Researchers developing drugs must navigate the vast complexity of human biology - predicting how a small molecule will behave among thousands of proteins, metabolic pathways, and cellular interactions. Expert academic understanding can provide the idea for a drug, and skilled translational scientists bring it through development, but still it there are multiple pain points.

We envision AI as assisting with the following pain points:

Any AI tool will need to show a meaningful improvement along these metrics, particularly success rate, time, and cost.

When we look at what current AI-driven biotechs are focusing on, they are largely geared towards optimising in-vitro preclinical work to create better candidates, including finding better targets or enabling novel modalities.

There are fewer companies working directly on in-vivo work or clinical trial optimisation, even though these are areas with the highest failure rates and lack of understanding. However, increased complexity and poorer data quality and quantity inhibit the development of AI tools for these stages of development.

Stage	AI Solutions	Limitations	% of companies
Target identification	- Literature data mining - Integration and analysis of –omics datasets - Disease modelling, ‘virtual organs’	- Quality and quantity of dataset contributing to understanding of disease can vary	15%
Target validation	- Dynamic modelling of binding pockets, pathway interactions - Identification of biomarkers	- Must be validated by wet lab results
Lead discovery	- Rapidly screen wide chemical space - Dynamic modelling of interactions, kinetics, etc - Generation of un-intuitive options	- Best supported by high-throughput screening - Non-static molecular dynamics remains difficult for AI to model, but is crucial to understanding drug binding	80%
Lead optimisation	- Optimise within parameters of toxicity, selectivity, bioavailability, etc	- Best supported by high-throughput screening
Preclinical testing	- Virtual animal models	- In silico translatability to animal models is unproven - In vivo timescales limit quantity of data that can be acquired to train ML	3%
Clinical trial twins	- Integrate and analyse more patient data; subpopulations - Digital twinning: control group is digital version of treated patient	- Long regulatory path to digital twins	2%

For this and following data, we analysed 100+ AI drug development companies in Europe and the US, which were representative of the overall distribution of clinical stage and funding status. We excluded SaaS companies whose primary focus was to sell a drug discovery solution.

Considering the dual nature of good tech and good biology, we consider that “good” looks like the following:

Biotech

• Therapeutic area: Subject to same expectations as non-AI biotechs: disease of focus is relevant; market size established; target is validated; modality suited to disease, etc
• Application of ML: ML aims to solve a problem other tools cannot; whether a problem with drug development generally or challenges of a specific therapeutic area

Artificial Intelligence

• Compute: must secure appropriate processing power and critically evaluate how complex a model is required to be fit for purpose
• Data: ML algorithms are general and not protectable by IP. Good data is the rate-limiting factor and should be actively generated, acquired, and protected
• Explainability: the closer to a patient an application of AI is, the greater the desire that the model be explainable; less likely to be a problem in target ID or lead optimisation

Other

• Team: Talent in both biotech and ML; capable of simultaneously building tech infrastructure and asset pipeline. Best TechBios have a “hybrid” culture across both
• Data protection: AI faces regulatory pressure over source/protection of data; however life sciences field already well versed in data protection, so unlikely to be significant problem

 

An ML model is only as good as the data it’s trained on. It must be…

Proprietary data thus becomes a key value driver, and we see this reflected in the valuation a company receives.

 

Given the platform nature of AI-driven biotechs, they tend to have more candidates than resources to bring them all to the clinic. Potential models include:

The internal pipeline model allows the company to capture full revenue potential. However, this requires all expertise to be in-house and has long timelines to revenues; in an environment that is increasingly skeptical and wants early proof-points from TechBios, this may present difficulties.

The partnership model addresses these concerns, as pharma can pay out milestones periodically, and also brings strong clinical expertise. However, there may be difficulties managing the additional team, and revenue potential is lower. Companies may also find themselves more at the whim of what pharma wants, and not be able to develop core capabilities. Nonetheless, partnerships bring additional value and contribute to exits.

In our review, we found the indications pursued by TechBios largely matched the general biotech landscape: areas like oncology and neurology are busy, with rare disease becoming attractive if a company has a particular solution.

For the purpose of training a model, busy areas like oncology have high data quality and quantity, well validated models of disease, and high pharma interest. On the other hand, areas like neurology are difficult to address but could benefit from targeted use of models with high rewards if successful. Where an indication is difficult, a differentiating factor for ML may be ability to generate unique insights from data.