Accelerating drug discovery from bed to benchside

Silicon Valley giant NVIDIA is teaming up with pharma company AstraZeneca and the University of Florida on new artificial intelligence research projects aimed at boosting drug discovery and patient care.

April 21, NVIDIA and AstraZeneca revealed a new drug-discovery model called MegaMoIBART, which is aimed at “reaction prediction, molecular optimization and de novo molecular generation.” MegaMoIBART will be deployable on NVIDIA’s platform for computational drug discovery, known as Clara Discovery, and will use a new kind of technology called transformer neural networks.

This is the new breed of press releases flooding the domain of drug discovery, until recently the field of pure pharma & life sciences companies, medical chemistry procedures and very time-consuming biologic research.

Indeed. I used to discover and develop novel candidate drugs myself. One made it into clinical trials. A humanised antibody to treat pediatric brain cancer, developed by Oncurious, a company I helped found. Zero of the work was digitised. Nowhere it was. Painstakingly slow. This was 7 years ago. Today, that is about to change. Since then, almost as a pet project, I have started to ‘collect’ emerging stand-ups and start-ups, which I believe could shake-up the current model of drug discovery.

Like the way we map novel digital solutions on patient journeys to create delight, the Healthskouts team started to envision a drug’s journey in a similar fashion. How could one create drug discovery delight and remove frictions from the current process.

Eventually reverse engineering from the patient’s needs back to the lab: from bed to benchside. Automating every repetitive step in that process, something AI is good at. Solutions which, when stitched together, could help to dramatically reform the current drug discovery process.

Time was on our side. Last year DeepMind’s breakthrough AI system AlphaFold2 was recognised as a solution to the 50-year-old grand challenge of protein folding, capable of predicting the 3D structure of a protein directly from its amino acid sequence to atomic-level accuracy. This has been a watershed moment for computational and AI methods for biology. Building on this advance, November 4 saw the announcement of the creation of a new Alphabet (Google parent) company: Isomorphic Labs. This commercial venture has a mission “to reimagine the entire drug discovery process from first principles with an AI-first approach and, ultimately, to model and understand some of the fundamental mechanisms of life.”

While we’re propagating a while already that the real revolution which will transform healthcare is the biological one – standing on the shoulders of the digital revolution, it is heart-warming to read the following in Isomorphic Labs ‘beliefs: “But just as mathematics turned out to be the right description language for physics, biology may turn out to be the perfect type of regime for the application of AI”.

Therefore, it may come as no surprise that AI drug development startups raised already $2.1B in the 1st half of 2021. So let’s dig into these newbies, and see where they fit in.

Also, multiple pharma partnerships illustrate the burgeoning interest in applying artificial intelligence tools to drug research and development. A good overview on that here.

We focus on developments that we classify in 4 grand domains:

1. Towards novel biology insights (beyond modelling)
2. Fast screening, hit & lead identification -> both for small molecules, and the broad area of biologicals, beyond antibodies
3. Preclinical research using digital twins 
4. Digital & remote Lab operations

1. Biology insights

Any drug development requires proper insight in interplay between candidate drug targets and pathways. Therefore, novel data science approaches which can deal with massive omics datasets and learn from them, are badly needed. Some come out of research labs, covering topics like a gut microbial Signature for Colorectal Cancer Identified Using Machine Learning, or the design of an AI tool called EVE (Evolutionary model of Variant Effect), which uses a sophisticated type of machine learning to interpret meaning of human gene variants as benign or disease-causing. But others are already turned into novel startups. The following companies exemplify this nicely:

• Immunai’s end-to-end single-cell platform enables high resolution profiling of the immune system at industrial scale, generating biological insights that accelerate R&D and power novel discoveries.

• CytoReason is an Israeli company developing a computational model of the human body for faster drug discovery and development. Their cell-centered disease models provide and improve the mechanistic understanding of the human body

• Drug Farm‘s IDInVivo platform allows utilizing genetic screens to directly identify novel drug targets in living animals with intact immune systems. Drug Farm’s MedChem5 platform harnesses the power of deep learning technologies to accelerate first-in-class drugs development. 

2. Fast molecule screening, design, selection and/or optimisation

Here we see very focused players, concentrating on one aspect as well as a new breed of companies applying data science approaches to the entire drug discovery process.

Artificial intelligence is now capable of generating novel, functionally active proteins, thanks to recently published work by researchers from Chalmers University of Technology, Sweden. This should lead to faster and more cost-efficient development of protein-based drugs.

