Just a few weeks ago, it might have been hard to imagine how AI and drug discovery’s integration could be any more prominent. That is until a Nobel Laureate, a former FDA commissioner, the CEO of Johnson & Johnson, executives from Meta, Illumina and Roche and some of drug discovery’s most prolific innovators from Washington School of Medicine’s Institute of Protein Design and Stanford University unveiled Xaira Life Sciences, armed with $1 billion of capital and headed up by none other than Marc Tessier-Lavigne, former CSO of Genentech.¹ Xaira is the latest example epitomising the industry’s immense appetite for AI-enabled drug discovery, with its platform combining cutting-edge R&D practices, machine learning and data generation. With the AI and drug discovery marriage once again in the spotlight and coupled with lofty promises of transforming global R&D, it’s helpful to take a step back and survey the current AI discovery landscape, learning where we can currently reap the benefits, and the key, practical takeaways Biotechs can take on board for deepening their integration with AI right now. 

‍

The current landscape for AI and drug discovery integration is extensive. In drug target identification alone, AI can enhance predictive modelling (when it learns from existing data to make more accurate predictions about the behavioural qualities of a particular target ahead of any bench experiments), expedite high throughput-screening (thanks to AI’s ability to analyze vast numbers of compounds in minutes), data integration and mining (thanks to seamlessly pulling data from multiple sources including from clinical trials, proteomics, genomics etc.) and more comprehensive drug design ( to predict the behaviour of proteins via deep analysis of its 3D structures), along with numerous others, including enhancing personalized medicine-based approaches, given AI’s ability to analyze patient-specific data to tailor therapeutic development. It also offers specific advantages for both small molecule and large molecule R&D. With the former, AI can predict the Absorption, Distribution, Metabolism, Excretion and Toxicity (ADMET) qualities of a small molecules, along with cheminformatics and compound property prediction (AI models can point to physicochemical properties of small molecules, e.g. stability, ahead of them even reaching the bench), synthesis prediction (whereby AI learns from existing data to predict the outcomes of chemical reactions and suggest optimal synthesis routes) and can design small molecule libraries; collections of structurally diverse chemical compounds to aid scientists in enabling them to test thousands of compounds quickly to identify ones with biological activity against specific targets. Large molecules typically involve more complex target identification processes, with current AI modelling methods more in their infancy here. AI can still play a crucial role the identification, selection and prioritization of biological drug targets, evaluating large amounts of transcriptomic, proteomic and clinical data in minutes, predicting the bioactivity and chemical properties of biological targets, detecting drug-target interactions for toxicity analyses and of course, revolutionary protein structure predictions, as seen with AlphaFold.^2,3

‍

This is by no means an exhaustive list of AI applications in drug discovery but reflects the rapid translational activities in the industry from POCs to operations. Even with the intertwinement of the two sectors still relatively in infancy, the AI enabled drug development has been bolstered by multiple high-profile successes. UK-based Exscientia currently has four AI-designed candidates in the clinic⁴, Biologics firm AbSci was the first “to create and validate de novo antibodies in silico” (i.e. through computer simulation) using generative AI⁵, and Insilico medicine received the first ever Orphan Drug Designation from the FDA for an AI-discovered and designed candidate, along with receiving FDA approval for its MAT2A small molecule inhibitor IND application just two weeks ago⁶. Most recently, researchers at UC San Diego revealed a new ML algorithm, POLYGON, capable of simulating complex chemistry and able to identify candidates rapidly, and chiefly focusing on multi-target molecules, thus making for a novel way to reduce adverse effects typically seen with traditional combination therapies. With this new platform, scientists were able to synthesise 32 potential multi-target therapeutics.⁷

