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.
Abstract
Drug discovery and development underpins healthcare but remains costly and failure-prone. A critical bottleneck lies in predicting molecular properties such as solubility, potency, and toxicity, which directly determine whether a candidate can advance from preclinical to clinical trials. Artificial Intelligence (AI) has accelerated this process, yet its reliability is often undermined by distribution shift, as experimental conditions frequently diverge from training data. In addition, conventional point predictions provide only single-value estimates, offering limited guidance for high-stakes experimental design. We address these challenges with a conformal prediction framework tailored to label shift. By weighting conformal scores using marginal label probability ratios, our method produces statistically rigorous prediction intervals without retraining. This enables robust uncertainty quantification even when property distributions drift, directly tackling one of the most pervasive obstacles to applying AI in real-world drug development. By moving beyond accuracy alone to provide actionable confidence measures, our approach enhances the trustworthiness of AI-driven predictions. This further aligns predictive modeling with regulatory demands for transparency and uncertainty reporting and ultimately supports more reliable decision-making in billion-dollar development pipelines.
Machine learning applications in preclinical drug development have been focused on automated covariate selection in pharmacometric modeling and high-throughput screening processes early in drug discovery. While inherent drug property prediction has made significant improvements in the past decade, fusing early target-based drug discovery methods to preclinical stage pharmacokinetic (PK) property predictions has been limited. This scoping review investigates the current state of PK property prediction of small molecules in drug discovery using machine learning methods and a combination of machine learning and mechanistic models. We identified major obstacles hindering the development of superior prediction models for small molecule behavior in biological systems. These encompass data accessibility, quantity, and quality, architectural constraints such as poor interpretability and model inherent assumptions, and the lack of robust evaluation and uncertainty assessment methods. To mitigate data-related constraints, we advocate for the use of collaborative federated learning frameworks. Furthermore, we propose leveraging the pattern recognition capabilities of deep learning models in conjunction with the biological interpretability provided by mechanistic approaches to strike an optimal balance between accuracy and biological explainability guided by the intended application of the prediction model. Addressing these limitations will advance reliable modeling pipelines and enable effective extrapolation to novel chemical space, additional species, and emerging drug development scenarios.
Lucille Tomin, Vida Bodaghi-Namileh, D. Schwartz et al.· Journal of Chemical Informat...· 0 citations
Machine learning (ML) models for molecular property prediction are increasingly deployed in drug discovery, yet their adoption in real-world scenarios requires an understanding of the conditions in which a model succeeds or fails. While standardized benchmarks are powerful instruments to measure and unlock progress in ML research, they should not be blindly treated as the end goal. Especially static and retrospective benchmarks, in which no true unknown test set is employed, limit our ability to robustly validate a model’s performance. Building on the collective expertise of a cross-industry consortium, we present a model validation framework consisting of five recommendations that would enable the community to move beyond aggregate metrics toward understanding where and why molecular property prediction models fail. We connect evaluation choices to real-world applications and case studies encountered in pharmaceutical research. The framework proposes splitting strategies that mimic realistic distribution shifts and expose common failure modes. We apply the recommended framework on a recently released dataset of absorption, distribution, metabolism, and excretion (ADME) properties. Across two complementary model algorithms, our case studies reveal four distinct failure modes (extrapolation, interpolation, representation, and evaluation) showing that model errors arise not only from distribution shift but also from limitations in molecular representations. Our results show that commonly used evaluation protocols can significantly overestimate performance and may not detect important model failure modes. All software and data are released via https://github.com/srijitseal/polaris.
