Enhancing the confidence of potential targets enriched by similarity-centric models: the crucial role of the similarity threshold.

Background: Computational target fishing (TF) tools have made tremendous progress in narrowing down the set of potential targets, thereby expediting time- and resource-consuming wet-lab experiments. Among these tools, similarity-centric TF methods are particularly prominent and extensively employed to guide target identification in modern research. Despite substantial progress, similarity-centric models still have significant limitations, particularly regarding the confidence of enriched targets. Methods: We constructed several baseline similarity-based TF models to explore supplementary aspects that could enhance the confidence of enriched targets. A high-quality library was first constructed. Multiple fingerprint representations and scoring schemes were applied to construct individual or ensemble models. The leave-one-out-like cross-validation and rigorous validation metrics were used to measure the performance. Based on the performance under different conditions, multiple influential factors, focusing on the similarity threshold, were investigated. Results: Evidence showed that the similarity between the query molecule and the reference ligands that bind to the target could serve as a quantitative measure of the target reliability. The distribution of effective similarity scores for TF was fingerprint-dependent. To highlight the identification of true positives by filtering background noise and to maximize reliability by balancing precision and recall, the corresponding similarity thresholds for each fingerprint type were identified. Furthermore, additional influential factors, including the choice of different fingerprints, the integration of different models, the target-ligand interaction profile, and the promiscuity of the query molecule, were investigated. Conclusion: Collectively, our findings provide novel insights into enhancing the confidence of enriched targets by applying the similarity threshold and other perspectives. These results also lay the groundwork for developing more robust and reliable target prediction models in the future.