17 January 2022, Cambridge – Artificial Intelligence (AI) is paving the way for speech and image recognition and the launch of AlphaFold DB has revolutionised protein structure prediction. However, the technology seems to be lagging when it comes to other areas, including translation into drug discovery. Despite a huge amount of media attention for its potential to accelerate this field, AI is yet to be proven as an effective solution. What needs to change for AI to advance drug discovery?

By Isidro Cortes-Ciriano, Research Group Leader at EMBL-EBI and Andreas Bender, Reader for Molecular Informatics at the University of Cambridge

AI could make the strongest impact on drug discovery by reducing the number of drugs failing in clinical trials. Currently AI is largely focused on method development using preclinical data – data from research that takes place before human clinical trials – rather than focusing on applying and generating the clinical data we need to make a real impact on drug discovery. Here we outline where in the drug discovery process AI is working well, how we can improve its use and what needs to change for AI to start to benefit future drug discovery pipelines.

AI – from ligand discovery to drug discovery

One of the most successful examples of using AI in the bigger context of drug discovery focuses on ligand discovery in preclinical settings – identifying molecules that bind to a biological target of interest such as a protein. For this, AI has been effective in helping to validate a target that may be suitable to treat a disease.

However, predicting an appropriate ligand does not guarantee a successful drug, that is, an efficacious and safe drug in vivo. Our current methods, and in particular the data they are based on, need to be developed further if drug discovery is to truly benefit from using AI. 

Understand the biology

Complex biology is context-dependent, which makes it difficult to understand, and it only gets harder when you try to model changes spatially and over time, for example in gene expression. This is the weak point of current computational drug discovery and poses big problems when you add AI to the mix. 

We need to both be able to model differences in patient biology, so to take complex data into account; but at the same time we need to be able to generalise, so to find features which agree among patients which can be used for decision making in drug discovery. Therefore, to improve drug discovery we need to focus on how AI can help given the data we already have and by generating data that would help us leverage the full potential of AI for drug discovery.

More clinically-relevant data

To validate AI in the downstream clinical stages of drug discovery, we need to move to more complex biological systems and eventually apply these methods to real life scenarios in the clinic. At the computational level, this means including more predictive endpoints in the models we use, related to both efficacy – a drug's ability to produce the desired result – and safety.

We have run computational simulations to determine the effect of reducing the cost, increasing the speed, and improving the quality of decisions on the success rate of drug discovery. These show that improving the quality of compounds taken forward has the biggest impact on the overall success of a trial. This indicates that for AI to make a real impact on the success of drug discovery we should focus on improving the quality of decisions – that is, better selecting which compounds enter clinical trials – rather than simply focusing on increasing the speed or reducing the cost at which we fail.

The future of AI for drug discovery

The data we currently have doesn’t always allow us to make a judgement on the efficacy and safety of a potential new drug, for example using data derived from preclinical vs clinical data. When these types of data, which are predictive for the in vivo situation from both the efficacy and safety angle, are available on a sufficient scale, they can be used by AI in the drug discovery decision-making process. Once we reach this stage, AI in drug discovery could be elevated to a whole new level.

Improved uses of AI in drug discovery will come from a more integrated understanding of the complex links between genes, proteins and disease. Once we are generating and using data that truly reflect the biological aspects of drug discovery and relate these to relevant in vivo endpoints, we can apply AI methods to make real progress in the field. Improving our understanding of efficacy and safety within drug discovery and applying AI to these data could then have a profound impact on reducing the number of drugs failing to get through clinical trials.

Edit