Framing the DDI Problem Across Modeling Paradigms

Pharmacokinetic drug–drug interactions emerge when co-administered agents contend for shared routes of absorption, distribution, metabolism, or excretion. A victim drug’s exposure shifts because a perpetrator alters enzymatic capacity, transporter availability, or local microphysiology along the route to systemic circulation. Mechanistic pharmacokinetic modeling encodes these processes through differential equations that represent organs, flows, and clearances as linked compartments. Population pharmacokinetics augments this picture by quantifying variability across individuals and occasions to explain divergent concentration–time profiles within a target group. Machine learning addresses the same question inductively by learning patterns linking molecular, experimental, and clinical descriptors to observed interaction outcomes. The point of comparison is not stylistic preference but epistemology, since each paradigm carries distinct assumptions about causality, generalization, and data dependency.

Early in discovery, the central bottleneck is knowing whether a scaffold will collide with established therapies through a shared enzyme or transporter. Structural alerts and physicochemical heuristics help, but they struggle when subtle substituent placements toggle metabolic fate. Mechanistic models can project interactions once in vitro inputs are in hand, yet those inputs often arrive after synthesis campaigns have committed resources. Conversely, machine learning can triage designs using features available from a drawing tool, which moves risk assessment closer to ideation. The tension is clear: one method asks for biology first and grants interpretability, while the other asks for historical labels and grants speed. Practical pipelines must navigate this tension rather than declare a single winner.

As projects transition to first-in-human studies, the modeling question shifts from possibility to magnitude, timing, and dose adjustment. Physiologically based models can propagate system changes over time, such as enzyme turnover or transporter induction, and tie them to predicted concentration–time curves. Population approaches can then estimate covariate effects of a comedication on clearance or volume while acknowledging random effects. Machine learning can be retrained on the growing internal corpus to refine decision boundaries and regress dose-dependent outcomes. Each update becomes a form of scientific bookkeeping, where new evidence reweights prior beliefs or reshapes the function class that maps inputs to outputs.

Clinical pharmacologists, modelers, and chemists ultimately need a shared decision surface, not a model-of-the-month. That surface should reveal when a mechanistic forecast is well-posed, when a data-driven predictor is within its applicability domain, and when neither should be trusted without new measurements. In practice, the craft lies in defining handoffs that minimize delay and maximize information gain at each stage. A design-time classifier can flag a concern, a mid-stage mechanistic model can simulate a study, and a late-stage population analysis can quantify effects in the label. The next sections examine how each paradigm is built, what it needs, and how to stitch them into a coherent workflow.

Mechanistic Grounding: PBPK and Pop-PK in Practice

Physiologically based pharmacokinetic models represent the body as a network of tissue compartments linked by blood and lymph, each with defined volumes, perfusion, and binding properties. Drug-specific parameters enter as solubility, permeability, fraction unbound, and intrinsic clearance, while system parameters encode enzyme abundance, transporter localization, and organ physiology. Interactions arise when an inhibitor reduces catalytic capacity, an inducer increases synthesis rates, or a substrate competes for transport, and the model integrates these effects into concentration–time predictions. Experimental values from microsomes, hepatocytes, or transporter assays inform catalytic constants and induction scales, with sensitivity analysis exploring plausible ranges. Clinical studies then verify or recalibrate those parameters, yielding a middle-out synthesis that honors both bench and bedside. The output is a set of dynamic predictions that can be interrogated at the level of organ fluxes, intracellular concentrations, and systemic exposure.

Population pharmacokinetics takes a top-down stance by fitting nonlinear mixed-effects models to observed concentration data in well-defined cohorts. Fixed effects describe typical clearances and volumes, random effects capture interindividual and residual variability, and covariates encode comedication status as time-varying factors. Within this framework, a nested drug–drug interaction study becomes a categorical or continuous modifier of parameters, and the model estimates the magnitude of that effect while accounting for sampling schedules and dosing histories. Because the path starts with actual profiles, the method is descriptive by construction, though simulations can extend its reach to untested regimens. The virtue lies in estimating what variability is explained by the interaction versus other sources, which is invaluable for labeling, trial design, and clinical guidance. The drawback is reliance on data collection that occurs relatively late, after exposure to risk and cost.

Mechanistic and population methods are not rivals so much as complements under a shared mathematical umbrella. A physiologically based model can propose a clinical design by simulating plausible perpetrator effects across doses and intervals, and a population model can then analyze the resultant data for precision and covariate structure. The two can iterate when observed effects diverge from expectations, prompting re-examination of metabolism routes, transporter saturation, or induction dynamics. In particularly complex cases, tissue-level predictions from the mechanistic model can rationalize why total plasma exposure shifts modestly while site-of-action concentrations change meaningfully. Decisions about dose adjustment or contraindication then rest on mechanistic plausibility backed by empirical variability estimates. This division of labor yields traceable rationales that withstand regulatory scrutiny.

