Machine-learning–assisted virtual screening has become an important strategy for accelerating structure-based drug discovery, especially when screening large chemical libraries where exhaustive molecular docking is computationally expensive. However, machine-learning models may fail to recover some high-scoring virtual hits, particularly when compounds occupy chemically diverse or underrepresented regions of chemical space. This study investigated chemical diversity between recalled and non-recalled virtual hits in machine-learning–boosted virtual screening using the HASTEN workflow. Two protein targets were evaluated: LpxH using the 7SS7 crystal structure and AmpC β-lactamase using the 1L2S structure. Docking datasets were generated using Schrödinger-based protein preparation, LigPrep, and Glide workflows. Chemprop-based HASTEN simulations were performed on approximately 500,000-compound reference datasets, and model performance was evaluated using recall of the top 1% of compounds based on true docking scores. Chemical-space and diversity analyses were performed using Morgan fingerprints, UMAP, Tanimoto similarity, and Bemis–Murcko scaffold analysis. Regional models based on UMAP-defined regions were also evaluated and compared with the global reference models. The reference models achieved moderate recall performance, recovering 61.7% and 60.1% of the top 1% compounds for the 7SS7 and 1L2S datasets, respectively. The regional models produced comparable but not improved recall values, with recalls of 58.1% and 60.4% for the 7SS7 regions and 57.5% and 59.2% for the 1L2S regions. However, the regional models recovered additional compounds missed by the reference models, indicating complementary recovery behavior. Chemical-space, scaffold, and Tanimoto similarity analyses showed that missed compounds were structurally diverse and broadly distributed rather than confined to a single scaffold family or isolated region. Evaluation on independent unseen datasets showed substantially lower recall, demonstrating limited generalization to novel chemical space. Overall, this study shows that machine-learning–boosted virtual screening can enrich high-scoring compounds within known chemical space, but accurate ranking near the top-scoring cutoff and generalization to unseen datasets remain important challenges. Regional modeling may provide complementary recovery of additional high-scoring compounds, but it does not substantially improve overall recall or transferability.