Srijit Seal, Akshat Shirish Zalte, David Alencar Araripe et al.· bioRxiv· 0 citations
Artificial intelligence (AI) is rapidly compressing the timeline from molecular discovery to regulatory submission, meaning pipeline therapies will reach formulary committees faster and in greater numbers than ever before. For managed care organizations (MCOs), coverage and contracting decisions must be made earlier, especially for drug classes where the clinical and economic profile remains deeply uncertain. Yet the pipeline monitoring frameworks MCOs rely on, that is, static cost-effectiveness models, manufacturer dossiers, and binary pipeline assumptions, may not keep pace, leaving teams to evaluate therapies with limited visibility into the upstream uncertainty that shapes a drug's real-world value. No current framework links early AI-generated predictions of molecular performance to factors like formulary placement, rebate dynamics, and payer decision-making, leaving MCOs reactive at precisely the moment when anticipatory intelligence would be most valuable. This article proposes a probabilistic AI framework that carries uncertainty from molecular discovery through human trials, health economics, and managed care decision-making. The framework preserves and updates uncertainty across each stage rather than simplifying it into deterministic outcomes. By integrating molecular modeling, Bayesian clinical forecasting, dynamic health economic modeling, and negotiation simulations, the framework can generate early forecasts of formulary placement, cost-effectiveness profile, and more. This article uses lesinurad as an illustrative case study of how drugs can clear every preclinical and regulatory hurdle yet still fail at the formulary-precisely the kind of late-stage surprise anticipatory intelligence is designed to prevent. The 4-layer framework can give MCOs earlier visibility into a therapy's likely clinical profile, cost-effectiveness distribution, and formulary placement probability before a manufacturer's dossier arrives. Such a framework allows budget forecasting and contracting strategy to keep pace with growing AI-accelerated pipelines.
Rishi Sharma· Journal of Managed Care & Sp...· 0 citations
Drug efficacy prediction remains a cornerstone of drug development and precision therapy. However, integrating heterogeneous biomedical data, including multi-omics profiles, pathological imaging, electronic health records, and pharmacokinetic-pharmacodynamic (PK/PD) time-series, faces three fundamental barriers, namely cross-domain distribution shifts between preclinical and clinical data, relational mismatches between isolated vector representations and biological networks, and feature heterogeneity across disparate modalities. To address these challenges, three AI paradigms have emerged, transfer learning for cross-domain alignment, graph neural networks for structured relational modeling, and Transformers for global cross-modal feature interaction. Importantly, these techniques form a many-to-many complementary system rather than a one-to-one correspondence, a key insight that this review explicitly formalizes. We further elaborate encoding workflows for PK/PD data to bridge static molecular signatures with dynamic in vivo exposure trajectories. Four graded clinical applications are outlined, including personalized monotherapy, combination optimization, drug repurposing, and preclinical-to-clinical evaluation of novel candidates. We also dissect persistent bottlenecks such as data harmonization, model interpretability, and prospective validation, and propose five actionable directions, namely privacy-preserving benchmarks, causally interpretable models, temporal dynamic frameworks, cross-domain generalization, and lightweight clinical tools. By integrating theoretical rationales, methodological synergies, and hierarchical translational scenarios, this review provides a unified roadmap to accelerate the clinical deployment of multimodal drug response prediction.
Jing-Wen Fang, Zihan Wang, Jun-Hao Shao et al.· Drug Discoveries & Therapeut...· 0 citations
Polypharmacy requires accurate prediction of drug-drug interactions to prevent adverse events, yet existing models often lack reliability and explainability. We propose T-DDI, a descriptor-based deep learning framework for multi-class drug-drug interaction prediction. Rather than relying on complex graph embeddings, T-DDI uses explicit physicochemical descriptors and an uncertainty-aware estimator to handle severe class imbalance. Evaluated on 868,069 drug pairs spanning 178 interaction types, T-DDI achieves a Macro F1 of 0.8452 on the held-out test set, improving to 0.8992 within the high-confidence subset (87.91% of test samples), outperforming all evaluated baselines within the architectures and datasets considered here. An illustrative prospective case-study assessment on five newly FDA-approved drugs from late 2025 showed that T-DDI can generate mechanistically plausible DDI hypotheses for compounds not used during model development. T-DDI pairs confidence-stratified predictions with LIME-based feature-level explanations and a web application for screening, supporting more reliable drug safety monitoring.
Q. Kha, Duc-Quang-Anh Nguyen, Phi Pham Van Hoang et al.· npj Digital Medicine· 0 citations
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.
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
MIT News · Artificial Intelligence· news.mit.eduAug 24, 2026
A new method for surgically removing training examples from a model reveals that as datasets grow, the link between what a model learns and what it produces dissolves.