Still, mechanistic frameworks do not eliminate uncertainty; they relocate it into parameters and structure. In vitro to in vivo extrapolation remains a delicate step, since enzyme kinetics and transporter behavior vary across assay systems and human tissues. Some processes, such as intestinal metabolism or canalicular efflux, resist clean measurement and must be inferred from composite behaviors. Parameter identifiability can also be fragile when multiple processes generate similar exposure shifts, leaving distinct mechanistic stories compatible with the same data. These realities argue for embedding uncertainty propagation directly into simulation and for maintaining a disciplined connection to emerging experimental systems that better emulate human physiology. When such inputs mature, the models gain both realism and credibility.

Inductive Engines: Machine Learning for DDI Risk

Machine learning reframes DDI prediction as a supervised mapping from features to labels, with features drawn from chemistry, ADME assays, knowledge graphs, and curated interactions. For classification tasks, the label may be the presence of a clinically relevant interaction under a given context, while regression tasks target fold-changes in exposure or parameter shifts. Linear models capture additive effects of descriptors, tree ensembles learn non-linear rules, and neural networks approximate complex functions over high-dimensional inputs. Crucially, the model’s coefficients or learned representations are re-estimated whenever new data arrive, which automates adaptation to evolving chemical space and therapeutic practice. Representation learning can compress chemical structure into vector embeddings, while graph-based models integrate drugs, enzymes, transporters, and pathways into a relational canvas. This inductive posture makes the approach attractive where broad historical data exist but mechanistic details are incomplete or slow to obtain.

Early discovery is a natural home for structure-based predictors that operate on SMILES strings, fingerprints, or learned embeddings. Similarity profiles relative to known perpetrators or victims can function as signals that a new compound resides near an interaction-prone region of chemical space. When augmented with physicochemical descriptors and early ADME assays, these models can separate designs likely to avoid a shared pathway from those that warrant mechanistic follow-up. Text mining and natural language processing can expand training sets by extracting interaction assertions from the literature, which strengthens generalization without demanding new experiments. Knowledge graphs add explicit relational context by linking molecules to enzymes and transporters, letting the model reason over network neighborhoods rather than isolated descriptors. The result is a design-time filter that steers medicinal chemistry away from collision courses.

Later in development, the feature set can evolve to include doses, dosing intervals, and measured pharmacokinetic parameters for both agents under study. Regression models can then forecast quantitative shifts in area under the curve or peak concentration under specified regimens, guiding dose selection for formal studies. When embedded in model-informed drug development, these predictions can pre-screen scenarios before investing in heavier mechanistic builds. Explainability methods such as Shapley values make the contribution of individual features legible, revealing whether predicted risk stems from lipophilicity, basicity, fraction unbound, or known pathway membership. Uncertainty can be expressed through ensembles or probabilistic layers, which communicates confidence to decision makers and suggests where additional data would most reduce ambiguity. In this regime, machine learning is not replacing mechanistic reasoning; it is accelerating triage and focusing attention.

Applicability domain is the central constraint for inductive models. Predictions are reliable when a new sample resembles the training distribution in feature space and become fragile as novelty increases. Chemical series with unusual ring systems, rare functional groups, or atypical stereochemistry can sit at the frontier where the model extrapolates poorly. The same holds for interaction mechanisms that are underrepresented in the corpus, such as atypical gut metabolism or transporter-mediated enterohepatic cycling. Curating diverse, well-annotated datasets mitigates these risks, but they do not vanish, especially for emerging modalities and routes. A disciplined workflow must therefore include checks for similarity and coverage before acting on a machine prediction.

Bias and confounding also require explicit management. If the training set overrepresents certain enzyme families, lipophilic scaffolds, or dosing ranges, the model may internalize shortcuts that fail under broader deployment. Feature selection and debiasing strategies help, but they are not substitutes for balanced, representative data collection. Synthetic minority oversampling and related resampling methods can stabilize classifiers under imbalance for binary tasks, yet they complicate multiclass labels that delineate interaction magnitude or mechanism. One escape hatch is to cast such problems as regression when continuous labels exist, allowing the model to learn along a spectrum. However, regression amplifies the need for careful curation of measurement uncertainty and alignment across studies.

Data, Uncertainty, and Extrapolation in Comparative Focus

Comparing paradigms begins with data timing and granularity. PBPK needs drug-specific and system-specific inputs with physical meaning, which arrive through in vitro systems, preclinical studies, and early clinical work. Pop-PK needs serial concentration data under designed covariate contrasts, which only exist once volunteers or patients have been dosed. ML can begin with structural descriptors and curated interaction labels drawn from the literature and internal archives, and then fold in richer features as the program matures. These trajectories position the methods at different points along the timeline, each adding value where the others cannot yet operate. Recognizing those positions prevents futile comparisons and encourages sensible sequencing.

