AI & Pharma: curing the productivity slump
A cohort of exciting startups are attempting to revive a sluggish industry
Welcome back to HealthTrends. Know someone who loves hearing the latest health technology trends? Share HealthTrends with them now!
š¤·āāļø Problem
The problem with drug discovery is multi-faceted:
The low hanging fruit has gone. The number of new drug filings with the FDA has been decreasing since the 90ās.
The cost to develop new drugs is increasing. $2.6 billion is the estimated development cost of each FDA-approved drug. R&D investment in 2014 tripled that of 1995, and doubled the amount spent in 2000.
The time it takes to bring a new drug to market is increasing. The current average from discovery to launch is 10-12 years.
The success rate is low. The probability of a drug candidate successfully completing clinical testing is around 12%.
š” Solution
Use AI to:
Accelerate drug discovery and reduce the time required to get to clinical trials
Improve predictions about the safety and effectiveness of drugs so less fail in clinical trials
š Terms
New molecular entities (NMEs) that receive regulatory approval are one of two main types:
Small molecules - small, simple molecules (like aspirin) produced through chemical reactions. They normally act within cells to provide their therapeutic effect.
Biologics - Large, complex structures (like antibodies) that are manufactured from specially engineered cells. They normally exert their effect in the bloodstream or on the surface of target cells.
Although the majority of new drug approvals are for small molecules, the share of the total pharmaceutical market held by biologics has increased significantly over the last decade.
š History
The use of AI in pharmaceuticals isnāt new. Back in 2009, researchers at Aberystwyth University and the University of Cambridge created āAdamā - a computer laboratory combo that could hypothesise about the function of certain yeast genes, and run experiments to test those hypotheses. An AI researcher from Stanford had this to say about Adam at the time:
Adam is operating only at the level of a graduate student. Still, the robot is moving closer to the goal of an artificially intelligent machine that can cooperate with other scientists and write up their results in natural language, he says. āThatās probably far off, but it seems likely that we will get there. This is yet another step on the way. - Will Bridewell
The same team went on to develop āEveā at the University of Manchester - which had the purpose of speeding up drug discovery for tropical diseases like malaria and Chagasā disease. Unveiling their creation in 2015, it was perhaps the first example of automating early stage drug design without just using the brute force of conventional high throughput screening. Eve was able to learn during the screening process to improve the āhit-rateā in future experiments.
š¼ Use cases
Rare disease treatments
Identifying more promising drug candidates
Raising the percentage of candidates that make it through clinical trials and gain regulatory approval
Speeding up discovery to licensing time
Side effect prediction
Drug repurposing
Identifying drug efficacy based on patient genetics
(plus many more..)
š„ Players
Benevolent use AI to generate a knowledge graph to help with drug target identification and molecule design
Atomwise use convolutional neural networks to gain information about protein structures and binding properties to help with lead optimisation
Standigm has developed two products - āInsightā and āBestā - to improve drug target prediction and novel compound creation respectively
CytoReason computationally model cell environments in order to help stimulate drug discovery
Genome Biologics use AI to find RNA targeted therapies
Bullfrog AI focus on drugs in late stage clinical trials in order to predict which patients will respond best to specific candidates
Deepcure use AI to facilitate drug discovery, lead optimisation and portfolio diversification mainly focussed on the pre-clinical stage
Exscientia focus on drug target selection using AI techniques
Cyclica take a holistic approach, using AI to reduce off-target drug effects and provide a wider understanding of a moleculeās activity
š Impact
Bristol Myers Squibb have used used machine learning to better predict which drugs might inhibit the enzyme CYP450 - which has an important role in ensuring drugs donāt reach toxic levels in the bloodstream. This allows them to drop problematic drugs early on before they cause (expensive) headaches later on.
Insilico Medicine, a startup, combined multiple AI approaches to discover molecules that could inhibit a receptor involved in fibrosis and take a promising candidate into animals studiesā¦ All in only 46 days. (Whether this drug successfully traverses the depths of clinical studies remains to be seen.)
š® Predictions
Successful filings. Multiple drug candidates identified through AI technologies are now making their way through the necessary regulatory stages required to achieve a successful filing with the FDA or MHRA. It might not be long before we have the first child of the AI drug revolution.
Mergers & Acquisitions. If some of the startups in this space start shipping successful compounds, that could trigger a spate of acquisitions as the larger pharma companies move away from just inking partnerships and start acquiring the more promising AI companies.
Precision medicine. Companies like Berg are stress testing biological samples from both diseased and healthy patients to generate a whole array of genomic, proteomic, metabolomic data (all the -omics basically), which can be used not just to help identify promising proteins to target but can ultimately determine which patients down the line would benefit from such a drug.
Cloud migration. AI-enabled drug discovery is data hungry and computing power is key to running AI models and simulations in reasonable time frames. More and more companies will partner with the big cloud providers including Amazon, Microsoft and Google in order to acquire the computational resources they need.
š Opportunities
Rare disease treatments. Previously neglected conditions could see renewed focus by companies which believe they have can find promising drugs without breaking the bank. (See Sensyne Healthās recent collaboration with Bristol Myers Squibb.)
Collaborations. Small startups can leverage the deep pockets and resources of larger institutions to help bring novel drugs to the market. Thereās already plenty of activity in this space with 20 team-ups so farā¦
One molecule, multiple targets. Complex conditions like Alzheimers disease (AD) are the result of complex physiological processes where multiple proteins need to be targeted in order to achieve a significant clinical effect. AI methods might facilitate the discovery of molecules that can have a therapeutic effect on multiple targets, effectively killing two or more birds with one stone.
Small data problems. Surprisingly, the structure and properties of about 90% of human proteins are still not known which makes it difficult to determine if a possible drug candidate will work. Companies like Exscientia are trying to solve this small data problem by using their database of billions of protein interactions to identify where the knowledge gaps are. They can then hunt for the necessary tissue samples to help fill those gaps.
š Links
This Nature article nicely summarises the way AI can help the drug discovery slump
This Deloitte insight paper is extensive, well researched and covers the history and future of AI in drug discovery
Donāt get HealthTrends FOMO - catch up on previous posts here
Thanks for reading. Help HealthTrends reach a wider audience by sharing it with a friend or colleague.
ātill next time,
Sandy š