3(3), 4(1), 5(4), 6(10), 10(6)	70%
Kidney	RG230	Enalapril-1,500 mg/kg-5 d | Enalapril-1,500 mg/kg-3 d	1(1), 2(5), 4(9), 6(3), 7(4), 9(10)	60%
Bone Marrow	RU1	Ifosphamide-143 mg/kg-3 d | Ifosphamide-143 mg/kg-5 d	1(1), 2(2), 3(4), 4(3), 6(7), 10(9)	60%
Heart	RG230	Doxorubicin-3 mg/kg-5 d | Doxorubicin-4 mg/kg-28 d	1(1), 2(2), 3(4), 7(10), 10(5)	50%
Skeletal Muscle	RU1	Lovastatin-1,500 mg/kg-3 d | Lovastatin-450 mg/kg-3 d	1(7), 2(5), 4(1), 7(9)	40%
Spleen	RU1	Cytaribine-23 mg/kg-3 d | Cytaribine-23 mg/kg-5 d	1(2), 2(9), 3(1)	30%

The top 10 KEGG pathways were assigned number 1-10 to denote their ranking. Then, the rankings for each KEGG pathway that appeared in both the measured and the predicted data results are shown in column 4, in which the measured ranking for the KEGG pathway is shown first and then the predicted ranking of that KEGG pathway is shown next to it in parenthesis. The percentage of the top 10 KEGG pathways that were replicated in both the measured and predicted data is shown in the last column. ToxCompl-predicted differentially expressed gene (DEG) lists yielded average KEGG pathway overlaps of 30%–70%, whereas randomized gene lists (matched to ToxCompl DEG counts) yielded average overlaps of only 0%–24% across 5 runs per chemical. For the full names and lists of the KEGG pathways, please see the Supplemental Results.

Regarding the order of the KEGG pathways that overlapped, the most statistically significant KEGG pathway was the exact same for enalapril (kidney), ifosphamide (bone marrow), and doxorubicin (heart) for measured vs predicted data. Additionally, in order to assess whether the observed pathway overlap reflected biologically meaningful signal rather than chance, we performed a parallel analysis for each chemical using randomized gene lists matched in size to the number of ToxCompl-predicted DEGs with at least 1.5 fold up- or down-regulated compared to control sample expression. Across 5 runs per chemical, randomized gene lists yielded average KEGG pathway overlaps of only 0%–24%, compared to the higher 30%–70% overlap for the true ToxCompl-predicted DEGs, supporting that the predictions capture biologically relevant pathway-level responses.

A further review of the specific pathways identified in the different predicted pathway sets indicated an alignment with known effects of the chemicals in respective tissues. For example, in the case of the predicted clofibrate expression in the liver, some of the top pathways include PPAR signaling, peroxisome, and fatty acid metabolism (Yan et al. 2010). In the heart, the predicted doxorubicin data shows an alteration of complement system proteins and coagulation cascades, which relate to the inflammation and tissue damage response elicited by the chemical (Wen et al. 2021; Grover et al. 2023). The predicted cytarabine data in the spleen suggests a significant alteration of pathways in cancer, which is related in part to its antimetabolite mechanism of action that leads to a DNA damage response, the primary mechanism of action of the chemotherapeutic (Namoju et al. 2014; Lewicky et al. 2022).

Transcriptional biomarkers of tissue toxicity

To determine if the predicted data produced the expected transcriptional biomarker activation, we examined the purely predicted data for up-regulation of these genes across various target tissues.

Bone marrow, RU1 platform

Cdkn1a is a known transcriptional biomarker for bone marrow toxicity that acts through a genotoxic mechanism (Cazzalini et al. 2010; van den Boogaard et al. 2022). A review of the measured Cdkn1a bone marrow gene expression data reveals that treatments that most strongly up-regulate Cdkn1a are toxic to the bone marrow and are typically genotoxic chemicals. Such treatments include doxifluridine, doxorubicin, ifosfamide, cyclophosphamide, and daunorubicin. In the predicted data for the corresponding measured treatments, all showed up-regulation of Cdkn1a (see ToxCompl application). This suggests that the inference approach performs as expected for bone marrow data.

For the data that was predicted without corresponding measured data, the following treatments were anticipated to elicit the greatest up-regulation of Cdkn1a: Cyclopropane carboxylic acid, doxorubicin (with a different duration and dose than the measured data), etoposide, gatifloxicin, and indarubicin. All of these compounds, with the exception of gatifloxicin, are well-documented genotoxic chemicals that cause bone marrow suppression. Overall, these findings suggest a moderate to high reliability for using solely predicted data to identify treatments that cause bone marrow suppression through a genotoxic mechanism when evaluated via up-regulation of Cdkn1a.

