As AI tools gain traction in drug development, there is growing enthusiasm around their potential to predict drug-induced liver injury (DILI). However, the reality is more nuanced. While AI-based models bring value in recognizing patterns across large datasets, they often fall short when it comes to mechanistic understanding and physiological realism. That is where quantitative systems toxicology (QST) modeling, particularly with DILIsym®, delivers a substantial advantage.
DILIsym is the flagship QST platform developed by Simulations Plus and designed to simulate hepatotoxicity using a physiologically-based and mechanistic framework. Although the future of QST modeling will leverage both traditional QST approaches and AI techniques, which will surely outperform exclusively AI-driven models, we explore six core reasons why a traditional QST model—specifically DILIsym—provides more comprehensive and clinically actionable predictions of DILI risk than current AI-only models.
-
Primary Methodology: Mechanistic Simulation vs. Pattern Recognition
DILIsym’s QST modeling begins with physiologically-based pharmacokinetic (PBPK) simulations to estimate drug and metabolite concentrations, most importantly in the hepatic sinusoidal space and hepatocytes. A mechanistic sub-model of the hepatocyte life cycle and associated biomarkers are integrated with mechanistic sub-models of DILI pathways—including oxidative stress, mitochondrial dysfunction, and bile acid and phospholipid transporter inhibition. Simulated populations, or SimPops®, representing interindividual variability in exposure and DILI susceptibility factors are utilized to explore the risk for hepatotoxicity on a population level by combining drug exposure with the mechanistic sub-models included in DILIsym.
In contrast, AI models rely on correlating in vitro assay data with known DILI outcomes based on drug labeling. These correlations often lack a mechanistic foundation and cannot simulate actual liver biology, account for individual patient variability, predict dose-dependent toxicity, nor simulate dynamic treatment scenarios. As a result, AI-based approaches may suffer from limited predictive power when evaluating novel compounds.
- DILIsym simulates real-world physiology; AI models recognize patterns without mechanistic biological context.
-
Inclusion of Both DILI-Positive and DILI-Negative Compounds
DILIsym models have been validated across a wide array of compounds—those that cause DILI and those that do not. This balanced approach strengthens its predictive accuracy and helps distinguish true positives from false alarms.
AI models typically derive insights from datasets biased toward DILI-positive compounds because they are relying on drug labeling data. The absence of negative controls diminishes their predictive specificity and makes them prone to overpredicting risk.
- DILIsym’s balanced training and validation datasets allow for both sensitivity and specificity in predictions.
-
Mechanistic Insights: DILIsym Identifies Synergistic Contributions
The hepatotoxic effects predicted by DILIsym often arise from multiple interacting mechanisms. For instance, a compound might weakly inhibit bile acid transport but severely impair mitochondrial function—effects that are synergistic in vivo. Furthermore and along similar lines, DILIsym can predict hepatotoxic risk for combination therapies where individual compounds may weakly impact hepatotoxicity mechanisms with the potential for the combination of compounds to result in liver injury.
AI models cannot detect such multi-mechanism interactions nor predict the DILI risk for combination therapies; their predictions reflect single-parameter associations, and the underlying causality is often unknown or inaccessible.
- DILIsym can dissect and simulate complex mechanistic interactions—AI cannot.
-
Physiological Relevance: Liver-to-Plasma Concentration Ratios Matter
Because DILIsym explicitly captures drug and metabolite exposure in hepatocytes, it can predict and help identify dosing strategies that balance efficacy with safety. This is particularly important for drugs that concentrate in the liver to toxic levels despite appearing safe in plasma-based assays.
AI models typically do not factor in intracellular drug or metabolism concentrations, and therefore they cannot capture key exposure-risk relationships critical to accurately predict hepatotoxicity.
- DILIsym accounts for intracellular hepatocyte concentrations crucial for driving liver toxicity—AI models do not.
-
Identifying Susceptible Subpopulations
Not all patients are equally at risk for DILI. DILIsym includes SimPops that capture variability in liver function, enzyme levels, transporter expression, and disease status. The simulated populations allow researchers to identify potentially susceptible individuals and optimize clinical trial design.
AI models offer no insights into interindividual variability and cannot stratify patients by risk, limiting their usefulness for precision medicine.
- DILIsym enables exploration of hepatotoxicity associated with patients who have varied levels of compromised liver function or status—AI offers only population-level generalizations.
-
Transparency and Interpretability
A critical advantage of DILIsym is its transparent, mechanistic basis. Users can see which pathways contribute to toxicity, at what dose levels, and under what conditions. This enables rational decision-making and supports regulatory submission.
AI models are often black boxes to users, delivering binary outputs (DILI risk: yes/no) without an understandable rationale. For stakeholders seeking mechanistic validation, this lack of interpretability is a serious drawback.
- DILIsym empowers users with mechanistic clarity—AI leaves them guessing.
Summary Table: DILIsym vs. AI Models for DILI Prediction
| DILIsym (QST Model) | AI Models | |
| 1. Primary Methodology | PBPK-based, mechanistic modeling in SimPops | Pattern recognition based on in vitro data and DILI-labeled drugs |
| 2. Compound Scope | Includes DILI-positive and DILI-negative compounds | Typically limited to DILI-positive compounds |
| 3. Mechanistic Detail | Captures synergistic effects of multiple toxicity pathways and potential to predict toxicity with combination therapies | Cannot assess interactions between mechanisms or between multiple compounds |
| 4. Liver Exposure Prediction | Models hepatocyte concentrations and liver-to-plasma ratios | Cannot estimate intracellular drug levels |
| 5. Patient Variability | Accounts for variability and disease status with virtual populations (SimPops) | Ignores interindividual differences |
| 6. Interpretability | Transparent mechanism-based and clinically relevant predictions with clear rationale | Opaque model predictions without mechanistic explanation |
AI is a powerful tool—but it cannot replace physiologically grounded, mechanistic modeling in contexts where causality and biological relevance matter most. When it comes to predicting DILI risk, the QST framework utilized by DILIsym offers a far more robust, transparent, and clinically relevant solution. For drug developers and safety scientists aiming to reduce late-stage failures and protect patients, DILIsym does not just offer predictions—it delivers scientific confidence and enables exploration of “what if” scenarios.
Importantly, recent innovations like the Liver Safety+ package from Simulations Plus show how AI and machine learning approaches can be thoughtfully integrated with QST models to enhance early-stage risk assessment. Liver Safety+ leverages predictive AI models to quickly flag potential concerns, while DILIsym’s physiologically-based, mechanistic framework simulates the complex biological processes underlying hepatotoxicity. Together, they provide a powerful, complementary toolkit for early prioritization and de-risking of drug candidates.