To What Extent Can We Extrapolate Proteochemometric Models: A Case Study for the SLC6 Transporter Family

Proteochemometrics (PCM) modeling combines protein and ligand information to create predictive models for biological activity. It aims to extrapolate information across targets, enabling its application in screening drug candidates across a whole family of proteins. In this study, we investigate the ability of PCM models to extrapolate information from data-rich proteins to data-poor ones and present a series of PCM models for the SLC6 transporter family showing reasonable performance (Q2 values up to 0.79). Moreover, feature importance analysis pointed towards residue position A173 in hSERT, corresponding to G149 and G153 in hNET and hDAT, respectively, to be relevant for subtype selectivity. However, examining the impact of different data splits on model validation metrics highlights potential over-optimism when only considering target stratification splits. Target stratification split only maintains the train/test ratio across all targets, preserving per target balance without considering chemical similarity. However, when performing leave-one-transporter-out studies, a considerable drop in performance was observed. This points towards the need for more complex technologies to exploit the potential of PCM and identify new drug candidates.