Brain, RU1 platform

Spp1 is a recognized transcriptional biomarker that is up-regulated following overt neurotoxicity and traumatic brain injury in rats (Spradling et al. 2011). Reviewing the measured brain data, several well-known neurotoxic chemicals were found to up-regulate Spp1. These include high doses of caffeine (Walker 1989), nicotine (Ferrea and Winterer 2009), haloperidol (Ferrea and Winterer 2009), and thimerosal (Clarkson and Magos 2006). In the predicted data corresponding to these treatments, all showed up-regulation of Spp1, indicating that the inference approach functions as anticipated for brain tissue as well.

For the fully predicted brain data without corresponding measured data within brain tissue, the treatments predicted to cause the greatest up-regulation of Spp1 were progesterone, chlorambucil, meloxicam, hydralazine, and sodium nitroprusside (see ToxCompl application). Among these, only chlorambucil (Iavazzo et al. 2020) and sodium nitroprusside (Iadecola 1997) have documented neurotoxic effects. Thus, these results suggest limited reliability in using solely predicted data to identify neurotoxic treatments.

Heart, RG230 platform

Myh7 is a transcriptional biomarker that is up-regulated in the heart following a toxic challenge, which elicits histopathological changes such as single cell necrosis (Mikaelian et al. 2010). A review of the measured heart Myh7 expression data identified treatments that exhibited the strongest up-regulation as sulindac, doxorubicin, dexamethasone, phenytoin, and cyclosporin A (see ToxCompl web application). Of those 5 chemicals, doxorubicin, dexamethasone, and cyclosporin A are recognized cardiotoxic agents (Mladěnka et al. 2018). Sulindac, in the context of DrugMatrix, likely caused GI perforation in the animals, leading to high levels of systemic inflammation which may have induced a secondary stress on the heart, resulting in up-regulation of Myh7 (Kakihana et al. 2016). Although phenytoin is not widely recognized as a cardiotoxic agent, it can induce cardiotoxicity at high doses through sodium channel inhibition (Cave and Sleigh 2003).

Predicted expression levels for these chemical exposure treatments were consistent when Myh7 expression was excluded from the model and subsequently predicted. When evaluating the expression of Myh7 in treatments without corresponding measured data, the chemicals showing the highest up-regulation of the Myh7 gene were miconazole, erythromycin, 1-napthyl isocyanate, benoxaprofen, and chlorpromazine (see ToxCompl application). Of these, only chlorpromazine is a well-documented cardiotoxicant (Li et al. 2021b). In the DrugMatrix dataset, 1-napthyl isocyanate did result in an increase in relative heart weight, suggesting potential cardiotoxicity (Kim et al. 2016). Overall, these findings suggest that either the predicted data might inaccurately represent cardiotoxicity as related to Myh7 expression, or there are some potentially poorly understood effects of these chemicals on the heart at high doses, which could arise through direct or indirect mechanisms.

Intestine, RU1 platform

When the intestinal lining is damaged, transcription up-regulation of various inflammatory signaling genes occurs (Martínez-Augustin et al. 2008). One of the genes that is a universal indicator of inflammation is Lcn2 (Abella et al. 2015). In DrugMatrix, intestinal only a limited number of treatments were assessed, making prediction of gene expression particularly challenging. In the measured data there were a limited number of chemicals that caused up-regulation of Lcn2. These include bisacodyl, loperamide, and azasetron (see ToxCompl application). At the high doses that were used to create the data in DrugMatrix, these chemicals are all plausibly toxic to the intestines.

Kidney, RU1 and RG230 platforms

One of the most well-established transcriptional biomarkers of kidney toxicity is Havcr1 (also known as kidney injury molecule-1, Kim-1), which is induced in the kidney following acute damage (Song et al. 2019). Havcr1 is present on both the Codelink RU1 and the Affymetrix RG230 platforms, with more solely predicted data available for RG230. A review of the measured data reveals an up-regulation of Havcr1 in the kidney following exposure to several well-established nephrotoxic treatments that cause histopathological effects in the proximal tubule.

Examining the purely predicted treatments for kidney that show the most significant up-regulation, we find treatments like cisplatin (28 d, 0.05 mg/kg), bacitracin (3 d, 380 mg/kg), and allopurinol (1-, 3-, or 5-d, all at 175 mg/kg). Some intriguing predictions also include an up-regulation of Havcr1 by isoprenaline (3 d, 15 mg/kg), thioacetamide (5 d, 200 mg/kg), and sildenafil (5 d, 14.6 mg/kg) (see ToxCompl application). Isoprenaline, at high doses, can reduce the glomerular filtration rate due to an increase in pre- and post-glomerular resistance (de Ponte et al. 2018), thereby damaging the kidneys (Zhou et al. 2022). Thioacetamide, while being a reactive chemical primarily known for its liver effects, can also be toxic to the kidneys at high doses (Barker and Smuckler 1974). Sildenafil may induce hypotension, leading to kidney hypoperfusion and subsequent hypoxia, similar to the effects seen with non-steroidal antiinflammatory drugs (NSAIDS) (Liu et al. 2018). Notably, several other NSAIDs are also predicted to elevate Havcr1 expression.

