Molecular function-guided unsupervised similarity-based enrichment for low-data virtual screening

Identifying bioactive compounds for orphan or understudied protein targets remains a significant hurdle in drug discovery due to the lack of sufficient {SAR} data. We present MolFuSE (Molecular Function-guided Similarity Explorer), an unsupervised, projection-based enrichment framework designed for low-data virtual screening. MolFuSE leverages manifold learning - specifically {UMAP} - to create low-dimensional similarity spaces based on a reference set of ligands known to bind proteins with the same molecular function as the target.Through systematic evaluation across nine tyrosine transferase targets, we demonstrate that {UMAP} embeddings of Mordred 2D descriptors act as a potent ''potency filter'', significantly outperforming traditional topological fingerprints in the early enrichment of high-potency inhibitors ($<$100~nM). While {ECFP4} fingerprints achieve superior global ranking ({ROC}-{AUC} 0.961 vs.\ 0.863), the descriptor-based approach provides superior early recognition ({EF1}) when the reference set is appropriately curated. Crucially, we identify a critical point in data scaling: performance peaks at a reference set size of approximately 10,000 compounds and degrades for larger sets due to descriptor-space dilution from structurally diverse weak binders. MolFuSE provides a computationally efficient, interpretable tool for prioritizing high-affinity candidates when target-specific data is unavailable.\\ \ Contribution\\ \ introduce a data-pooling strategy for low-data virtual screening that uses molecular-function annotations to aggregate ''pseudo-active'' reference sets, bypassing the need for target-specific ligands or structural data. Evaluating this framework across eight diverse protein families, we show that non-linear {UMAP} projections of physicochemical descriptors act as a potent potency filter, preferentially prioritizing high-affinity binders ($<$~100~nM) in early enrichment compared to traditional topological fingerprints. We provide evidence of a critical scaling threshold where reference sets exceeding approximately 10,000 diverse compounds dilute the descriptor space, establishing a practical blueprint for unsupervised screening across distinct target classes.