AI-Assisted Drug Discovery

The enormous potential and possible impact of artificial intelligence (AI) applications have been widely recognised by the public since the release of “ChatGPT”. Right after AlphaGo’s victory over the South Korean professional Go player Lee Sedol[1], the company “DeepMind” stepped up its AI work on protein sequences and structures and with “AlphaFold” made great progress in predicting the three-dimensional structure of proteins from their amino acid sequences. The structure of many proteins can now be predicted quickly and with high accuracy, although the method is not equally successful for all proteins and structural states.[2] This is a major validation of AI’s potential in biology.

AlphaFold was recognised as a solution to protein folding by CASP, a community in which researchers share their progress in predictions against real experimental data. Every two years CASP holds a competition in which research groups test the accuracy of their predictions against real experimental data. Teams are given amino acid sequences for proteins whose 3D shapes have been mapped but not yet made public. The teams then submit their best predictions to see how close they come to the structures later revealed. At CASP13 (in 2018), AlphaFold came first.

At CASP14 (in 2020), the latest version of AlphaFold showed such a high level of accuracy that the community considered the problem of predicting the three-dimensional structure of single protein chains largely solved. Since then the AlphaFold methods paper has received more than 10,000 citations, placing it among the 100 most cited papers of the past decade and the 900 most cited papers of all time.[3] This methodological breakthrough has led dozens of other scientists to use the method to determine the shapes of other important proteins.[4] (Update: In 2024 Demis Hassabis and John Jumper of DeepMind shared the Nobel Prize in Chemistry with David Baker for their work on protein structure.)

Proteins are made up of amino acids joined together. The number of amino acids that come together and the chemical bonds linking them determine a protein’s primary structure. The amino acid sequence in this primary structure is determined by the gene encoding the protein. Genes contain the information for protein structure and are not static but dynamic in evolutionary terms; in other words, they can mutate. All genes have the potential to mutate. These mutations can affect the primary structure of the proteins they encode and, in turn, change their secondary and tertiary structures. Such changes may prevent some proteins from performing their function. Some proteins that contribute to basic functions in the cell may also be affected by such changes. Another group of proteins, by contrast, can continue to function despite the changes. Amino acid changes resulting from mutations or gene alterations can, however, affect protein function positively or negatively. They can, for example, lead to susceptibility or resistance to disease.

Accurately predicting protein structures is important in drug and vaccine development. Changes in candidate vaccine proteins against disease-causing viruses can prevent the human immune system from responding effectively to these proteins. Interaction between proteins takes place via specific amino acid motifs. Disruption of these motifs or a change in the three-dimensional structure can disrupt interactions between proteins. So if we can predict protein structures in advance, we can strongly predict protein–protein interactions. This in turn allows us to develop more effective drug and vaccine candidates. [5]

Determining protein structure by computational methods can, besides answering the most difficult questions in molecular biology, contribute significantly to improving and accelerating the drug development process.

Besides predicting the overall suitability of a target, AI can help, within the framework of structure-based drug design, to identify drug candidates that, because of their spatial structure, should show a particularly high binding affinity for the target and thus promise an optimal effect. In this way, the long drug development process, often ending in failure, can be made more efficient in time and cost.

In future, even broader potential uses of AI applications in drug research will emerge. “RoseTTAFold”, developed at the University of Washington, and “ESMFold”, recently released by Meta AI, offer alternatives for structure prediction, each with its own strengths and weaknesses. “ProteinMPNN” and “ProtGPT”, meanwhile, have been developed to design customised, purpose-optimised proteins from scratch by generating the corresponding amino acid sequences.[6]

A study by Assoc. Prof. Tunca Doğan and colleagues reported that, as part of efforts to develop new personalised treatment options in healthcare, such AI models are intended to be used in the molecular analysis of genetic diseases and in the discovery of new drugs to target them specifically.[7]

Such AI applications can be used to investigate protein dysfunctions and regulatory disorders responsible for immune diseases, metabolic disorders, neurological disorders and other pathologies. They can be used to elucidate the three-dimensional structure of viral capsid proteins and to design specific antibodies that could successfully neutralise diseases such as Ebola, smallpox or SARS.

The importance of innovative, high-throughput and accelerated drug development became apparent as never before during the coronavirus pandemic. The progressive use of AI applications could in future make it possible to respond faster and more effectively to pandemics, to counter the growing health threat posed by some multi-resistant diseases and, of course, to develop treatment approaches for previously untreatable diseases.[8]

From the perspective of military medicine, and especially given the risk of infection in regions with different climatic and environmental conditions, the use of similar AI applications will help to protect health and maintain operational readiness.

In future such systems could lead to revolutionary progress in the fight against disease; but if misused, they could also lead to the development of unprecedented toxic substances.

Risks of Misuse

AI does, however, carry a significant risk of misuse: in 2022 researchers showed that an AI application designed to identify new active substances could be used to find compounds that are as harmful as possible and to optimise them for toxicity with relative ease. According to the researchers, the repurposed AI generated some 40,000 toxic molecules within six hours, including many previously unknown compounds.[9]

Measures taken to prevent such potential misuse are also of critical importance for national security. The development of AI-based technologies concerns not only science and health but many aspects affecting national security. In particular, the misuse of these technologies could lead to potential threats such as the theft or manipulation of countries’ strategically important information or exposure to cyberattacks.

For AI-based technologies to be used effectively from a national security perspective, it is therefore important to set ethical standards and establish regulatory requirements. National defence strategies should be integrated by defining the areas where AI applications may be used, in order to minimise potential risks and strengthen the possible advantages of these technologies. Researchers and drug developers involved in developing these technologies should also have guidelines that steer them in line with national security policies and objectives.

In conclusion, understanding the potential risks and opportunities of AI-based technologies from a national security perspective is a critical step towards offering a perspective that guides their development. This conscious approach is necessary both to protect national security and to make the best use of advances in AI.