Liver, RG230 platform

Gpnmb is a transcriptional biomarker that is up-regulated in response to overt cytotoxicity in several tissues, including the liver (Glaab et al. 2021). It is known to modulate inflammation through its effects on macrophages (Saade et al. 2021). The chemicals that elicited the strongest response in the measured data were miconazole, n, n-dimethylformamide, and carbon tetrachloride (see ToxCompl application). All of these are recognized liver toxicants and demonstrated increased liver enzyme levels in DrugMatrix. Notably, when these treatments were excluded from the training data and subsequently predicted, they showed strong up-regulation of Gpnmb, reinforcing the accuracy of the predictive model.

Chemicals displaying the most significant up-regulation of Gpnmb in the purely predicted data (i.e., without any corresponding training data for liver tissue) include doxapram, carbon tetrachloride, thioacetamide, pralidoxime chloride, and atorvastatin (see ToxCompl application). All these treatments revealed various forms of liver histopathology in the DrugMatrix database, with some eliciting liver necrosis. This observation suggests that up-regulation of Gpnmb in the purely predicted data is indicative of overt liver damage.

Spleen, RU1 platform

The spleen and bone marrow often demonstrate correlated toxicological manifestations. This is because toxicity to the bone marrow can induce stress in the spleen, which then needs to remove damaged blood cells originating from the marrow. The gene Cdkn1a, also known as p21, serves as an indicator of genotoxic damage in both the bone marrow (as discussed earlier in the “Bone Marrow” subsection) and, secondarily, the spleen (Wang et al. 2006). The measured treatments that resulted in increased expression of Cdkn1a in the spleen are doxorubicin, mitomycin c, carmustine, leflunomide, and chlorambucil (see ToxCompl application). All these treatments are either direct-acting genotoxic chemotherapeutics or anti-metabolite therapeutics (Swift and Golsteyn 2014), making the findings of Cdkn1a up-regulation consistent with the genotoxic impacts these agents have on spleen cells. For example, leflunomide, a pyrimidine synthesis inhibitor, is used as an immune suppressant (Fox et al. 1999). It causes lymphoid cellular depletion in the spleen by inhibiting DNA synthesis. Such inhibition often triggers a response from the p53 signaling pathway (Huang et al. 2011), which is up-stream of Cdkna1, making the observed up-regulation of Cdkn1a with this treatment expected.

When the data for these treatments were excluded from the model and then predicted, all showed increased expression of Cdkn1a, suggesting the model effectively predicts expression indicative of genotoxic effects in the spleen. In instances where spleen expression data was solely predicted (i.e., no corresponding measured data was available for the spleen), the chemicals projected to induce the most substantial upregulation of Cdkn1a were daunorubicin, cyclopropane carboxylic acid, etoposide, doxorubicin (at a different dose level than the measured data), thioacetamide, and doxifluridine (see ToxCompl application). Daunorubicin, etoposide, doxorubicin, and doxifluridine are all recognized genotoxic chemotherapeutics that induce splenic toxicity secondary to bone marrow toxicity (Wang et al. 2006; Weber 2015; Kathare et al. 2025). Cyclopropane carboxylic acid resulted in a decreased leukocyte count and diminished lymphocytes in DrugMatrix, consistent with a chemical causing bone marrow toxicity. Thioacetamide has also been linked with spleen toxicity, likely secondary to its effects on the bone marrow (Al-Bader et al. 2000).

Skeletal muscle, RU1 and RG230 platform

Increased expression of the Arrdc2 gene has been linked with chemical-induced cellular injury in skeletal muscle, specifically muscle atrophy (Wu et al. 2010). In DrugMatrix, Arrdc2 expression was measured using both the RU1 and RG230 platforms. For this analysis, we focused on the RU1 platform due to its more comprehensive alignment between gene expression and histopathology within DrugMatrix. In the measured data, treatments resulting in the most significant up-regulation of Arrdc2 include simvastatin, pravastatin, fluvastatin, cerivastatin, and lovastatin (see ToxCompl application). All of these chemicals also demonstrated varying degrees of skeletal muscle toxicity, as evidenced by the histopathology recorded in the DrugMatrix database. Furthermore, they are known to cause rhabdomyolysis (breakdown of skeletal muscle tissue) (Mendes et al. 2014). When these treatments were excluded from the training dataset and subsequently predicted, all exhibited the anticipated high level of Arrdc2 expression. This suggests the model is adept at predicting Arrdc2 expression in skeletal muscle.

Treatments that were purely predicted and exhibited the strongest up-regulation of Arrdc2 include 1naphthyl isocyanate, omeprazole, tetracycline, chlorpromazine, and lansoprazole (see ToxCompl application). Out of these 5 chemicals predicted to significantly induce Arrdc2, 3 can cause rhabdomyolysis. These are the proton pump inhibitors (omeprazole and lansoprazole) (Tanaka et al. 2014) and chlorpromazine (Torres et al. 2015). The remaining 2 treatments, 1-naphthyl isocyanate and tetracycline, are not known to cause skeletal muscle damage. As such, the Arrdc2 gene expression predictions for these 2 chemicals might be somewhat unreliable for predicting skeletal myotoxicity.