The new tool, published in the journal Nature Machine Intelligence is an AI-based approach called ProteinGAN, which uses a generative deep learning approach. In essence, the AI is provided with a large amount of data from well-studied proteins; it studies this data and attempts to create new proteins based on it. At the same time, another part of the AI tries to figure out if the synthetic proteins are fake or not. The proteins are sent back and forth in the system until the AI cannot tell apart natural and synthetic proteins anymore. This method is well known for creating photos and videos of people who do not exist, but in this study, it was used for producing highly diverse protein variants with naturalistic-like physical properties that could be tested for their functions.

In the first category, a special mention to

• As an AI/Biotech hybrid, Kuano is using a novel approach to generate better enzyme inhibitors. Using quantum computing, their uniquely positioned, proprietary, discovery platform addresses key liabilities such as specificity, potency, and resistance. With 45% of cancer drugs being enzymes … the space to improve is vast

• Entos combines machine learning and automated chemistry to revolutionize small-molecule therapeutics design.

• Cyclica uses a ‘computational approach to poly-pharmacology’. Their platforms, Ligand Express® and Ligand Design™, use a proteome-wide lens to evaluate multiple novel and rare on- and off-target interactions simultaneously.

• BigHat biosciences integrates a wet lab for high-speed characterization with machine learning technologies to guide the search for better antibodies

• Arbor Biotechnologies employs a diverse set of technologies and techniques – including artificial intelligence, genome sequencing, gene synthesis and high-throughput screening – for accelerating the discovery of proteins for improving human health and sustainability

In the second category, a special mention to

• Exscientia On 30th January 2020 Exscientia announced the first molecule designed by Artificial Intelligence (AI) to enter a Phase 1 Clinical trial. The compound is being progressed by collaborator Sumitomo Dainippon Pharma as a long-acting and potent serotonin 5-HT1A receptor agonist, with its phase I clinical study to treat obsessive-compulsive disorder as an indication candidate. The project required less than 12 months to complete the exploratory research phase.

• Recursion (‘the Digital biology company’) operates a synchronized combination of hardware, software and data used to industrialize drug discovery

• Valo has created a unique in silico and in lab-experimental platform that rapidly iterates to design drugs, making targeted and specific small molecules ‘engineerable’. Valo can screen billions of molecules empirically in a matter of days to weeks.

• Deepcure optimizes all phases of drug-discovery, including early hit identification, hit-to-lead, lead optimization, patent strategy, and preparation for IND filing, yielding shorter timelines to develop the highest-quality drug candidates. DeepCure’s proprietary molecular database (MolDB), is the world’s largest medicinal chemistry database that covers over one trillion compounds. The company operates an automated wet lab as well to synthesize selectd molecules.

• Insitro’s approach to rethinking drug discovery and development is fueled by three strategic pillars: 1) machine learning-enabled statistical genetics on deeply phenotyped human cohorts to discover targets and patient segments with potential to inform clinical strategy, 2) predictive cell-based disease models to discover targets, patient segments and drugs; 3) and machine learning-enabled therapeutics design.

• Metagenomi has a proprietary metagenomics approach that mines the Earth’s natural environment to discover next-generation gene editing tools, akin the CRISPR nucleases discovered in bacteria. Moderna, the ones with Corona vaccine fame, just signed a deal with this genome-editing startup

• Isomorphic Laboratories Google’s parent company Alphabet Inc. is diving into the drug discovery game with an AI company built on the protein-folding simulation successes achieved at DeepMind. Last year DeepMind’s breakthrough AI system #AlphaFold2 was recognised as a solution to the 50-year-old grand challenge of protein folding, capable of predicting the 3D structure of a protein directly from its amino acid sequence to atomic-level accuracy. This has been a watershed moment for computational and AI methods for biology. The new Isomorphic Laboratories aims to deliver an “AI-first approach” for tackling biopharma research, with the goal of serving as a commercial partner to drugmakers and figuring out how medicines will react within the body.

• Modulus Therapeutics‘s Convergent Design™ platform combines experimental gene-editing and computational advances into a systematic and repeatable process for cell therapy optimization.

