Expert Systems (ExSys) has a state-of-the-art platform for small-molecule-based drug discovery models that combines public and proprietary cheminformatic tools and machine learning (ML) algorithms. Our platform supports a wide variety of ML models that evaluate target-based bioactivity, several physicochemical properties, ADME properties, and cell- and target-based toxicity. These ML models are proactively developed in response to the FDA Modernization 2.0 initiative, as well as the European Union 3R policy, regarding New Approach Methodologies (NAMs). Our efforts focus on validated ML models to help reduce or replace animal studies while benefiting early drug discovery. At present, NAMs are subject to rapidly evolving regulatory expectations as the demand for integrated, predictive, human-centered methods is increasing. This talk will briefly describe the platform, present several models relevant to ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) property prediction, and highlight one of the shortcomings in the development and application of ML-based NAMs. Based on external set predictions, we observe an inverse relationship between training set size and uncertainty quantification (UQ). These discrepancies, spanning one to two orders of magnitude, illustrate how model robustness and predictive accuracy can be significantly compromised when training data are unevenly distributed across properties. There is an urgent need for harmonized, large-scale datasets where multiple ADMET properties are measured on the same compounds. Addressing this gap will be essential to ensure that ML-driven NAMs deliver reliable, reproducible, and regulatory-relevant predictions. Until such discrepancies are properly factored in, adoption of ML-based NAMs in drug discovery may be at risk.