Predicted apical endpoint characterization

A component of the original sparse matrix of measured data encompassed hematology (19 endpoints), clinical chemistry (19 endpoints), and histopathology data (dose group severity for 146 pathologies across 8 tissues). In the current version of the matrix completion, histopathological predictions tend to be overestimated due to targeted histological assessment (i.e., in DrugMatrix histopathology was not performed unless there was an expectation of effect). This discrepancy is evident in Table 4, where the measured assessment of histopathology from a 1-naphthylisocyanate-60 mg/kg-7 d exposure identified 4 effects with an average severity score greater than 1 out of a maximum possible severity score of 4. In contrast, the prediction histopathology indicated 27 effects with an average severity exceeding 1. However, notably, the 4 histopathological outcomes identified in the measured data had remarkably similar average severity scores when compared to the predicted findings list (see Table 4).

Table 4.

Histopathology with scores greater than 1 comparing measured versus predicted values for exposure to 1-naphthylisocyanate-60 mg/kg-7 d.

Histopathology for 1-naphthylisocyanate-60 mg/kg-7 d	Average Severity
	Measured	Predicted	% Difference
Liver, HEPATOCYTE, CENTRILOBULAR, LIPID ACCUMULATION, MACROVESICULAR	–	3.17	NA
Liver, BILE DUCT HYPERPLASIA*	3.00	2.99	−0.3%
Liver, PERIPORTAL, INFLAMMATORY CELL INFILTRATE, MIXED CELL*	2.67	2.55	−4.4%
Intestine, EPITHELIAL ATROPHY	–	2.33	NA
Kidney, TUBULE, HYALINE DROPLETS	–	2.16	NA
Mesentery, LYMPH NODE PROLIFERATION	–	2.06	NA
Mesentery, LYMPH NODE PLASMACYTOSIS	–	1.97	NA
Liver, PERIPORTAL, FIBROSIS*	2.00	1.96	−2.0%
Liver, HEPATOCYTE, DIFFUSE, CYTOPLASM, EOSINOPHILIA	–	1.86	NA
Liver, HEPATOCYTE, CENTRILOBULAR, NECROSIS, ONCOCYTIC	–	1.73	NA
Kidney, TUBULE, DEPOSIT	–	1.71	NA
Liver, HEPATOCYTE, CENTRILOBULAR, GLYCOGEN ACCUMULATION	–	1.58	NA
Intestine or Mesentery, PERITONITIS	–	1.58	NA
Liver, HEPATOCYTE, CENTRILOBULAR, LIPID ACCUMULATION, MICROVESICULAR	–	1.41	NA
Liver, HEPATOCYTE, NONZONAL, NECROSIS, ONCOCYTIC*	1.33	1.40	5.0%
Skeletal Muscle, RIGHT BICEPS FEMORIS, MYOCYTE, NUCLEAR PROLIFERATION	–	1.28	NA
Spleen, LYMPHOID DEPLETION	–	1.25	NA
Liver, HEPATOCYTE, DIFFUSE, HYPERTROPHY	–	1.23	NA
Liver, HEPATOCYTE, PERIPORTAL, LIPID ACCUMULATION, MACROVESICULAR	–	1.17	NA
Liver, HEPATOCYTE, PERIPORTAL, CYTOPLASM, EOSINOPHILIA	–	1.16	NA
Intestine, INTESTINE AUTOLYSIS	–	1.16	NA
Liver, SINGLE HEPATOCYTE NECROSIS	–	1.12	NA
Liver, HEPATOCYTE, MIDZONAL, LIPID ACCUMULATION, MICROVESICULAR	–	1.10	NA
Heart, INTRAMYOCARDIAL ARTERIES, DEGENERATION	–	1.09	NA
Liver, SUBCAPSULAR, FIBROSIS	–	1.06	NA
Intestine, VILLUS TIP FUSION	–	1.05	NA
Liver, HEPATOCYTE, NONZONAL, LIPID ACCUMULATION, MACROVESICULAR	–	1.04	NA
Heart, INTRAMYOCARDIAL ARTERIES, PERIVASCULAR EDEMA	–	1.04	NA

The histopathology severity scores are higher for more pronounced histology changes, with a maximum possible score of 4. Histopathological outcomes that had both measured and predicted scores were evaluated for the percent difference between the predicted value versus the measured value. NA indicates that the percent difference was not able to be calculated due to no measured data being available for comparison, and bolded numbers indicate that the percent difference was within ±5%.