‍

Naturally, with the field receiving this degree of traction and publicity, and forming such a key part of current company strategy rhetoric, it’s important to understand what AI can and cannot do for a Biotech firm right now - and the key practical takeaways for integrating it into drug discovery processes. As Aaron Daugherty points out, AI cannot actually “solve the real problem, which is the complexity of human disease. AI, at its best, processes data with unthinkable accuracy and speed at a massive scale”⁸ whilst Rahul Das, VP of AI and Life Science Solutions at Norstella points out to Pharma Voicethat “Just because something has been designed by AI doesn’t mean it’s going to take away all the uncertainties”⁹ and Matthew Featherstone, vice president, and Rowan Saada discuss whether AI will ever “completely re-write the playbook” of drug discovery and conclude that for now, the most extractable benefits are seen in its ability to expand drug discovery’s scope and accelerate its timelines. At a minimum, the cost and timeline reductions associated with drug discovery and pre-clinical are significant - Insilico Medicine identified its INDS018-055 candidate, a small molecule inhibitor, with AI platforms PandaOmics and Chemistry42 in just 18 months at the cost of around $1.8M - a radical reduction from the typical 3 - 6 year timeline for preclinical candidate selection. Exscienta designed several drug candidates which reached the clinic in under 12 months and UK-based AI Biotech BenevolentAI’s BEN-8744 candidate for ulcerative colitis reached pre-clinical studies after just two years in discovery. With a conservative outlook, AI is, at a minimum, “expanding its [drug discovery’s] scope and accelerating its progression” at an unprecedented rate with BCG and the Wellcome Trust, jointling stating that the time and cost reductions associated in discovery - preclinical development could be at least 25–50%.¹⁰  BCG also undertook a study analyzing 20 Pharma firms between 2010 - 2021 all of which were ahead of the curve in terms of exploiting AI applications. Out of the 15 drug candidates that reached the clinic during this period, BCG honed in on eight of them and discovered that even with Drug Discovery AI still in the beginning stages, five had reached the clinic in below-average time.¹¹

‍

However, capital and timeline reductions alone are likely to be a gross underestimation of AI’s potential to transform drug discovery, with strong early evidence suggesting that we should look forward to clinical successes, as demonstrated by Insilico, Benevolent AI and others. Aaron Daugherty points out AI’s particular potential to disrupt drug discovery’s traditionally reductionist nature, with its ability to analyze potential targets against significantly more multimodal data at once to paint a vastly more complete picture of how a disease will be affected by a particular treatment, and thus increasing chances of clinical success. He points out however, that these potential gains are highly dependent on a successful, symbiotic integration in drug discovery teams. A key factor behind some early failures or disappointment has come from too much onerous being placed on AI’s potential to identify drugs faster and more accurately, “without acknowledging that high quality lab and clinical science is a prerequisite for ultimate success”.⁸ The years of experience honed by scientists needs to be complimented and “supercharged” by AI tools. This integration should be two-pronged, with experienced data scientists and computational biologists having a seat at the table alongside drug discovery researchers and scientists, and with the latter also receiving training and upskilling on their utilisation of AI.¹² Whether internally or through external courses/by external educational bodies. As Dr Gerard van Westen, Professor of AI and Medicinal Chemistry at Leiden University points out, “Artificial intelligence won’t replace scientists” but to still be truly effective in their roles in the future, they need to become cognizant and highly aware of how to collaborate with AI to enhance their capabilities.¹³

‍

A key factor as to why scientists and AI will need to collaborate in harmony points to the second key takeaway to consider for a truly successful AI and drug discovery integration - dataset accuracy and effective translation from discovery data, notations and findings to ML problems. Naturally, AI-generated predictions can only be as accurate as the datasets which under-pin them. In discovery, this can pose a challenge as benchmark datasets can be low-quality and poorly annotated, resulting in AI identifying non-viable drug candidates.¹⁰ Further, poor data labelling cohesion and a lack of standardised knowledge representations can be a major root cause behind inaccurate AI-generated predictions - for example, drugs can be represented by Simplified Molecular Input Line Entry System strings (SMILES) (a notation describing the chemical make-up of a structure in a line of text, including its atoms, bonds etc.), Extended Connectivity Fingerprints (ECFPs), (another cheminformatics notation, capturing a structure’s atom neighbourhood and hashing algorithm), and in Molecular Graphs (whereby molecules are depicted where atoms are nodes and bonds are edges on a graph). Equally, with proteins, high quality experimental data is the single greatest factor behind accurate AI-generated 3D structures¹⁰ and without that, AI has little efficacy in protein analysis. Proteins can also be represented in numerous ways, from 1D amino acids to protein sequence representations, as well as 3D-structures, thus also posing a problem for consistent, accurate AI generations. A major takeaway therefore is for organisations to align on representations and on generating high-quality, cohesively and comprehensively annotated and labelled datasets. Standardisation of benchmarking datasets, such as ImageNet, a vast visual database to underpin visual object recognition software, developing an “ImageNet”-like design for molecules could be a possible way forward. Another more standardized benchmark includes Stanford University’s “MoleculeNet” devised to test ML methods for properties of molecules by curating several dataset collections and developing a software suite including a variety of features and algorithms that have already been used in the field. Researchers have also theorized that over-fitted ML models (i.e. models which perform well on training data but poorly on unseen data due to their overly complex nature tailored to specific dataset quirks) can still have utility by potentially being able to generate novel data-driven hypotheses by detecting subtle patterns or anomalies in the data that more generalized models overlook, which could suggest novel biological mechanisms or relationships that haven't been considered before.² The importance of collaboration between scientists and AI is also to be re-emphasized here as these would need to be validated by biologists. Of course, organisations must consider their organizational structures, change management and importantly, the state and AI integration potential of their current tech stack, but starting with the right team and accurate data sets prime for a successful adoption of AI into discovery.

