AI in Drug Discovery Pipelines: Generative Chemistry and Closed-Loop Experimental Validation

• Author(s): Deepak Saini; Jatin; Ishita; Aman Kumar; Devendra Prasad
• Paper ID: 1719859
• Page: 435-449
• Published Date: 28-02-2023
• Published In: Iconic Research And Engineering Journals
• Publisher: IRE Journals
• e-ISSN: 2456-8880
• Volume/Issue: Volume 6 Issue 8 February-2023

Abstract

Bringing a single new drug to market costs on the order of two to three billion dollars and typically takes ten to fifteen years, with roughly nine of every ten candidates that enter human trials ultimately failing on efficacy or safety grounds. Artificial intelligence has been proposed as a remedy at nearly every stage of this pipeline, from target identification through generative molecular design to preclinical validation. This paper examines the technical architecture of modern AI-driven drug discovery, with particular focus on two complementary components: generative chemistry models that propose novel candidate molecules optimized against multiple pharmacological objectives, and closed-loop validation systems that iteratively test, score, and refine those proposals against wet-laboratory or computational feedback. We survey representative generative approaches, including SMILES-based recurrent and transformer models, graph-based generative networks, reinforcement-learning-driven optimization, and diffusion and structure-based design methods exemplified by AlphaFold-derived engines. We then examine the design-make-test-analyze cycle and its AI-augmented variants, including Bayesian active learning, multi-objective optimization, and robotic self-driving laboratories, and we propose a unified closed-loop reference architecture linking generative proposal, in silico triage, and experimental feedback into a single iterative system. Using case studies of Insilico Medicine's Chemistry42 and rentosertib program, Isomorphic Labs' IsoDDE engine, Exscientia's Centaur Chemist, and Recursion's LOWE platform, we assess what has been achieved to date — including hit rates markedly above traditional high-throughput screening and at least one candidate reaching Phase II trials in under three years — against what remains unresolved, including the continued absence, as of this writing, of any AI-originated drug with full regulatory approval. We conclude that generative chemistry substantially compresses early discovery timelines but that the dominant bottleneck has shifted downstream, to the same efficacy and safety uncertainties that have long limited traditional pipelines, and that closing this gap will require tighter integration between generative design and experimental validation loops rather than further scaling of generative models alone.

Keywords

Generative Chemistry, Drug Discovery, Active Learning, Design-Make-Test-Analyze Cycle, Molecular Generation, Reinforcement Learning, Structure-Based Drug Design, Self-Driving Laboratories

Citations

IRE Journals:
Deepak Saini, Jatin, Ishita, Aman Kumar, Devendra Prasad "AI in Drug Discovery Pipelines: Generative Chemistry and Closed-Loop Experimental Validation" Iconic Research And Engineering Journals Volume 6 Issue 8 2023 Page 435-449

IEEE:
Deepak Saini, Jatin, Ishita, Aman Kumar, Devendra Prasad "AI in Drug Discovery Pipelines: Generative Chemistry and Closed-Loop Experimental Validation" Iconic Research And Engineering Journals, vol. 6, no. 8, Feb. 2023

APA:
Deepak Saini, Jatin, Ishita, Aman Kumar, Devendra Prasad (2023). AI in Drug Discovery Pipelines: Generative Chemistry and Closed-Loop Experimental Validation. Iconic Research And Engineering Journals, 6(8).

MLA:
Deepak Saini, Jatin, Ishita, Aman Kumar, Devendra Prasad "AI in Drug Discovery Pipelines: Generative Chemistry and Closed-Loop Experimental Validation" Iconic Research And Engineering Journals, vol. 6, no. 8, Feb. 2023.

BibTeX

@article{1719859,
author = {Deepak Saini, Jatin, Ishita, Aman Kumar, Devendra Prasad},
title = {AI in Drug Discovery Pipelines: Generative Chemistry and Closed-Loop Experimental Validation},
journal = {Iconic Research And Engineering Journals},
year = {2023},
volume = {6},
number = {8},
pages = {435-449},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1719859.pdf},
abstract = {Bringing a single new drug to market costs on the order of two to three billion dollars and typically takes ten to fifteen years, with roughly nine of every ten candidates that enter human trials ultimately failing on efficacy or safety grounds. Artificial intelligence has been proposed as a remedy at nearly every stage of this pipeline, from target identification through generative molecular design to preclinical validation. This paper examines the technical architecture of modern AI-driven drug discovery, with particular focus on two complementary components: generative chemistry models that propose novel candidate molecules optimized against multiple pharmacological objectives, and closed-loop validation systems that iteratively test, score, and refine those proposals against wet-laboratory or computational feedback. We survey representative generative approaches, including SMILES-based recurrent and transformer models, graph-based generative networks, reinforcement-learning-driven optimization, and diffusion and structure-based design methods exemplified by AlphaFold-derived engines. We then examine the design-make-test-analyze cycle and its AI-augmented variants, including Bayesian active learning, multi-objective optimization, and robotic self-driving laboratories, and we propose a unified closed-loop reference architecture linking generative proposal, in silico triage, and experimental feedback into a single iterative system. Using case studies of Insilico Medicine's Chemistry42 and rentosertib program, Isomorphic Labs' IsoDDE engine, Exscientia's Centaur Chemist, and Recursion's LOWE platform, we assess what has been achieved to date — including hit rates markedly above traditional high-throughput screening and at least one candidate reaching Phase II trials in under three years — against what remains unresolved, including the continued absence, as of this writing, of any AI-originated drug with full regulatory approval. We conclude that generative chemistry substantially compresses early discovery timelines but that the dominant bottleneck has shifted downstream, to the same efficacy and safety uncertainties that have long limited traditional pipelines, and that closing this gap will require tighter integration between generative design and experimental validation loops rather than further scaling of generative models alone.},
keywords = {Generative Chemistry, Drug Discovery, Active Learning, Design-Make-Test-Analyze Cycle, Molecular Generation, Reinforcement Learning, Structure-Based Drug Design, Self-Driving Laboratories},
month = {February}
}