Since there were 27 histology changes predicted to have an average severity score over one, compared to only 4 histology changes that were reported, it is possible that up to 23 out of the 27 histopathology predictions may be false positives. Unfortunately, this is difficult to assess due to how this type of measured data is reported. For histopathology, only histological changes are reported by evaluators; a lack of observation of a particular change is not noted, so it is not possible to know for certain whether a specific outcome was evaluated for or not. Nevertheless, the apparent overall prediction accuracy for the apical endpoints was suboptimal if the non-reported endpoints are true negatives, indicating that the ToxCompl model may still require significant refinements to enhance its predictive capabilities for histopathology data.

For clinical chemistry, an example comparing measured and predicted clinical chemistry results for 1-naphthylisocyanate-60 mg/kg-7 d (the same exposure as the histopathology example) is provided in Table 5, which shows that the predicted data aligns closely with the measured data (see ToxCompl application). As anticipated, there’s a pronounced up-regulation of liver damage biomarkers (alanine aminotransferase, aspartate aminotransferase, and bilirubin) in both the measured and predicted datasets. Additionally, the predicted data indicates a slight increase in another tissue damage biomarker, lactate dehydrogenase. Both datasets also reflect an increase in cholesterol, which is expected for a compound like 1-naphthylisocyanate that inflicts damage to the biliary tract. This means that the biological observations align with clinical chemistry expectations for this exposure.

Table 5.

Clinical chemistry comparing measured versus predicted data for exposure to 1-naphthylisocyanate-60 mg/kg-7 d.

Clinical Chemistry Endpoint	Lower bound of normal	Upper bound of normal	Measured value	Predicted value	Measured Category	Predicted Category	Accurately predicted in- vs out-of-bounds?
ALANINE AMINOTRANSFERASE (ALT)	29	75	180*	211*	High	High	Yes
ALBUMIN (Alb)	3.6	4.5	3.8	3.8	Normal	Normal	Yes
ALKALINE PHOSPHATASE (ALP)	212	526	516	589*	Normal	High	No
ASPARTATE AMINOTRANSFERASE (AST)	61	119	418*	395*	High	High	Yes
BLOOD UREA NITROGEN (BUN)	11	19	12	15	Normal	Normal	Yes
CARBON DIOXIDE (CO2)	23	43	27	27	Normal	Normal	Yes
CHLORIDE (Cl)	97	107	94*	95*	Low	Low	Yes
CHOLESTEROL (Chol)	51	94	385*	340*	High	High	Yes
CREATINE PHOSPHOKINASE (CPK)	125	966	340	393	Normal	Normal	Yes
CREATININE (Creat)	0.6	0.11	0.10	0.21*	Normal	High	No
GLUCOSE (Clu)	127	209	96*	95*	Low	Low	Yes
LACTATE DEHYDROGENASE (LDH)	65	615	287	132	Normal	Normal	Yes
LIPASE (Lip)	8	28	12	14	Normal	Normal	Yes
PHOSPHORUS (P)	9.2	13.6	8.75*	9.15*	Low	Low	Yes
POTASSIUM (K)	5.09	7.40	6.62	6.65	Normal	Normal	Yes
SODIUM (Na)	139.8	151.0	137.9*	134.4*	Low	Low	Yes
TOTAL BILIRUBIN (T-Bil)	0.1	0.6	8.2*	6.7*	High	High	Yes
TOTAL PROTEIN (Tprot)	5.5	6.8	6.5	6.5	Normal	Normal	Yes
URIC ACID (UA)	0.4	2.4	0.9	1.0	Normal	Normal	Yes

An asterix* indicates that the measured and/or predicted clinical chemistry value was outside of normal bounds, and the last column indicates whether the predicted value and measured value agreed about whether the clinical chemistry endpoint was normal, high, or low.

Furthermore, when evaluating whether the clinical chemistry endpoint was within or outside of the normal bound values, the predictions matched the measured findings for 17 out of the 19 endpoints (Table 5). This was consistent for both values that were higher than normal (ALT, AST, total bilirubin), those that were normal (10 total endpoints), and those that were lower than normal (chloride, cholesterol, glucose, sodium). For the 2 clinical chemistry endpoints that did not agree on whether the values were within vs outside of bounds (ALP and creatinine), the measured values were on the high end of normal, and the predicted values were estimated to be higher than normal bounds. Thus, the directional trend of the values both trending towards high for those 2 endpoints is encouraging for how well the model is performing. There were no serious errors such as an endpoint being measured as high but predicted to be low, or vice versa. Thus, the clinical chemistry predictions were very robust with high confidence for this exposure scenario.

Web application for ToxCompl access and analysis

The full predicted DrugMatrix enables further analysis such as clustering drug treatments by their profiles, so we performed this clustering and constructed a web application for accessing DrugMatrix original data and ToxCompl generated predictions. An instance of the application is available at https://rstudio.niehs. nih.gov/toxcompl/ and the public code repository is accessible at https://doi.org/10.5281/zenodo.19455132.