3. Preclinical models using digital twins

While digital twins constitute a topic for a future blog by itself – note that I co-lead the healthcare working group in the Digital Twin Consortium – the concept of digital twins has made its entry in the field of drug discovery and clinical trial execution as well. A few examples:

• Nova Discovery with Artificial Intelligence and virtual populations at the heart of their Jinko platform, JINKO’s workflows are designed to closely resemble that of real life clinical studies. Scenarios represent variations in disease, treatment options, and other study-related factors that, combined with the virtual patient’s own characteristics, make up the input to a model simulation.

• MD Clone helped put Synthetic data on the research agenda. Synthetic data is non-reversible, artificially created data that replicates the statistical characteristics and correlations of real-world, raw data.Utilizing both discrete and non-discrete variables of interest, synthetic data does not contain identifiable information because it uses a statistical approach to create a brand new data set. While it’s possible to identify an individual with anonymized data or de-identified data by inferring characteristics, cross-referencing data similarities, or reversing the data approach, MDClone’s synthetic data is an anonymization method that fully prevents re-identification. This way one can balance patient privacy, legal, compliance, and security issues which often hinder individuals, teams, and organizations as a whole from working together to share data. 

• Unlearn.AI, a start-up based in San Francisco, creates digital twins to model disease progression in Alzheimer’s disease. The virtual twins serve as control arms in clinical studies. Unlearn uses data from past clinical trials and has created a machine learning tool that can simulate accurate control subjects. Using these twins could reduce bias that usually comes with human trials.

4. Digital & remote Lab operations

Probably the most archaic pictures of a lab: people in a white coat – staring at a tube – and erlenmeyers with a blue liquid. I invite you to find such a lab in the real world. That being said, labs are ready for automation big time. Here are a few players working on that.

• LabFellows helps Artificial Intelligence in Drug Discovery companies access Life Science APIs to connect their LabOps securely. Their Life Science APIs and open framework integrates with systems one already use and trust. From accounting and invoicing software to project management, barcoding, and HR tools, FUSION brings them all together.

• HelixAI builds voice-powered digital assistants for scientific laboratories. Your lab’s assistant can respond to natural language queries with real-time auditory responses; allowing access to your lab’s custom information in a completely hands-free way. In one feature Helix helps scientists complete their experimental procedures by relaying the steps of a protocol in a stepwise fashion. Through the dashboard, scientists can add protocols, Standard Operating Procedures (SOPs), and solution recipes to their digital assistant. This information can then be accessed by simple voice prompts.

• Strateos is reimagining and redefining laboratories as smart data centers by enabling both remote-control and on-demand access to  cloud labs for rapid and clean data generation, and the ability for organizations to deploy their software to automate, command and control their own private laboratories even in labs with no automation currently.

• Opentrons makes open source, flexible, user-friendly robots for life scientists. Our mission is to provide the scientific community with a common platform to easily share protocols and reproduce each other’s results. Their robots automate experiments that would otherwise be done by hand. 

An interesting emerging tool I like to mention here is as well is developed by Labster. Aimed for students, Labster allows to work through real-life case stories, interact with lab equipment, perform virtual laboratory experiments and learn with theory and quiz questions. Accessible on laptop and desktop computers, Labster simulation can be played without installing any browser plugins. Focused on education for now – “empowering the next generation of students to change the world” – imagine you providing access to your operations in such a way as well. See the engagement and future recruitment potential? And can you imagine this next happening in the so-called metaverse (the MESH, or Web 3.0)? Stay tuned for how that will look like.

As said, stitching these solutions from the above domains together, allows to revamp the drug discovery process dramatically. However, gaps remain, not covered yet by digital solutions, or at least not by companies we’re aware of. Therefore if you see yourself fit in the bigger scheme, please reach out and we’re happy to add you to our Healthskouts database for pharma companies to find you.

Lastly, in a recent longread by Andreesen’s Vijay Pande, he argues the new era of industrialized bio — enabled by AI as well as an ongoing, foundational shift in biology from empirical science to more engineered approaches — will be the next industrial revolution in human history. In the bio and healthcare market, AI can be extremely helpful: It helps turn the things that used to be expensive, human-labor intensive, less efficient, and less accessible, into less costly, more efficient, and even more effective “compute”. 

As a result, we will now see the emergence of bio’s version of GAFA – playing off the “Google Amazon Facebook Apple” of the leading companies in computing, social, mobile – but for bio. 

I believe so too. Therefore watch companies like Ginkgo Bioworks, programming cells to make everything from food to materials to therapeutics. Or Tempus, making precision medicine a reality through the power and promise of data and artificial intelligence. We’re working on a predictor to see them rise before others do 🙂