AI routines, drug discovery and the potential for IP

AI routines are powering the discovery of new treatments, even for conditions as severe as Stephen Hawking’s. Christoph Behrens and Lukas Bischoff at Meissner Bolte review AI's potential for creating IP, highlighting how ventures like Insilico are taking a lead

---

Share

Christoph Behrens and Lukas Bischoff

---

Today, the average person may still wonder how artificial intelligence may impact intellectual property, as the most publicly visible applications, such as ChatGTP, are mainly concerned with the processing of text. Such subject matter is not patentable, as it does not go beyond data processing and visualisation.

However, we are now seeing the emergence of ventures who are capturing the potential for developing and using AI in the generation of IP. One of the forerunners, Insilico Medicine, is using AI to speed up processes (or routines), which were conventionally driven by manual research and serendipity or, in practice, discoveries by unplanned strokes of fortune.

By using such AI routines, Insilico with its academic partners was able to start tackling the disease from which Stephen Hawking most visibly suffered, ALS (amyotrophic lateral sclerosis), which has always been hard to treat. Such identification of potential drugs with AI has so far led to the discovery of 28 new targets in a relatively limited time. One of them, FB1006, is now in clinical trials. Since patents, as an IP asset, are granted on a first-to-file basis, a reduction of the time to identify new drugs and targets is expected to give a significant advantage to companies who follow Insilico in its use of AI.

Where AI and IP have favourable overlap

What enabled Insilico to develop as such a strong player in drug development, when the company was founded only in 2014, and the field is notoriously governed by big pharma companies such as Pfizer, Merck or Novartis? As its initial starting point was in finding an alternative to animal testing for research and development of pharmaceuticals, its later development could probably not have been foreseen. However, since it was interested in the development of AI more generally, it worked its way into drug discovery itself, where the routines, which they had initially developed, can be applied.

Even so, going into Al for generating IP may seem counterintuitive at first sight. The set of rules or algorithms on which AI depends are excluded from most patent laws (including the European).

However, if AI can take over routine tasks such as information processing, which have conventionally been performed by humans, it can benefit the generation of IP. Al is especially suited to the analysis of vast sets of data, where humans falter in seeking the needle in the haystack. Insilico is a good example of how to establish a strong IP position by using AI as a tool to obtain knowledge on which a patent application can be later based.

By pairing AI with IP generation, each development stage is accelerated. Less time is taken to eliminate targets and more attention is given to the most promising candidates. The advantages of Al become apparent in assisting and accelerating the generation of IP, whose subject, a newly developed drug, can then benefit from longer protection to pay off initial investments.

How AI supports drug development

When thinking about implementing Al, a first step is to recall how conventional drug development usually proceeds. Here a protein and the function it performs in the human body are often the starting point for the scientist. Disease in many cases is an imbalance in the interplay of processes in the body. Treatment strategies are mostly targeted at re-establishing this balance, for example, by regulating the function of proteins which are overactive, in which case compounds to deactivate or block the protein are developed.

Such mechanisms rely on competitor molecules that have similarity to a natural substrate of the protein, but which bind more strongly and prevent the protein from performing its current function. Alternatively, the binding of compounds may initiate a small change in the three-dimensional structure of the protein, which decreases its ability to bind to a natural substrate.

For both mechanisms, the 3D structure makes it possible to perform computer-based docking experiments. Compounds are artificially bound to the protein and the strength between the two is assessed. In this manner, it is possible to screen large compound libraries for their possible binding capabilities.

The bottleneck in this process is finding the 3D protein structures. Conventionally, proteins are crystallised and analysed by x-ray diffraction. However, not all can be crystallised and alternatives are highly expensive. So 3D structures can be hard to establish. Until recently, x-ray structural data had only been available for a small fraction of proteins that represent possible targets for drug development.

AI for protein structure modelling

A pioneer AI routine now predicts the most likely structure of a protein from its amino acid sequence. Alphafold was introduced by DeepMind, a subsidiary of Alphabet, in 2018 and an improved version published in 2020. This programme has become the benchmark for the prediction of all 3D structures in the human genome and is now included in public databases, such as UniProt. For any protein, Alphafold compares its sequence with those that are already known and confirmed. Based on the probable interactions of amino acid residues, it makes a prediction of the most likely structure of the protein.

