JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) has become a game-changing force in drug discovery, transforming target identification, lead optimization, and precision medicine. Conventional drug development is usually limited by excessive cost, labor-intensive experimental verification, and uncertain therapeutic effects. AI-based models like AlphaFold, AtomNet, and Insilico GANs have proven to be promising in forecasting drug efficacy, toxicity, and molecular interactions. However, their use is still constrained by inconsistency in cross-therapeutic generalizability and a failure to generalize across various disease spaces. Existing AI algorithms excel at particular tasks, like protein structure prediction (AlphaFold) or virtual screening (AtomNet), but tend to work in isolation, limiting their applicability in broader contexts. The problem lies in designing an AI system that can combine several computational approaches to maximize predictive accuracy and therapeutic applicability. This work presents HybridAI, a combinational AI architecture that integrates geometric deep learning (GDL), reinforcement learning (RL), and federated learning to address the shortcomings of standalone AI models. HybridAI bridges the gaps in AI-assisted drug discovery by enhancing cross-therapeutic flexibility, predictive robustness optimization, and speedup in precision medicine. By combining information from various sources, such as ChEMBL and DrugBank, HybridAI can make more precise predictions of drug-target interactions, toxicity profiles, and repurposing potential. This research will (1) systematically contrast the predictive performance of current AI models, (2) assess the performance of HybridAI in drug discovery, and (3) illustrate its practical applicability using a case study on non-small cell lung cancer (NSCLC). Through bridging the gap between computational innovation and medical application, the study underlines the power of hybrid AI architecture in enabling personalized treatments, reducing trial-and-error inefficiencies, and redefining the future of pharmaceutical research based on AI

Objective:

The accelerated growth of artificial intelligence (AI) in drug discovery calls for critical assessment of its predictive validity and therapeutic relevance. The present study proposes to compare the performance of various AI-based models in predicting the outcome of drug therapy and to present a new combinational AI approach, HybridAI, for improving predictive strength and cross-therapeutic versatility.

Methods:

Seven AI models, including AlphaFold¹, AtomNet², and Insilico GANs³, were comprehensively evaluated for predicting drug efficacy, toxicity, and binding affinity in four therapeutic areas: oncology, antimicrobial resistance, neurodegenerative diseases, and autoimmune disorders. The evaluation was performed using normalized metrics like receiver operating characteristic (ROC-AUC), root mean square deviation (RMSD), and hit-rate accuracy. HybridAI, a novel combinational model that combines geometric deep learning (GDL)⁴, reinforcement learning (RL)⁵, and federated learning⁶, was validated on a 150 structurally diverse compound dataset derived from ChEMBL⁷ and DrugBank⁸.

Results:

Comparative analysis indicated that the current AI models have 78–85% accuracy in target-specific drug design but display wide variation (12–28%) in cross-therapeutic generalizability. HybridAI surpassed single models by predicting 92% drug-kinase interactions (vs. 79% with AlphaFold¹) and making a 34% decrease in errors in predicting toxicity compared to standard ADMET predictors. HybridAI was cross-validated by case study through repurposing of kinase inhibitors on non-small cell lung cancer (NSCLC correct prediction of afatinib¹⁰ through 89% binding affinity and later confirmed in vitro within a time frame of 14 days.

Conclusions:

The results emphasize the limitation of individual AI models in drug discovery and point to the need for hybrid AI architectures to provide higher predictive reliability. Through multi-modal learning methodologies, HybridAI provides a scalable and flexible platform for the acceleration of precision medicine, the minimization of inefficiencies in drug development, and personalization of therapeutic approaches.