The application allows a user to explore the gene expression, histopathology, clinical chemistry, and hematology data for each treatment, if available. Users may search for data for specific treatments by specifying the chemical, dose, duration, tissue, microarray platform, and/or gene name. Users may also opt to see a treatment’s measured data, predicted data, or both. The application also enables users to map the rat genes from each treatment to corresponding human genes using the RG230 and RU1 microarrays and enrich those genes using the Enrichr gene enrichment analysis tool at https://maayanlab.cloud/Enrichr/ (Chen et al. 2013; Kuleshov et al. 2016; Xie et al. 2021). Specifically, ToxCompl leverages the Enrichr tool to enrich the genes using human-relevant toxicological annotations from the KEGG (2021), BioPlanet (2019), Reactome (2022), and Gene Ontology (2023) databases (Ashburner et al. 2000; Kanehisa and Goto 2000; Huang et al. 2019; Kanehisa 2019; Gillespie et al. 2022; Aleksander et al. 2023; Kanehisa et al. 2023).

Finally, the ToxCompl provides multiple ways to visualize the treatment data and their associated annotations. For example, Fig. 5 shows an interactive cluster graph for the liver, RG230 cluster of treatments. This particular visualization highlights the related chemicals of fenofibrate, clofibric acid, clofibrate, gemfibrozil, and pirinixic acid, which are all PPARα activators (Bocos et al. 1995; Merk et al. 2015; Xie et al. 2025). We can see that treatments using these chemicals clustered closely together relative to the rest of the liver, RG230 data based on similar gene expression responses. Furthermore, the application allows a user to view tabular visualizations for the gene expression (see Table 6). This example shows liver gene expression data for fenofibrate at both 215 and 430 mg/kg for 5 d. The results show notable correspondence between the measured and predicted data.

Fig. 5.

The ToxCompl web application’s view of the dimension-reduced Uniform Manifold Approximation and Projection (UMAP) cluster of the liver, RG230 microarray data. This plot distinguishes between measured data (given as circles) and predicted data (given as triangles) mapped into a 2-dimensional space. Data markers for different chemicals are color-coded.

Table 6.

Top 10 measured versus predicted up-regulated and down-regulated liver gene expression data (value = log10 ratio treated to control comparison) for fenofibrate at high doses for 5 d on the RG230 chip pulled from the Shiny application.

Fenofibrate Liver Gene Expression Mapping Using RG230 Data	Probeset ID	Human gene	Rat gene	Dose (mg/kg) -5 d	Measured Liver Value	Predicted Liver Value	% Difference	Variability
Top 10 Most Up-regulated Genes	P006001	ACOT1	Acot1	215	3.22	3.34	4%	0.22
	P006001	ACOT1	Acot1	430	3.13	3.24	4%	0.14
	P001780	AQP3	Aqp3	215	2.53	2.45	−3%	0.09
	P001780	AQP3	Aqp3	430	2.24	2.43	8%	0.16
	P018341	ACOT1	Acot1	215	2.24	2.32	4%	0.10
	P018341	ACOT2	Acot2	215	2.24	2.32	4%	0.10
	P013770	HDC	Hdc	215	2.21	2.23	1%	0.02
	P018341	ACOT1	Acot1	430	2.07	2.02	−2%	0.09
	P018341	ACOT2	Acot2	430	2.07	2.02	−2%	0.09
	P013770	HDC	Hdc	430	1.98	2.10	6%	0.02
Top 10 Most Down-regulated Genes	P012822	RTN4RL2	Rtn4rl2	215	−1.39	(N/A)	(N/A)	(N/A)
	P005968	CYP2C9	Cyp2c11	215	−1.41	−1.41	0%	0.08
	P032779	UGT2B17	Ugt2b	215	(N/A)	−1.41	(N/A)	0.25
	P026572	STAC3	Stac3	215	(N/A)	−1.42	(N/A)	0.10
	P013602	TSKU	Tsku	215	−1.42	−1.38	−3%	0.04
	P027287	SLC6A6	Slc6a6	430	−1.44	(N/A)	(N/A)	(N/A)
	P005891	CYP2C18	Cyp2c7	430	−1.58	−1.62	3%	0.02
	P001730	APOA4	Apoa4	430	−1.65	−1.65	0%	0.08
	P001730	APOA4	Apoa4	215	−1.67	−1.60	−4%	0.08
	P005891	CYP2C18	Cyp2c7	215	−1.71	−1.67	−2%	0.02
	P006874	DHRS7	Dhrs7l1	215	−1.91	−2.00	5%	0.08
	P006874	DHRS7	Dhrs7l1	430	−2.66	−2.71	2%	0.02

The percent difference column shows the percentage by which the predicted data was greater or less than the measured data, and bold values indicate that the percent difference was within ±5%. Variability was calculated for n = 5 completion runs with randomly initialized P and Q factor matrices, and for an endpoint we computed the largest deviation from the mean. For example, if 5 runs gave predicted values of 3.0, 2.8, 3.1, 3.0, 3.1, then the mean is 3.0 and the variability would be 0.2.

