The article introduces a novel approach called Functional Representation of Gene Signatures (FRoGS) to enhance the prediction of compound-target interactions in bioinformatics. FRoGS represents gene signatures based on their biological functions rather than their identities, similar to how word2vec works in natural language processing. This approach is designed to improve the accuracy of compound-target predictions by leveraging pre-existing knowledge about gene functions. The authors trained a deep learning model using FRoGS and applied it to the Broad Institute's L1000 datasets, demonstrating that it outperforms models based solely on gene identities. By integrating additional pharmacological activity data sources, FRoGS significantly increases the number of high-quality compound-target predictions. The study highlights the utility of FRoGS in machine learning-based bioinformatics applications, particularly in drug target prediction and mechanism of action studies. The approach is validated through simulations and real-world data, showing its ability to extract weak pathway signals and recall known compound targets. The combined model of FRoGS and activity-based models further enhances the prediction accuracy, leading to a high-quality compound-target network. Experimental validation of predicted kinase inhibitors and aryl hydrocarbon receptor ligands supports the effectiveness of the FRoGS approach. The study concludes by discussing the broader applicability of FRoGS in various biomedical problems and the potential for transfer learning and graph neural networks to improve future applications.The article introduces a novel approach called Functional Representation of Gene Signatures (FRoGS) to enhance the prediction of compound-target interactions in bioinformatics. FRoGS represents gene signatures based on their biological functions rather than their identities, similar to how word2vec works in natural language processing. This approach is designed to improve the accuracy of compound-target predictions by leveraging pre-existing knowledge about gene functions. The authors trained a deep learning model using FRoGS and applied it to the Broad Institute's L1000 datasets, demonstrating that it outperforms models based solely on gene identities. By integrating additional pharmacological activity data sources, FRoGS significantly increases the number of high-quality compound-target predictions. The study highlights the utility of FRoGS in machine learning-based bioinformatics applications, particularly in drug target prediction and mechanism of action studies. The approach is validated through simulations and real-world data, showing its ability to extract weak pathway signals and recall known compound targets. The combined model of FRoGS and activity-based models further enhances the prediction accuracy, leading to a high-quality compound-target network. Experimental validation of predicted kinase inhibitors and aryl hydrocarbon receptor ligands supports the effectiveness of the FRoGS approach. The study concludes by discussing the broader applicability of FRoGS in various biomedical problems and the potential for transfer learning and graph neural networks to improve future applications.