FEAOF: A transferable framework applied to prediction of hERG-related cardiotoxicity

Inhibition of the hERG (human ether-a-go-go-related gene) channel by drug molecules can lead to severe cardiac toxicity, resulting in the withdrawal of many approved drugs from the market or halting their development in later stages. These findings highlight the pressing need to evaluate hERG blockade during drug development. We propose a novel framework for feature extraction and aggregation optimization (FEAOF), which primarily consists of a feature extraction module and an aggregation optimization module. The model integrates diverse ligand representations, including molecular fingerprints, descriptors, and graphs, as well as ligand-receptor interaction features. Based on this integration, we further optimize the algorithmic framework to achieve precise predictions of compounds cardiac toxicity. Two independent test sets exhibiting pronounced structural dissimilarity from the training data were constructed to rigorously assess the model's generalization ability. The results demonstrate that the FEAOF model exhibits strong robustness compared to seven baseline models, achieving F1 score of 66.1 % and 68.1 %. Notably, when benchmarked against five existing models on two external test sets, FEAOF also achieved the highest or near-highest scores across all key evaluation metrics. Importantly, this model can be easily adapted for other drug-target interaction prediction tasks. It is made available as open source under the permissive MIT license at https://github.com/ConfusedAnt/FEAOF.