Discussion

Using complex predictive AI modeling, we have created a new method called ToxCompl that uses the 72 million measured data points from DrugMatrix rat exposure studies to generate predictions for the remaining 508 million missing entries within DrugMatrix. This includes filled-in predicted data for gene expression values, hematology, clinical chemistry, and histopathology for chemical exposures in tissues that were not measured within the original DrugMatrix in vivo experiments, (see ToxCompl website application). As ToxCompl is purely computational and does not require further animal testing for this new predicted data generation, its success points to the value of leveraging large data sets to predict complex biological effects of chemicals using inferential models.

In addition to robust statistical evaluation, we also conducted rigorous analysis of the predicted data from a biological and toxicological perspective, which demonstrate that ToxCompl convincingly recovers much of the missing data in a way that should be useful for hypothesis formulation around biological effects and toxicological mechanisms for a number of tissues, including unstudied tissues such as bone marrow, skeletal muscle, spleen, brain, and intestine.

Our toxicological evaluation findings and their interpretations are extensively described within the Results section. The 5 different types of assessments performed strongly.

1. The connectivity analysis of the top 500 predicted differentially expressed genes showed very strong connectivity between the measured and predicted data for 4 out of the 5 chemicals that were evaluated (Table 1).

2. Pathway analysis of fenofibrate across all tissues showed that the predicted data performed extremely well for tissues with more prolific measured data (particularly liver, kidney, and heart), although tissues with far more sparse measured data (including spleen, brain, and intestine) do need further improvement in future renditions (Table 2).

3. The pathway analysis of the top 10 KEGG pathways for measured vs predicted data across different tissues for key reference chemicals also indicated that tissues with more measured data available have more reliable predictions, with liver tissue performing the best and spleen performing the poorest (Table 3).

   Notably, when a parallel analysis was conducted using a randomized list of genes for the same number of DEGs that the ToxCompl method had predicted for those chemical exposures, the average KEGG pathway overlap across 5 runs for each of the 6 chemicals ranged from only 0% to 24%, meaning that all of the true predictions outperformed the randomized test list (in which the real predicted data had a KEGG pathway overlap range of 30% to 70%).

4. The analysis of the measured vs predicted transcriptional biomarkers data demonstrated that the inference approach functions as anticipated when training data are omitted from the model and biomarker expression is subsequently predicted.

5. Predicted Apical Endpoint Characterization also performed very well for histopathology data for 1-napthylisocyanate when there was both predicted and measured severity data available to compare; the predicted average severity scores matched the measured average severity scores within +/− 5% (Table 4).

An important caveat to the histopathology predictions is that there were 27 histology changes predicted to have an average severity score over one, compared to only 4 histology changes that were reported. Thus, it is important to note that up to 23 out of the 27 histopathology predictions may be false positives, indicating that the ToxCompl model may still require significant refinements to enhance its predictive capabilities for histopathology specifically. At this time, researchers could subset the predicted average severity histology scores to the more extreme cases, such as only using predicted endpoints with a severity above 2.50 (instead of a cutoff of above 1), for which there was far greater agreement. In that case, 2 of the 3 predicted findings were confirmed with measured data.

For the clinical chemistry values of 1-napthylisocyanate, this analysis performed very well; when evaluating whether the clinical chemistry endpoint was within or outside of the normal bound values, the high vs normal vs low predictions matched the measured findings for 17 out of the 19 endpoints (Table 5).

Strengths

One of the strengths of this work includes the variety of matrix completion methods that were evaluated to determine which approach works the best. Other approaches explored included random forest, deep neural networks, and matrix completion with side features, and our experiments showed these do not perform as well as the matrix completion approach that we used (see the results subsection “Feature Importance” for the random forest results and see the Supplemental Methods subsection on “deep neural networks”).

Another strength of this approach is that the database structure enables complex relationships and gene expression patterns to be predicted by the AI model. For example, the transcriptomics data is grouped by organ type and gene within the rows (ex. liver Cyp3a4, kidney Cyp3a4, etc). This means that a gene with variable expression in one tissue type will have predictions reflecting that variability for different exposures, but if that same gene never shows expression in a particular tissue type for any of the chemical exposure conditions, then that gene would be predicted to have no differential expression within that tissue for any chemical for which the data is being predicted. Furthermore, a strength of the hybrid sampling approach is that it is able to provide far more reliable predictions for differentially expressed genes, which are very rare compared to genes with no significant change in expression compared to unexposed controls.

Limitations

Despite the promise of the approach, we do caution the need to be judicious in the application of the data when attempting to draw conclusions related to toxicity, in particular histopathology predictions which currently appear to over-predict corresponding measured findings (making that category of predictions currently lower confidence compared to other predicted data). Accurately quantifying the uncertainties and errors in the predicted matrix is essential to practitioners utilizing the predicted DrugMatrix to formulate biologically sound research hypotheses. Future work will include more refinement of the uncertainty metrics, which will help further clarify the limits of applicability of the inferred predicted data.

