Page 43 - AIH-2-2

Artificial Intelligence in Health New drug discovery in the AI era

Leveraging GenAI for drug discovery and pre-clinical cisplatin, lysergic acid diethylamide, meprobamate, and
development, in silico medicine identified a molecule chlorpromazine is no longer expected. Notably, a study
target, generated novel drugs, assessed target binding and published in 2012 indicated that 24% of all marketed
pre-clinical efficacy, and predicted clinical outcomes for drugs and 35% of anticancer drugs have originated from
lead candidates. Following pivotal pre-clinical studies, serendipitous discoveries. 51
“INS018_055” was selected and is now in phase IIa clinical The success of AI depends on data. Large datasets are
trials. Just 18 months after the project began, in February essential for effectively training AI-driven approaches.
2021, the pre-clinical candidate was chosen. Insilico’s Unfortunately, data are sometimes limited, low in quality,
Biology42: PandaOmics and Chemistry42 – generative inconsistent, or biased, compromising the reliability and
chemistry platforms were used to create INS018_055 for accuracy of the findings.
treating idiopathic pulmonary fibrosis. It was developed
from scratch in just 3 weeks, with another 3 weeks to GenAI models trained on skewed or partial data or on
validate the compound for treating fibrosis. This process prior trials of similar medications will reflect these biases
50
would have taken at least 2 years if it had followed the in their results. While GenAI algorithms can explore and
traditional discovery route. Further, to reach clinical develop unique chemical structures previously unexplored
evaluation, it would have taken >$400 million and up by human researchers, they will produce only similar
to 6 years for NCE if pursued through traditional drug chemotypes (me-too moieties) if trained on datasets
discovery methods. These milestones were achieved by primarily consisting of one type of molecular property.
in silico medicine in a third of the time and at a tenth of Consequently, they will be unable to generate results in
the expense. underrepresented chemical spaces, which are vast and
multidimensional.
4. Small but significant challenges To build a robust and dependable AI platform for
Developing AI/ML tools is cost-intensive, with a significant in silico drug discovery, AI systems should be trained on
portion of drug development expenses allocated to clinical the entire drug evolution process, from hit identification to
trials. Although the cost and duration of clinical trials may lead optimization, clinical candidate selection, and market
remain unchanged with AI/ML, these technologies greatly authorization, rather than solely on approved marketed
facilitate the customization of clinical trial protocols, patient products. However, a significant portion of historical
selection, stratification and retention, real-time clinical data from various discovery programs is privately owned
data analysis, and forecasting of safety and efficacy trends. by innovators. The drug discovery and development data
Thus, investing significant time, money, and resources in available in the public domain are stored in silos and
creating these tools is expected to meaningfully reduce the have not been properly connected or integrated. Many
bench-to-bedside timespan and cost. AI businesses are grappling with massive amounts of
disconnected data spread across too many verticals, leaving
However, the advanced coding and programming skills them to learn by doing.
required for AI/ML tool creation make it challenging for
many small and mid-size pharma R&Ds to develop these Occasionally, AI-driven in silico drug development
tools in-house. Consequently, they often rely on in-licensing initiatives produce molecules with structures that are
tools from software tech giants or partnering with them to challenging for medicinal chemists to synthesize in reality.
access AI/ML tools. Using AI/ML tools from software tech Combining GenAI with conventional experimental
companies under non-exclusive agreements carries risks of techniques will enhance the drug development process,
making it faster and more affordable while generating
intellectual property loss or data breaches unless they are more effective and customized candidate molecules.
operated on-premises, such as with “PandaOmics Box.” 36
However, modern AI-based approaches cannot completely
With the advent of AI and ML, NCEs are designed replace traditional experimental techniques as well as
in silico, and their physiochemical characteristics, PK the invaluable knowledge and experience of human
parameters, in vitro and in vivo efficacy, and toxicity researchers. A recent report revealed that the success rate
52
properties are predicted using advanced computational of AI-derived molecules is 80 – 90% in phase I trials but
algorithms. From an initial selection of 50 – 100 molecules, drops to approximately 40% in phase II trials. GenAI can
53
only 5 – 10% that meet the highly desirable predicted only make predictions based on currently accessible data,
parameters are subjected to wet-laboratory profiling. This and experienced human drug hunters are still needed for
approach significantly reduces animal usage and eliminates result validation and interpretation. Thus, GenAI alone
the chances of serendipitous drug discovery. Therefore, may not be reliable in aspects that directly affect people’s
the discovery of molecules such as penicillin, warfarin, health. Nevertheless, there is an opportunity to expedite

Volume 2 Issue 2 (2025) 37 doi: 10.36922/aih.4423