AI Race in Drug Discovery Intensifies as Insilico Medicine Brings AI-generated Molecule To First-in-human Trial

by Andrii Buvailo

Disclaimer: All opinions expressed by Contributors are their own and do not represent those of their employers, or
Contributors are fully responsible for assuring they own any required copyright for any content they submit to This website and its owners shall not be liable for neither information and content submitted for publication by Contributors, nor its accuracy.

   1434    Comments 0

Getting a step closer to treating Idiopathic Pulmonary Fibrosis 

A Hong Kong-based company applying a state-of-the-art artificial intelligence (AI) platform for drug discovery, Insilico Medicine, has announced the start of a first-in-human microdose trial of ISM001-055 — a potentially first-in-class small molecule inhibitor of a novel biological target developed by the company for the treatment of idiopathic pulmonary fibrosis (IPF). This news may become a ray of hope for an increasing number of people, primarily the elderly, diagnosed with IPF -- an irreversible progressive disease with median survival time ranging from 2.5 to 3.5 years after the first diagnosis if untreated, with few cures currently available on the market.

Insilico Medicine, a company earlier highlighted in BiopharmaTrend’s yearly review “R&D Trends To Follow In Pharmaceutical Industry In 2021 And Beyond'', reported in a press release that the first healthy volunteer in an Australian study has received a limited, intravenous dose of ISM001-055. This small molecule showed promising efficacy for IPF and a good safety profile that led to its nomination as a preclinical drug candidate in December 2020 for IND-enabling studies. The molecule has been de-novo designed using Chemistry42, a subsystem of the company’s Artificial Intelligence-based end-to-end R&D platform Pharma.AI.

The molecule ISM001-055 is an inhibitor for a novel pan-fibrotic target earlier also discovered by Insilico Medicine using PandaOmics -- a target discovery subsystem of Pharma.AI, connected with Chemistry42 in an integrated workflow. Pharma.AI also includes its third major subsystem -- InClinico, which can predict clinical trials success rates, recognize the weak points in trial design, while adopting the best practices in the industry.

Notably, Insilico Medicine completed the entire discovery process from target discovery to preclinical candidate nomination within 18 months on a budget of $2.6 million, demonstrating a record-setting performance of its R&D platform. “We believe this is a significant milestone in the history of AI-powered drug discovery because to our knowledge the drug candidate is the first-ever AI-discovered novel molecule based on an AI-discovered novel target. We have leveraged our end-to-end AI-powered drug discovery platform, including the usage of generative biology and generative chemistry, to discover novel biological targets and generate novel molecules with drug-like properties. ISM001-055 is the first such compound to enter the clinic, and we expect more to come in the near future," Insilico Medicine’s Chief Scientific Officer, Feng Ren, Ph.D., said in a statement.

Blazing the trail to AI-enabled drug discovery model

Insilico Medicine was founded in 2014 with the idea of adopting deep learning to solve the challenges of pharmaceutical research. Those were the times of a global rise of deep learning, particularly in the area of image recognition and video processing -- ignited by a wave of successes in theoretical and practical machine learning. For instance, in 2012 a convolutional neural architecture AlexNet scored a solid win in the ImageNet competition, demonstrating the outstanding ability of deep learned neural nets to work with images. The same year, Google announced its AI system managed to spontaneously identify a cat in a series of youtube videos -- via unsupervised learning.  

Generative Adversarial Networks (GANs)

The GAN model architecture includes two sub-models: a generator model for designing new examples and a discriminator model (adversary) which tries to classify if generated examples are real (from training dataset), or fake -- generated by the generator model. The two models compete against each other until GAN excels at generating examples non-distinguishable from training data, and belonging to the same distribution as the training data. 

In 2014 Ian Goodfellow introduced an innovative deep learning architecture -- generative adversarial networks (GANs) as a general concept to solve de-novo design problems in various fields. Just a couple of years later, Insilico Medicine pioneered a successful application of GANs to specific problems in drug discovery, such as de-novo molecular design.

The team at Insilico developed several improvements and new features to their GAN-based AI platform for drug design and started patenting findings. In 2017 they built multiple working GAN models, including druGAN for fingerprints, ORGAN for SMILES, various recurrent neural networks (RNN) architectures with reinforcement learning and LSTM, agile temporal convolutional networks (ACTNs), and a reinforced adversarial neural computer (RANC).