Extrapolation divides the camps for principled reasons. Mechanistic models explain why exposure changes and can simulate out-of-distribution settings as long as biology is correctly captured and parameters are plausible. Machine learning is strongest within its observed domain and weakest when forced to infer beyond it, since the mapping is statistical rather than causal. That asymmetry implies that mechanistic forecasts are better suited for special populations, organ impairment, or extreme dosing schedules, where statistical patterns from average cohorts may mislead. It also implies that purely structural classifiers should not be stretched to predict quantitative dose adjustments in the absence of dose features. Hybrids can narrow these gaps by letting mechanistic outputs serve as ML inputs, effectively transferring deductive insights into an inductive learner.

Uncertainty must be modeled rather than ignored. Bayesian variants of PBPK and Pop-PK propagate parameter distributions through simulations to yield credible bands and probabilities of exceeding thresholds. ML can mirror this stance through ensembles, dropout-based approximations, or explicitly probabilistic layers that output distributions rather than points. Calibration then becomes a first-class objective so that nominal confidence aligns with empirical reliability across risk strata. When both sides speak the language of uncertainty, cross-model reconciliation becomes less about dueling point estimates and more about overlapping distributions. Decisions can then be framed in terms of risk tolerance and value of information.

Interpretability is often cited as a separating axis, but it is more nuanced than a slogan. Mechanistic models are interpretable by design because every parameter maps to a biological process, yet complexity and identifiability limits can cloud conclusions. Modern ML explainability can attribute predictions to features consistently, although attribution is not causation and can be gamed by correlated inputs. The practical target is actionable transparency: enough structure to design confirmatory experiments and enough attribution to diagnose failure modes. When combined with prospective validation, that standard keeps both camps honest and aligned with scientific intent. It also supports communication with regulators and clinicians who must implement the resulting guidance.

Toward Hybrid Intelligence and a Stage-Gated Workflow

A pragmatic workflow aligns method to question while allowing feedback between stages. In design, structure-based ML screens prioritize scaffolds with low interaction risk against common co-therapies, reducing wasted synthesis cycles. As ADME data accumulate, ML models incorporate permeability, fraction unbound, intrinsic clearance, and transporter phenotypes to refine risk assessments and flag scenarios for mechanistic builds. PBPK then simulates dynamic co-administration across dosing regimens, meal states, and impairment classes, proposing clinical designs that will be maximally informative. Pop-PK closes the loop by quantifying observed effects and anchoring variability estimates for labeling. Each stage hands actionable artifacts forward, including features, priors, and uncertainty summaries.

Bidirectional coupling boosts both speed and rigor. Mechanistic models can generate synthetic features that enrich ML training, such as predicted gut concentrations, hepatic inlet profiles, or intracellular burdens in relevant tissues. ML can, in turn, predict hard-to-measure mechanistic parameters from chemical and early in vitro data, reducing reliance on single assays with fragile translatability. Knowledge graphs can mediate this exchange by providing a shared schema of drugs, enzymes, transporters, and tissues on which both learners operate. Over time, the system becomes a living model-of-models, where updates in one component reverberate through the others in a controlled manner. Governance and versioning then become essential infrastructure, not afterthoughts.

Hybrid models are already feasible with standard toolchains. A pipeline might use learned embeddings of molecular structure, concatenate them with mechanistic summaries from PBPK simulations, and feed the result to a calibrated regressor for dose-specific interaction magnitude. The regressor’s explanations would reveal whether the predicted risk is driven by lipophilicity and fraction unbound, by a simulated intestinal metabolism bottleneck, or by known transporter overlap. If the mechanistic contribution dominates, the team can prioritize refining that physiology with targeted experiments. If the structural signal dominates, chemistry can explore analogs that move the compound out of the risky neighborhood. This interplay converts modeling into an experimental design engine.

The operational gain is not only accuracy but also faster cycles of learning. Early ML filters reduce downstream surprises, mechanistic simulations focus clinical questions, and population analyses turn experience into covariate wisdom for future programs. Postmarket, ML can monitor spontaneous reports and real-world data for emergent interaction signals when mechanistic parameters are incomplete or discordant across sources. Meanwhile, PBPK can explore explanations for those signals and suggest confirmatory studies with a clear mechanistic hypothesis. The ecosystem matures as a feedback-controlled system rather than a sequence of disconnected analyses. That is how modeling and simulation move from expert craft to organizational capability.

Study DOI: https://doi.org/10.1002/psp4.12870

Engr. Dex Marco Tiu Guibelondo, B.Sc. Pharm, R.Ph., B.Sc. CompE

Editor-in-Chief, PharmaFEATURES