Role of AI in early drug discovery

On average, it costs $2.8 billion and takes 10-15 years for a drug to enter the market. The significant cost can be attributed to the failure of most drugs at some stage in the drug development process; therefore, it is imperative to identify ineffective compounds early on in this journey. To reduce the cost and time to market, Artificial Intelligence models allow companies to automate some of the iterative processes, thereby discovering medicines rapidly.

The Role of AI in Early Drug Discovery

During early drug discovery, some thousands to millions of compounds are screened and huge volumes of data are produced. Deep Learning models can be trained using these data sets, which will enable them to recognise patterns and make predictions.

Fig (1) Various phases in early drug discovery

Knowing the structure of the target molecule is essential for successful targeting by the drug molecule. AI-models can add value during structure-based drug discovery by predicting 3D protein structures through analysing the distance between the amino acids and corresponding angles of the peptide bonds. The accuracy in predicting the structure has been found to be more than 50%. ¹

The affinity with which a drug binds to its target plays a crucial role in the efficiency of the compound, and can be predicted using AI-based methods. It is assumed that similarities in the structures of drug and target will help to determine affinity and similar drugs will interact with the same targets. One of the machine learning techniques, called MANTRA, groups the compounds based on the predictions of their common biological pathway or mechanism of action. ¹  AI can also play a major role in target identification through accessing the vast libraries of publicly available biological data.

Sometimes, compounds can interact with unwanted targets, which can lead to toxicity. Advanced AI-models like DeepTox can predict toxicity based on the similarity of these compounds to the ones already present on the market. Open-source tools such as PrOCTOR considers many drug properties as well as targets to forecast whether a drug can fail in clinical trials due to its toxicity.

Another important application of AI is in the repurposing of drugs. Repurposing an existing drug will allow it to enter directly into phase II clinical trials, thereby reducing the cost of development significantly. Various machine learning approaches consider similarities between drugs, disease, or target molecules while repurposing a drug.

Many chemical databases like PubChem, ChemBank, DrugBank and ChemDB show the positional information about molecules within the space. This allows for virtual screening using various in silico (computational) methods to identify bioactive compounds. Some algorithms such as Coulomb matrices and molecular fingerprint recognition are used in drug design to identify a lead compound. ²  They consider the physical, chemical, and toxicological profiles of a compound, which results in selecting lead molecules cost-effectively.

Some computational models are combined with AI models to predict absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of drug molecules. Lipophilicity and solubility of the compounds can be predicted using neural network programs such as ADMET predictor and the ALGOPS program. This helps in the optimisation of lead molecules. Predicting in vivo activity and toxicity of molecules is important in identifying the pre-clinical candidate molecules. Algorithms, such as extreme learning machines and Deep Neural Networks (DNN) were developed for this purpose. ^{3, 4}

De novo drug design is also made possible through computer-aided synthesis planning programs like Synthia, which effectively proposes possible synthesising routes or alternate synthesising strategies for patented compounds. ⁵  The involvement of AI in the de novo design of molecules can be beneficial to the pharmaceutical sector, leading to swift lead design and development compared to traditional methods. De novo drug design has two models: the Generative Model that generates unique molecules and the Predictive model that evaluates the properties of the generated molecules.

Many leading pharmaceutical companies like Roche, Pfizer, Novartis and Astrazeneca have collaborated with, and continue to do so, with AI organisations in various fields including Oncology, Cardiovascular diseases, and CNS disorders. ⁶

Conclusion

AI is a revolution in the field of pharmaceutical drug discovery and development due to its rapid hit, lead compound identification and suggested synthesis routes for these molecules. It can predict the desired chemical structure of efficient molecules as well as link their structure to their activity. AI can also predict the interactions of a drug with its target, thereby increasing the chances of developing an efficient molecule. It is likely that the number of AI applications and their efficiencies will also increase in the future, making the process of drug discovery even quicker and even more cost-efficient.

References

1. Lounkine, E. et al. (2012) Large-scale prediction and testing of drug activity on side-effect targets. Nature 486, 361–367.
2. Chan, H.S. et al. (2019) Advancing drug discovery via artificial intelligence. Trends Pharmacol. Sci. 40 (8), 592–604.
3. A´lvarez-Machancoses, O´ and Ferna´ndez-Martı´nez, J.L. (2019) Using artificial intelligence methods to speed up drug discovery. Expert Opin. Drug Discovery 14, 769–777.
4. Dana, D. et al. (2018) Deep learning in drug discovery and medicine; scratching the surface. Molecules 23, 2384.
5. Grzybowski, B.A. et al. (2018) Chematica: a story of computer code that started to think like a chemist. Chem 4, 390–398.
6. Debleena Paul et al. (2021) Artificial Intelligence in drug discovery and development. Drug Discovery Today, 26(1): 80-93.