GAN model

Image credit: DOI:10.1002/cpt.1795

“We pioneered this particular field of AI in drug discovery but I think that the application of the Generative Adversarial Networks (GANs) and combinations of GANs with Reinforcement Learning (RL) to the generation of novel molecular structures with the desired properties was the real breakthrough”, commented Dr. Alex Zhavoronkov for BiopharmaTrend interview back in 2017.

Reinforcement Learning (RL)

During the system’s (agent’s) interaction with the environment, where it is performing a certain task, the agent receives rewards for right guesses and punishment for wrong ones. This reward/punishment process guides the agent towards the most optimal performance

In 2018 the сompany built and validated a powerful deep generative model, generative tensorial reinforcement learning (GENTRL). GENTRL was a new AI system for drug discovery that dramatically accelerated the process of lead discovery from years to days. The code for GENTRL is publicly available on GitHub for exploration and building upon it. 

Eventually, Insilico Medicine's team managed to build an end-to-end AI platform Pharma.AI with three key components; a target discovery and multi-omics data analysis engine PandaOmics, a de novo molecular design engine Chemistry42, and a clinical trial outcomes prediction engine InClinico.

Image credit:

The Pharma.AI has been assembled over several years, and a number of notable validation studies were published to demonstrate the ability of the platform for efficient drug discovery. In 2018 Insilico Medicine published research revealing the first JAK3 inhibitor generated using the Entangled Conditional Autoencoder (ECAAE) with experimental validation. At that time, the engines could already achieve reasonable hit rates for GPCRs and other target classes.

In 2019 Insilico Medicine made headlines with a significant proof-of-concept milestone where they predicted a molecule for a well-known target DDR1 in about 20 days — and successfully validated prediction in vitro and in vivo. 

Finally, in 2020-2021 the company made significant progress with its AI platform and in February 2021 announced the nomination of a preclinical drug candidate (ISM001-055) for IPF, a small molecule inhibitor designed by Chemistry42 for a novel pan-fibrotic target discovered earlier by PandaOmics -- a molecule which now entered the clinical program. Two months later, in April, the company announced a second preclinical drug candidate — for Kidney Fibrosis, which is also planned to be advanced to the clinical stage by the end of 2022. 

Insilico Medicine was created by its founders to be the “AI-first” platform-based company from day zero -- in contrast to many “traditional” drug discovery startups usually born around a particular biological idea or an academic breakthrough. It seems that the company’s investment into building the end-to-end platform early on has started paying back at present -- by the ability of the company to generate strong candidate molecules in a “conveyor” fashion.

The rise of drug discovery “super-platforms” powered by machine learning

Insilico Medicine is not alone in its quest to engage advanced artificial intelligence technologies for more efficient drug discovery. According to an analytical report by Deep Pharma Intelligence, the AI for drug discovery community now includes more than 250 companies backed by venture capital money, a flock of pharmaceutical corporations actively building AI strategies, and largest contract research organizations -- the latter also actively tapping AI potential to augment their own service capabilities. A separate group -- large technological corporations, such as Google, Microsoft, Tencent, NVIDIA, and others, which have made active advancements lately into the area of pharmaceutical research and healthcare. Such corporations possess state-of-the-art artificial intelligence technologies, data infrastructures, and agile innovation models, which gives them a strong edge in rapidly developing business in such data-rich fields as pharmaceutical research, biotech, and healthcare.

AI community

Image credit: Deep Pharma Intelligence

The perception of the role of artificial intelligence technologies in the pharmaceutical industry has been shifting from hype and skepticism around 2014-2016 to a notable rise of interest in 2017-2018, and skyrocketing growth of attention since 2019 -- a year that is referred to by Deep Pharma Intelligence analysts as “AlphaGo” moment for the pharmaceutical artificial intelligence. That year a number of proof-of-concept results of AI-assisted drug design by Deep Genomics, Exscientia, Insilico Medicine, Recursion Pharmaceuticals, and some other companies received considerable media attention and demonstrated the outstanding practical ability of integrated research platforms -- those connecting multiple AI modules for disease modeling, target discovery, generative drug design, drug repurposing, and biomarker discovery -- to substantially cut time and cost of drug discovery, as well as helping innovate. 

RELATED: The Overview of AI in Drug Discovery in 2019: The “Proof-of-concept Year”

For instance, Canadian-based Deep Genomics made headlines in 2019 with discovering a novel target and a novel RNA therapeutics candidate for rare Wilson disease using their AI Workbench platform -- all within 18 months of initiating target discovery effort. The same year, Insilico Medicine’s deep learning software had a major proof of concept run where it managed to “imagine” novel molecules -- potent inhibitors for a known target DDR1 kinase -- within just days. 

