Evidence-constrained mechanistic synthesis (ECMS), a framework that classifies what information each finding contains and converts only that information into restrictions on a family of mechanistic hypotheses, is developed for the pre-calibration stage of drug development.
Abstract
Mechanistic drug-development programmes often have more biological evidence than they can safely quantify. We developed evidence-constrained mechanistic synthesis (ECMS), a framework that classifies what information each finding contains and converts only that information into restrictions on a family of mechanistic hypotheses. Evidence shifts the frequency of supported events in a reproducible ensemble rather than being converted into unsupported coefficients or probabilities of biological truth. In a chronic spontaneous urticaria (CSU) implementation, a representative, non-exhaustive corpus of 114 atomic findings from 53 sources and 13 public data resources compiled 18 relation/context constraints and a frozen 4,096-hypothesis ensemble. Regimen evaluation was formulated as continuous multi-node target matching: researchers specify desired changes and importance coefficients for modeled nodes, while package-declared controls vary continuously. A deterministic Sobol-to-block-refinement search, validated on all 4,096 hypotheses, reduced target-matching loss by 27.3% relative to the best of 44 deterministic anchors under a prespecified heuristic demonstration profile; changing the objective profile changed the selected control vector without changing the evidence ensemble. A complementary D-only reference analysis localized decision-relevant uncertainty around the mast-cell-to-disease relation, illustrating that mechanistic prioritization depends on the declared objective. ECMS is intended for the pre-calibration stage of drug development: it makes heterogeneous literature computable while keeping evidence, uncertainty and decision preferences distinct.
ABSTRACT Introduction Drug discovery remains constrained by high attrition rates and the fragmented evaluation of exposure, efficacy, and safety. Mechanistic models offer a biologically grounded framework for connecting these determinants across multiple levels of biological organization. This may help improve translational decision-making by supporting earlier and more integrated assessment of candidate progression. Areas covered This narrative review examines the conceptual basis and current role of next-generation mechanistic models in drug discovery, with emphasis on physiologically based pharmacokinetic models, virtual cell-based assays, quantitative systems pharmacology, artificial intelligence (AI)-augmented mechanistic models, and emerging virtual-cell frameworks. It highlights how these approaches may connect efficacy and safety across biological scales, support in vitro-to-in vivo extrapolation, incorporate in silico predictions, and improve candidate prioritization. The literature was surveyed through PubMed searches conducted up to 25 May 2026. Expert opinion Next-generation mechanistic models are unlikely to transform drug discovery simply by increasing biological detail or computational sophistication. Progress in this direction will depend on standardized data streams, robust validation, explicit model calibration, reproducibility, tighter integration between models, and careful alignment between model design and context of use. Under these conditions, mechanistic frameworks may become important components of a more predictive and less attrition-prone drug discovery pipeline. GRAPHICAL ABSTRACTThe diagram illustrates how next-generation mechanistic models can be achieved through the integration of complementary modeling approaches, such as VCBA, PBPK, QSP, and AOP-based models, interconnected through artificial intelligence architectures. VCBA shows the cell partitioning component with intracellular distribution among structural proteins, aqueous fractions, and lipids, with uptake and elimination fluxes. PBPK represents whole-body drug disposition through arterial and venous circulation across major organs. QSP connects signaling networks, subcellular toxicity pathways, clinical pharmacodynamic responses, and trial design. The AOP-based panel illustrates the hierarchical cascade from a chemical stressor through the molecular initiating event (MIE), key events (KE₁, KE₂), to an organ-level adverse outcome depicted as a liver icon. Below these models, seven categories of fragmented data sources are shown: physicochemical properties, in vitro assays, omics, biomarkers, ADME, toxicity, and clinical data. These data streams connect to the models above, representing the integration of fragmented evidence into a unified mechanistic framework.Diagram of next-generation mechanistic models: VCBA, PBPK, QSP, AOP-based, linked to fragmented data sources.
