ABSTRACT
Activity-based lead screening and druggability-based structure optimization are usually torn into two independent processes, although both are related to the chemical structure of target compounds. It leads to an unsatisfactory success ratio and inefficient drug development. DeepRLADS, a de novo molecular design assay, was established based on deep reinforcement learning training, integrating activity screening and structure optimization into a single artificial intelligence (AI)-based drug discovery module. Targeting the relatively easy-verified porcine epidemic diarrhea coronavirus (PEDV) main protease (Mpro, 3CLpro), a diverse virtual library of potential 3CLpro inhibitors was created by splitting existing inhibitors, digitizing fragments, and de novo design, focusing on their potential biological activities and pharmacological properties. The AI-designed compound 11b has a novel catechol-pyrazoline structure, presenting efficient inhibition against Mpro, surprising protection against PEDV infection, low toxicity, and favorable pharmacokinetic properties in vitro and in vivo. Low-dose oral 11b (5 mg/kg) reversed PEDV-induced diarrhea in 5-day-old piglets and greatly improved the survival rate (from 0% to 100%). A preclinical study (560 cases) in eight independent farms indicated an inspiring recovery rate (survival rate: 95.8% in treated piglets VS 9.1% in untreated piglets). Compound 11b is the first proven chemical efficiently protecting piglets in a large-scale preclinical study with multiple centers. It suggested that the novel strategy used by DeepRLADS was a hopeful approach for further AI-based drug design.