This article was first published on The Pharma Letter. Click here to view the original article.
Understanding the AI-ENABLED DRUG DISCOVERY LANDSCAPE
AI-enabled Drug Discovery companies (“AIDDs”) have taken the bioscience world by storm, raising over $5.2 billion from investors in 2021, while 15 AI-enabled candidate drugs are in clinical trials, plus over 150 at the preclinical stage[1]. With hundreds of AIDDs worldwide, ranging from listed companies to brand new startups, how does one make sense of them? This article suggests some dimensions to help potential pharma collaborators or prospective investors differentiate AIDDs.
AIDDs can first be distinguished by their drug discovery role. “Treatment Concept Generators” determine where and how in the body’s complex system to intervene. Whereas “Intervention Engineers” design or improve the physical interventions. Some AIDDs do both in combination. Secondly, AIDDs adopt different Commercial Models.
Treatment Concept Generators
AI-enabled concept generators aim to ideate and validate new treatment paradigms, more efficiently and reliably than the traditional, haphazard process of manually trawling academic literature. They usually have a biological focus on certain diseases or biological processes.
Biological Focus | Example AIDD Company |
Aging | |
Cancer tumor cell dynamics | |
Immune system modulation | |
Liver disease and regeneration | |
Neurodegenerative processes | |
Metabolic disorder endotypes | |
Rare genetic diseases | |
RNA splicing errors |
Successful AI applications rely on good data. Concept generators are distinguished by their data source. Many start with public datasets, typically from academia. Such data may be incomplete or irreproducible, and given the diverse technologies and protocols for generating omics data in particular, challenging to compare across datasets. Savvy concept generators rely on proprietary data for their core insights. Some like Immunai and Insitro generate high throughput data at scale using automated cell or tissue experiments. A few use ex vivo organ perfusion like Ochre Bio. Others like BioAge and MultiOmic Health opt for patient-sourced samples coupled with longitudinal clinical phenotyping.
“Omics bandwidth” is an important consideration. Many concept generators use only genomics. An increasing number use genomics supplemented by one other omic. To address complex multi-factorial diseases, some companies perform integrative multi-omics analyses using three or more omics.
Biological focus is often the first criteria when considering a concept generator since pharma collaborators and life science investors tend to have strong disease preferences. The relative importance of data source and omics bandwidth reflects the underlying nature of the biological focus.
Concept generators can differ by opportunity type. Some identify novel drug targets requiring new molecular entities (“NMEs”) to modulate. Others identify repositioning opportunities for existing drug molecules. And some companies are agnostic since their platform creates both types. Certain investors prefer the nearer-term value creation afforded by repositioning, whereas others prefer the much higher risk-return profile of novel targets and NMEs.
Experienced pharma collaborators and investors are wary of insights derived purely in-silico. They usually prefer to work with those concept generators who deploy wet lab experimentation to validate and sharpen their AI-derived insights.
Intervention Engineers
Intervention engineers, specialized by drug modality, create molecules to outperform existing drugs or to modulate novel targets previously regarded as “undruggable”. Traditional modalities are categorized by molecular weight – small molecules, peptides, biologics – each requiring different skill sets and technologies. There are many niches within modalities; for example Cyclica is strong in small molecule polypharmacology for modulating multiple targets with a single drug.
Physical | Molecular Weight Range (Da) | Marketed Examples discovered traditionally | Delivery Format | Example AIDDs working |
Small molecule | 100s | Losartan (423 Da) for hypertension | Oral | |
Peptide | 1,000s | Semaglutide (4,113 Da) for dysglycemia and obesity | Mostly | |
Biologic | 10,000s to 100,000s | Evolocumab (141,800 Da) for dyslipidemia | Injectable | |
Oligonucleotide | 10,000s to 100,000s | Inclisiran (17,285 Da) for dyslipidemia | Injectable | |
Cell & Gene Therapy | N/A | Tisagenlecleucel for lymphoma | Injectable |
Within each category, they may specialize by molecule engineering stage. Some create the initial starter molecules (“lead generation”) whereas others refine those starters into drug candidates that balance efficacy, selectivity, pharmacokinetics and safety (“lead optimization”). In small molecules, Exscientia initially majored on lead optimization and Atomwise on lead generation, although they now do both.
Traditional drugs bind physically to target proteins. In contrast, oligonucleotides have emerged as an exciting new modality. Coming in two main flavors (antisense, RNAi), they are short pieces of genetic material that modify how the body translates genetic instructions to make proteins. Another fast-developing modality is cell & gene therapy, altering cell behaviors to treat disease, for example, training patients’ immune cells to attack their cancers.
Investors evaluate an intervention engineer based on its competitive advantages vis-à-vis the drug target types best suited to its technology, and the market attractiveness of the diseases where those target types are most relevant. Whereas pharma collaborators will look at whether it has made (or can make) drug molecules that match their R&D requirements.
Commercial Models
AIDDs could operate as service providers, earning output/time-based fees to support pharma customers’ R&D programs, with no claim on the resulting economic value. Many venture investors dislike service provision since 10x-100x returns seem unlikely. However, an attractive investment return might be achievable with low capital requirements, high customer lifetime value and a business model that outsources a substantial chunk of the customer’s project budgets.
Alternatively, AIDDs could pursue value sharing partnerships, earning upfront and contingent milestone payments, plus sales royalties on the marketed drugs. They justify their economic share by either: (i) creating novel research programs for exclusive licensing to pharma collaborators that subsequently drive clinical development and marketing, and/or (ii) substantially improving the outcomes of their collaborators’ R&D programs.
AIDDs might adopt a biotech model, developing wholly-owned drug assets that they take into Phase I and II clinical trials. On achieving clinical proof-of-concept, each asset is typically co-developed with an established pharma for the big ticket Phase III trials and subsequent worldwide co-marketing.
Each successive option above reflects an increasing risk-return tradeoff. A pragmatic approach is probably a hybrid model. Few AIDDs initially have sufficient capital to start with a biotech model. At the outset, they sell fee-for-service projects, prioritizing those pharmas where they believe value sharing partnerships could subsequently emerge. With sufficient funds from partnership revenues and investors, they evolve to a mixed portfolio of value sharing and wholly-owned projects, eventually eschewing service projects entirely. The hybrid model is a well-trodden path, as seen in earlier pharma technological shifts such as antibody engineering.
[1] Boston Consulting Group (2022) “Adopting AI in Drug Discovery”