AI for identifying disease markers

An area at the interface of AI and drug development, where Insilico is active, is the identification of proteins and markers associated with diseases, for which pharmaceuticals or antibodies can be subsequently developed. To this end, Insilico has developed an AI system called PandaOmics, which can analyse data from biological samples on the entirety of genes, proteins or messenger RNA (also called omics data). Comparisons are made between respective datasets from patients, who suffer from a given disease, and healthy people. If expression levels of protein are only found in the patients, they can represent a possible target for the treatment of the disease.

Given the size of datasets that are generated for even a single biological sample, AI has a particularly high potential for this kind of analysis. PandaOmics has been integrated by Insilico into a tool called Pharma.ai. It goes beyond data comparison, searching scientific publications and patents for information relating to a protein of interest, helping to rule out false links between the identified proteins and the disease.

Such AI analysis also has potential to identify patient groups, which may respond to a given treatment, and distinguish them from other patients, who suffer from the same disease, but are not receptive to the treatment. Breast cancer is a notable example. With such information, patients can be prescribed the right medication, which can prevent unwanted progression of the disease due to ineffective treatment.

AI in compound optimisation

A further area, where AI has been used by Insilico, is optimisation of compounds, which have been identified as targets. Here, Insilico has developed a tool called Chemistry42, which can learn from the simulation of the binding of large numbers of compounds to a given protein. Variants of the best binding compounds are generated and tested. In this way, compounds with the highest potential are identified as candidates for clinical phase testing.

Over time, it is expected that such routines will improve, as data from real-life testing is fed back to eliminate inaccurate predictions. The benefits of machine learning apply in particular to the treatment of diseases where there is more insecurity about the true structure and interaction of a given protein with a potential drug.

The Chemistry42 tool is also limited to the prediction of the chemical structure, but it can also can suggest synthetic routes for the production of compounds. It can even be linked to automated devices that can then synthesise the AI-generated compounds.

AI for clinical trial simulation

Insilico is also applying AI to the predictive modelling of clinical trials. Its tool, InClinico, has been trained with information and data from drugs, which have successfully passed clinical trials, as well as those that failed. Its routines are specifically directed at identifying the causes of failure (for example, the molecular substructures of candidates that might give rise to toxicity issues or adverse interactions with other proteins). Early results for predicting the outcomes of clinical trials had a high rate of accuracy, close to 80 percent. As clinical trials are usually the most expensive part of bringing a new drug to a patient, Al routines represent a valuable tool to prioritise the trials of the most promising compounds.

Potential time gains: what is possible with AI

Recently, Insilico has shown what may be possible with AI in drug development. By using a combination of its PandaOmics and Chemistry42 tools, they were able to identify CDK20, as kinase, which is overexpressed in many tumour cell lines including colorectal cancer, lung cancer and ovarian cancer, as a potential target for the treatment of hepatocellular carcinoma (HCC, a form of liver cancer). By using the 3D structure of CDK20 as generated by AlphaFold and Chemistry42 for the generation of possible inhibitor compounds for CDK20, they were able to identify a hit compound with a binding constant of about 9 micrometres. This target was selected with the synthesis and testing of only a single digit number of actual compounds, highlighting what is possible with using respective AI routines. It is no surprise that big pharmaceutical companies are significantly increasing their AI abilities to avoid being left behind.

AI in drug discovery: future developments

For the future, with the respective routines that are now available, a trend towards an integration of all of these routines into one system can be expected, covering all steps of drug discovery, starting from the identification of protein targets for a given disease to the preparation and optimisation of candidates, and then the choice of the most promising compounds for clinical testing.

Does this mean that eventually Al will do all the work in drug discovery? Most probably, it will not be the case as most AI routines have been developed and are trained on the basis of existing knowledge and data, which might be biased so that there is a tendency of Al to provide new data with similar characteristics. However, it can be expected that with time the Al routines will become better and better, so it is likely there will be no way around using such routines in the future.

It is already clear that the use of AI can speed of discovery of new drugs, which brings a huge advantage in term of IP protection. In addition, if by using AI, companies are able to focus on compounds that are on average more likely to reach the market, they will spend less on drug that eventually fail in clinical trials, leaving more financial resources for further research.

AI is speeding up how knowledge is gained as a basis for IP. In pharmaceuticals, what matters is being faster than a competitor. To

• 'AI, drug discovery and the potential for AI', an article by Christoph Behrens and Lukas Bischoff, first appeared in Managing Intellectual Property Today, 2025 edition, published by Novaro, ISBN: 978-1-0685644-1-3. See here for further details.