Gustavo Santos Sandes Felizardo, Vinícius Alexandre Fiaia Costa, Eder Soares de Almeida Santos et al.· Expert Opinion on Drug Disco...· 0 citations
HEDGEHOG is introduced, a unified six-stage filtration benchmark that is inspired by industrial hit identification workflows and exposes a central limitation of current molecular generators: molecules that appear acceptable under isolated criteria rarely satisfy medicinal chemistry, synthesis, docking, and 3D pose filters simultaneously.
Daria A. Ryabchenko, Pavel Gurevich, S. Kadyrov et al.· 0 citations
Artificial intelligence (AI) is increasingly used in drug repurposing to integrate chemical, biological, and clinical data and to prioritize candidate drug–disease associations. Yet current evaluation practices remain dominated by benchmark metrics such as AUROC, AUPRC, F1, or top- k retrieval, which do not by themselves establish pharmacological credibility. A useful prediction is not merely one that scores highly on retrospective datasets, but one that can be connected to a biologically plausible mechanism, safety-relevant context, and a traceable evidentiary basis for experimental follow-up. Mechanistic explainability should therefore be treated as a core objective for translationally oriented AI-enabled repurposing, while also emphasizing that explainability complements rather than replaces experimental validation. Specifically, explanations should connect drugs, targets, pathways, phenotypes, and clinical outcomes in ways that are intelligible to pharmacologists and compatible with experimental prioritization, translational decision-making, and emerging regulatory expectations. We identify a central gap in the literature: many explainability methods emphasize feature attribution or local model transparency, but rarely produce pharmacology-aligned evidence chains that also communicate uncertainty, robustness, and provenance. We introduce biomedical knowledge graphs and neurosymbolic approaches as a promising foundation for more credible repurposing systems because they can support relational inference, structured mechanistic reasoning, and provenance-aware explanation. We further argue that future evaluation should extend beyond predictive accuracy to include mechanistic coherence, uncertainty, robustness, and provenance (MURP) for experimental pharmacology, ideally within workflows that integrate computational prediction with laboratory and, where feasible, real-world validation. Reframing success in this way would improve rigor, translational relevance, and regulatory readiness in AI-driven drug repurposing.
Markus Bertl, Salvatore Sinno· Frontiers in Pharmacology· 0 citations
This work addresses one of the most pervasive obstacles to applying AI in real-world drug development by addressing conformal prediction framework tailored to label shift by weighting conformal scores using marginal label probability ratios and enhancing the trustworthiness of AI-driven predictions.
Hyeonsu Lee, Juyeong Kim, Erkhembayar Jadamba et al.· 0 citations
Clinical drug development is characterized by high attrition rates: approximately 90% of candidates entering Phase I fail to obtain approval. Inadequate efficacy accounts for 40 to 50 percent of terminations in Phase II and Phase III, while safety concerns account for roughly 30 percent. Most efficacy failures result from incorrect biological hypotheses rather than physicochemical deficiencies: drugs frequently bind their intended targets but fail to alter disease progression because the mechanistic assumptions underlying their action are invalid. This perspective introduces a feasibility-gating framework that utilizes physical constraints as explicit, early-stage filters. The framework does not claim that physics alone can predict therapeutic success; rather, it asserts that a systematic evaluation of thermodynamic, kinetic, transport, and microenvironmental constraints can identify candidates likely to fail on physical grounds prior to substantial investment in synthesis and translational research, thereby conserving resources for the inherently experimental validation of biological hypotheses. The individual constraints are based on well-established pharmacological principles; their contribution lies in organizing these principles into a checklist-style decision logic with specific thresholds, explicit guidance on interpreting each threshold, and a clear connection to quantitative systems pharmacology (QSP) and physiologically based pharmacokinetic (PBPK) modeling. We position this framework relative to these established quantitative disciplines, implement each gating criterion, and demonstrate the logic retrospectively using documented development histories. Physical reasoning serves as an essential yet limited filter: it identifies certain predictable physical risks early in the development process without guaranteeing success for candidates that pass through.
Sarfaraz K. Niazi· Frontiers in Drug Discovery· 0 citations
Atomica is presented, a web-based engineering-intelligence platform designed to integrate AI-driven molecular generation with reproducible cheminformatics validation and bioactivity-context retrieval in a single workflow that bridges algorithmic molecular generation and applied, collaboration-ready drug-discovery workflows.