AI’s Promising Role in Transforming Drug Discovery
Drug discovery is a notoriously complex and arduous process. David Pardoe, a computational chemist at Evotec, notes that, despite a century of modern medicine, treatments have only been found for around 500 of the approximately 7,000 rare diseases. This challenging process is both time-consuming and expensive. However, artificial intelligence (AI) offers exciting possibilities to address these issues.
AI has the potential to integrate the three-dimensional geometry and atomic structure of a potential drug molecule, visualizing its fit within its target protein. This allows for the modification of designs to enhance potency. Furthermore, AI can identify new targets for potential drugs. Another important aspect is the consideration of the biological environment within the patient’s body. This can help to avoid undesirable interactions with non-target proteins, decreasing side effects and increasing the effectiveness of the drug.
A robust foundation of high-quality data is essential for developing AI systems to enhance drug discovery. The pharmaceutical industry has a strong advantage over other sectors, with vast amounts of biological data constantly being generated worldwide.
Addressing the Challenges: Data Quality and Standardization
Although AI has the potential to transform the industry, the availability of data is not the only factor. The quality of the data is often a concern, as it is not always collected with machine learning in mind. Eric Durand, chief data science officer at Owkin, points out that inconsistencies in experimental methods pose a major problem for AI. Therefore, these discrepancies can be misinterpreted as being biologically significant by AI models.
Standardizing reporting and methods is an important aspect of improving data quality. For example, the Human Cell Atlas, a global project, has created reference maps of human cells in a standardized way and this offers consistent data for AI algorithms. Pat Walters, a computational chemist at Relay Therapeutics, notes that the lack of standardization leads to difficulties in comparing data sets.
Another initiative, called Polaris, is a benchmarking platform that aims to help with this data standardization by introducing guidelines for data set quality. A certification stamp is then awarded to sets that meet these standards.
The Value of Negative Results and Data Sharing
One challenge concerns the publication bias towards positive results in research. Miraz Rahman, a medicinal chemist at King’s College London, states that publishing only positive data may distort the view that AI systems have about drug discovery. This can be particularly crucial in instances where new drugs are being developed, and can lead to the AI suggesting ineffective compounds.
Sharing industry data and expertise is vital, as drug companies possess large amounts of data, including negative results. However, these data are typically kept confidential, as pharmaceutical companies are hesitant to share them with competitors.
Federated learning has been implemented in a collaborative project, the EU-funded project called Melloddy, that allows ten companies to train software without sharing sensitive data. Another opportunity for data sharing is to provide more funding to public databases, such as the UK Biobank.
Some researchers believe that better processing coupled with a large volume of data will help overcome the difficulties of utilizing AI for drug discovery. Alex Zhavoronkov, founder and CEO of Insilico Medicine, says that pharmaceutical companies should support projects such as the UK Biobank because they hold useful data. Insilico Medicine has created tools to evaluate the credibility of scientists by monitoring their previous publications and also tracking stock prices after clinical trial announcements.
Insilico has already achieved success using AI in drug discovery. In 2019, the company’s AI platform discovered a target for fibrotic diseases. It then employed its generative AI platform to find compounds that would block this target. Zhavoronkov states that Insilico has nominated 22 preclinical candidates since 2019.
To conclude, in order to capitalize on the transformative potential of AI, the pharmaceutical sector must prioritize data standardization, data sharing, and the incorporation of both positive and negative findings. By addressing these challenges, AI can genuinely drive breakthroughs in drug development and improve patient outcomes.
