An integrated deep learning model accelerates luciferase based high throughput drug screening.

Abstract

High-throughput screening presents clear advantages in accelerating drug development efficiency, but also faces challenges such as high costs, time-consuming processes, and labor-intensive procedures. To address these issues, we developed an integrated deep learning model to find patterns between the structural and molecular characteristics of compounds and our well-established luciferase based HTS values. We utilized about 100,000 HTS values from 18,840 compounds in five luciferase assays including STAT&NFκB system, PPAR system, P53 system, WNT system, and HIF system. Following AI-prediction for putative targeted hit compounds from 8,713 compounds, the in vitro and in vivo experimental validation was performed, and drug candidates (inhibitors or activators) with anti-inflammatory, anti-tumor or anti-metabolic syndrome were identified. T4230 exerts its anti-inflammatory effects by inhibiting the expression of inflammatory factors. The classification performance of the compounds after the joint screening exceeded the performance of the respective sub-models when screened independently and the screening accuracy and efficiency improved 7.08 to 32.04-fold across these five systems compared to our conventional HTS. The integrated AI-conducted HTS model we have developed could reduce R&D costs and accelerate the drug development process, making it a valuable referential pipeline for the artificial intelligence accelerated specific signaling pathway-luciferase HTS.