Additionally, it is very important to be mindful that tissues with only sparse measured data will not perform as well when making predictions compared to tissues with more measured data, which will have higher confidence predictions of missing data for particular chemical exposure scenarios within that tissue type. The predictions will also likely be best for chemicals with similar mechanisms to chemicals that have more complete measured data profiles, and weaker for chemicals with mechanisms of action that are not well-covered within the DrugMatrix measured data. And while we have performed many different types of toxicological evaluations on these predictions, we were limited by how many assessments we could reasonably perform. The extent of the number of chemicals and tissues within the matrix made it so we were unable to perform a complete evaluation but instead had to rely on highlighting a set of well-annotated reference chemicals within the matrix that were chosen a priori.

Due to the nature and structure of the measured data within DrugMatrix, our application of matrix completion theory rests on assumptions that are not fully satisfied by the DrugMatrix dataset. First, the theoretical guarantees underlying low-rank matrix completion (e.g., by Candes and Tao 2010; Recht 2011) generally presume that missing entries are missing at random (MAR). In DrugMatrix, however, data are missing by design: each constituent study was targeted based on prior knowledge of target organ toxicity, resulting in a Missing Not At Random (MNAR) pattern in which data-rich tissues (liver, kidney) and data-poor tissues (e.g., intestine, brain, bone marrow) reflect systematically different sampling processes rather than a uniform random subset of the true matrix. Second, and relatedly, the theoretical bounds we cite also require that the matrix satisfy incoherence conditions that informative signal be spread relatively evenly across rows and columns rather than concentrated in a subset of entries. We did not estimate the incoherence parameters (μ₀, μ₁) from the observed data, and the pronounced imbalance in tissue coverage (e.g., ∼39.4M measured entries for liver versus ∼0.17M for intestine, Fig. 1), combined with the targeted rather than random sampling design, raises the possibility that the effective coherence of DrugMatrix is higher than the theory assumes. We therefore present the cited recovery bounds as heuristic motivation for applying low-rank completion to DrugMatrix, rather than as formal guarantees specific to this dataset. A practical consequence of both issues is that predictions for data-sparse, MNAR-affected tissues may be biased, for instance, toward response patterns learned predominantly from data-rich tissues such as liver, rather than carrying uniformly distributed uncertainty across the matrix. We did not perform a dedicated sensitivity analysis to bound this bias, and we flag this as an important direction for future work.

Conclusions

In conclusion, our ToxCompl approach and completion of the DrugMatrix by predicting 508 million data points of the 580 million data point matrix has a number of potential uses including 1) improving understanding of inter-tissue response relationships to drug treatment (which is currently not possible with standard single-cell type in vitro experiments) 2) identification of pharmacological signatures across different tissues 3) building machine learning models to predict the gene responses from a drug treatment and 4) inferring chemical structures that will elicit a specific gene response profile. We have also provided a publicly available interactive website that makes this data readily accessible, and it includes search parameters for chemical name, exposure dose, duration of exposure, tissue type, technological platform for gene expression measurement (RG230 vs RU1), settings for displaying only measured data, only predicted data, or both. It will display tabs for the different data types (gene expression, histopathology, clinical chemistry, hematology) and it has the EnrichR tool built into the interface as well.

As was illustrated in the biological and toxicological validation, the predicted data show great promise for aiding in new hypothesis generation including for traditionally under-studied tissue types. However, since 88% of the completed DrugMatrix is predicted, we caution that real data should be generated to test hypotheses before making any final conclusions. Ultimately, we hope that these predicted rat data, which have been generated without the use of additional animal testing, will help drive new hypothesis-driven toxicological research and enable more strategic and efficient use of future research efforts and resources.

Author contributions

L.J.W. and G.C. both wrote the manuscript and created the tables and figures, G.C. created and ran the computational models, F.C. and D.L.S. organized the DrugMatrix data that was used in the modeling, P.C. built the web application, J.N.E. built the database to support the web application, R.M.P., W.M.C., and C.P.S. provided scientific input, S.S.A. conceptualized the work and oversaw the project. All authors reviewed and approved the final manuscript for publication.

Supplementary material

Supplementary material is available at Toxicological Sciences online.

Funding

This research utilized the resources of the Compute and Data Environment for Science (CADES) at Oak Ridge National Laboratory (ORNL), supported by the U.S. Department of Energy (DOE), Office of Science, under Contract No. DE-AC05-00OR22725. Funding for this work was also supported by the National Institutes of Health (NIH) and the National Institute of Environmental Health Sciences (NIEHS) through Intramural Research Project ZIAES103385 and Interagency Agreement No. AES24003001-24-001 between NIEHS and ORNL. Sciome was funded in part by the Department of Health and Human Services (HHS) Contract No. 75N96024C00003.

Conflicts of interest. None declared.

Data availability

The data is available publicly online at https://cebs.niehs.nih.gov/cebs/paper/15888 and within our web interface at https://rstudio.niehs.nih.gov/toxcompl/.