‍

Overall, the excitement and widespread hype around AI in drug discovery recently reflected in Xaira’s $1B raise and world-class team does feel justified, with AI offering immediate, game-changing time and capital-reducing benefits. For its potential to be truly recognized in the clinic with AI-identified and designed drug candidates showing efficacy and safety, there are two main starting points for organisations to focus on- the need for educating personnel and constructing a team which can effectively execute and then exploit this integration, along with careful consideration of the datasets underpinning these models. 

‍

References 

‍

(1) New AI drug discovery powerhouse Xaira rises with $1B in funding. Fierece Biotech. https://www.fiercebiotech.com/biotech/new-ai-drug-discovery-powerhouse-xaira-rises-1b-funding (accessed 2024-05-30).

(2) Qureshi, R.; Irfan, M.; Gondal, T. M.; Khan, S.; Wu, J.; Hadi, M. U.; Heymach, J.; Le, X.; Yan, H.; Alam, T. AI in Drug Discovery and Its Clinical Relevance. Heliyon 2023, 9, e17575.

(3) Niazi, S. K. The Coming of Age of AI/ML in Drug Discovery, Development, Clinical Testing, and Manufacturing: The FDA Perspectives. Drug Des. Devel. Ther. 2023, 17, 2691–2725.

(4) Exscientia Pipeline. https://www.exscientia.com.

(5) How Artificial Intelligence is Revolutionizing Drug Discovery. Bill of Health. https://blog.petrieflom.law.harvard.edu/2023/03/20/how-artificial-intelligence-is-revolutionizing-drug-discovery/. 

(6) Insilico Medicine’s AI-designed drug ISM3412 receives FDA IND approval. News-Medical. https://www.news-medical.net/news/20240425/Insilico-Medicines-AI-designed-drug-ISM3412-receives-FDA-IND-approval.aspx.

(7) AI Transforms Drug Discovery With Faster, Safer Cancer Treatments. SciTechDaily. https://scitechdaily.com/ai-transforms-drug-discovery-with-faster-safer-cancer-treatments/.

(8) AI: a great crash of hype into reality. Drug Target Review. https://www.drugtargetreview.com/article/108086/artificial-intelligence-ai-a-great-crash-of-hype-into-reality/. 

(9) Believe the hype? Mixed signals from AI’s impact on drug development. PharmaVoice. https://www.pharmavoice.com/news/artificial-intelligence-hype-ai-drug-development/704153/.

(10) AI In Drug Discovery And Development — Will It Live Up To The Hype? Drug Discovery Online. https://www.drugdiscoveryonline.com/doc/ai-in-drug-discovery-and-development-will-it-live-up-to-the-hype-0001.

(11) AI’s Potential to Accelerate Drug Discovery Needs a Reality Check. Nature 2023, 622, 217–217.  

(12) Generative AI in the pharmaceutical industry. McKinsey. https://www.mckinsey.com/industries/life-sciences/our-insights/generative-ai-in-the-pharmaceutical-industry-moving-from-hype-to-reality.

(13) Artificial intelligence as the co-pilot for drug discovery. Leiden University. https://www.universiteitleiden.nl/en/news/2023/05/artificial-intelligence-as-the-co-pilot-for-drug-discovery.