Data Quality as a Bottleneck in AI-Driven Natural Product Discovery: Bias, Completeness, Assay Reliability, Chemical Annotation, and Model Generalizability

Download PDF

Artificial intelligence is increasingly used to support natural product discovery through structure-based prediction, bioactivity modeling, target inference, omics interpretation, ADMET estimation, toxicity prediction, and candidate prioritization. However, the reliability of these approaches depends on the quality of the chemical, biological, assay, source, and annotation data from which models learn. This article develops an original data quality framework for AI-driven natural product discovery. The framework treats data quality as a multidimensional bottleneck involving provenance, completeness, representativeness, assay reliability, chemical annotation, biological context, harmonization, leakage prevention, external validation, domain applicability, uncertainty communication, and auditability. Particular emphasis is placed on sampling bias, publication bias, database bias, missingness, class imbalance, label noise, conflicting records, incomplete metadata, assay variability, uncertain compound identity, stereochemical ambiguity, target annotation quality, and model generalizability. The central argument is that AI-driven natural product discovery is constrained not only by model architecture but also by the reliability, provenance, completeness, and domain relevance of the data used to train, validate, and interpret models. The proposed framework supports cautious research prioritization by defining quality-control gates across raw data, curated data, and model-ready datasets. It is conceptual rather than empirically validated and is intended to guide transparent dataset evaluation, model interpretation, and uncertainty-aware discovery workflows.