Credit: Just_Super / iStock / Getty Images Plus

Some say that AI’s potential in drug development is overstated. Others say that it is, if anything, understated. Who is right? Only time will tell. In the meantime, both sides—and the undecided—eagerly await news of AI successes and failures. For those who suspect that AI’s potential is understated, a piece of news has just appeared that will no doubt be taken as confirmation, however partial or contingent it may seem to AI’s doubters.

The news is the appearance of a paper in Nature Biotechnology describing how AI not only enabled the discovery of an antifibrotic target, but also generated a small molecule inhibitor of the target. The inhibitor’s potential was then confirmed in multiple in vivo studies and in a Phase I trial.

The paper, titled “A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models,” describes how scientists at Insilico Medicine and their collaborators leveraged predictive AI technology to bring a drug candidate for idiopathic pulmonary fibrosis (IPF) to the threshold of Phase II development.

“We identify TRAF2- and NCK-interacting kinase (TNIK) as an antifibrotic target,” the article’s authors wrote. “[We] generated INS018_055, a small-molecule TNIK inhibitor, which exhibits desirable drug-like properties and antifibrotic activity across different organs in vivo through oral, inhaled, or topical administration.”

“This work,” the scientists reported, “was completed in roughly 18 months from target discovery to preclinical candidate nomination and demonstrates the capabilities of our generative AI-driven drug-discovery pipeline.”

Besides disclosing raw experimental data, the scientists described how they first trained PandaOmics, the target identification engine of Insilico’s AI platform Pharma.AI, on a collection of omics and clinical datasets related to tissue fibrosis. PandaOmics proposed a potential target list using deep feature synthesis, causality inference, and de novo pathway reconstruction. After that, the natural language processing models of PandaOmics analyzed millions of text files, including patents, publications, grants, and clinical trial databases to further assess the novelty and disease association.

After selecting TNIK as a primary target, the scientists utilized Chemistry42, Insilico Medicine’s generative chemistry engine, to generate novel molecular structures using a structure-based drug design workflow. Chemistry42 combines over 40 generative chemistry algorithms and over 500 pretrained reward models for de novo compound generation, and it can optimize both generation and virtual screening based on expert human feedback.

After multiple iterative screens were run, one promising hit candidate demonstrated activity with nanomolar IC50 values. The scientists further optimized the compound to increase solubility, promote a good ADME safety profile, and mitigate unwanted toxicity while retaining the compound’s affinity for TNIK. Ultimately, this work produced the lead molecule INS018_055, with less than 80 molecules synthesized and tested.

In subsequent preclinical studies, INS018_055 demonstrated significant efficacy in vitro and in vivo studies for IPF and showed promising results in pharmacokinetic and safety studies across multiple cell lines and multiple species. Furthermore, INS018_055 showed panfibrotic inhibitory function, attenuating skin and kidney fibrosis in two additional animal models. Based on these studies, INS018_055 achieved preclinical candidate nomination in February 2021.

“From my perspective, the progress of INS018_055 has significant implications for the drug discovery field,” said Alex Zhavoronkov, PhD, founder and CEO of Insilico Medicine. “It not only serves as a proof-of-concept for Pharma.AI … but sets a precedent for the potential of generative AI to accelerate drug discovery. Using the publication as a guide, one can extrapolate how generative AI drug discovery tools may streamline early discovery efforts.”

At present, two Phase IIa clinical trials of INS018_055 for the treatment of IPF are being conducted in parallel in the United States and China. The studies are randomized, double-blind, placebo-controlled trials designed to evaluate the safety, tolerability, and pharmacokinetics of the lead drug. In addition, the trials will assess the preliminary efficacy of INS018_055 on lung function in IPF patients. As this drug continues to advance, it drives hope for the roughly five million people worldwide suffering from this deadly disease.

Comments about the paper from Bud Mishra, PhD, professor of computer science at New York University, emphasized how the study’s different parts—target selection and drug design—employed different kinds of AI technology: “The first part uses heuristics that are based on the scientific experiences accumulated in the past (the target must be novel, easy to understand in terms of interactions with known pathways, and follow the approaches used by others in guiding drug discovery and clinical trials in the past) and hence ideal for NLP using LLMs. The second part uses randomized heuristics to search and optimize over complex combinatorial spaces using DNNs, capable of dealing with naturally occurring ‘easy instances of hard problems.’”

“Speculatively,” he continued, “the first part will become more difficult over time (hallucination vs. true novelty) and the second part, simpler, as Moore’s law will continue to improve the computational power exponentially.”