A digital biology company Recursion Pharmaceuticals, founded in 2013, managed to rapidly build an impressive pipeline of preclinical and clinical drug candidates for various indications using their AI-enabled automation and cell image analysis platform. In 2019, the company had a notable success story with Takeda, having reached option exercise for two of its AI-generated drug candidates for rare diseases in less than 18 months. Recursion went public in April 2021, banking more than $430 million.

Pharma AlphaGo moment

Image credit: Deep Pharma Intelligence

From 2019 till the present, the progress of pharmaceutical AI has been following a rapid growth trajectory both in terms of the number of new preclinical and clinical drug candidates discovered via AI-assisted de-novo design or drug repurposing, and in the amount of funding that investors poured into leading AI-driven drug discovery and biotech startups (see left-hand chart below). There is also a growing number of R&D collaborations, where largest pharmaceutical companies hire AI-platform providers for contract research type of work towards modeling diseases and biological systems (modeling as a service), discovering novel targets, or designing innovative lead candidates for pre-defined therapeutic areas or known targets of choice (see right-hand chart below).

Loading chart1...
Loading chart2...

Source of data: “The Landscape of Artificial Intelligence (AI) In Pharmaceutical R&D

Leaders of the pharmaceutical AI race

With many drug discovery companies advertising the use of artificial intelligence technologies, the availability of high-quality research publications and proof-of-concept validation studies becomes a critical part of assessing the innovation potential and robustness of novel R&D platforms such AI companies offer. 

There are clear leaders of the pack, which keep demonstrating research proofs and deliver novel preclinical and clinical candidates with regularity. Those are usually those types of companies that started building AI-based platforms from the early days of their foundation. They managed by now to create and validate sophisticated machine learning pipelines, multi-component systems, and well-concerted processes to run the whole spectrum of drug discovery. 

For example, Insilico Medicine started as a deep learning company and was initially focused on generative models for drug design. Over the years, the company managed to gradually build an “end-to-end” drug discovery system, including hundreds of components -- everything from automatically building disease models and target discovery to generative drug design and even clinical trial modeling. The system is capable of processing a wide range of multimodal data types, from research papers and patents to omics and molecular activity data.

BenevolentAI early in the development built its integrated Benevolent platform with Knowledge Graph being at the heart of target discovery and drug repurposing. 

Exscientia developed its end-to-end AI platform Centaur Chemist which works in a combination with human expertise. 

Recursion Pharmaceuticals was built as a digital biology company in the first place, combining AI-enabled modeling capabilities with the robotization of their biology research. 

Deep Genomics developed AI Workbench as a cornerstone of its research process, capable of novel target and lead discovery towards oligonucleotide therapies.  

Lantern Pharmaceuticals developed RADR platform that surpassed 10 billion data points, facilitating increased drug and cancer type-specific biomarker identification, the discovery of new indications, and the identification of additional drug candidates for cancer treatment. 

Standigm built an AI platform-enabled process of drug discovery, in combination with wet experiments, which allows it to design and discover lead compounds quickly even when target structure information is lacking. 

Insitro built a sophisticated AI platform capable of creating state-of-the-art cell-based disease models and utilizing population-scale data, and biology at scale to create novel medicines. 

RELATED: A New Breed of Biotechs is Taking a Lead

There are just under 50 companies on the market that now offer end-to-end AI platforms with capabilities to discover targets and drug candidates in both small molecule and biologics categories. A larger number of AI companies have more specialized AI platforms -- either for modeling biology and target discovery or specifically for drug design based on known targets. Even a greater number of companies, touting the AI-driven capabilities, are primarily using AI to augment existing R&D processes -- resembling more a “traditional” drug discovery or biotech companies with well-adopted computational processes. 

Considering the wave of hype in the AI field, and a great number of low-quality machine learning practitioners trying to ride the wave of AI trend to get easier time during fundraising, it is crucial for AI-driven companies to constantly deliver results, and publish research validation of their platforms. The presence of a solid preclinical or better -- clinical pipeline of rapidly generated drug candidates is the best quality mark. Last but not least, it is important to publish key findings and explain methodologies for the application of AI technologies -- a typical marker of leading AI companies. For example, Insilico Medicine has published more than 200 papers in peer-reviewed journals, AI conferences, and presented results at more than 200 conferences, including NurIPS -- a place to be for an AI pioneer.

Topics: Industry Trends    AI and Big Data   

Share this:              


There are no comments yet. You can be the first.

Leave a Reply

Your email address will not be published. Required fields are marked *