Unlocking the potential of artificial intelligence in drug discovery-https://lnkd.in/e9p2Qz8j. Highlights:
1. Barriers must be addressed to unlock the full potential of AI:
• Trust in AI is a major barrier in many of the settings explored in this report.
• Lack of high-quality data sets, access to mature tools, and relevant AI and drug discovery capabilities constrains the value being delivered from AI today.
• The challenges are particularly acute in applying AI to commercially less attractive therapeutic areas and for low-middle income countries (LMIC) researchers looking to harness AI. For example, longitudinal population datasets that can be mined to understand diseases and identify new targets may be scarce, of lower quality or absent in LMIC settings.
2. Initiatives are emerging to tackle these barriers:
• Efforts such as the World Economic Forum and University of Oxford’s AI Governance Research group are working to improve understanding and trust in AI across a range of settings including medical research.
• Initiatives are being established to create or enrich data and enable greater access. For example, the US NIH is also funding grantees to clean and standardise existing datasets to improve their applicability to machine learning techniques – which is particularly critical in data-poor therapeutic areas
3. To truly unlock the potential of AI in drug discovery, this report identified a number of opportunities across the areas, as described in Figure 18.
Whilst in most cases, solutions will differ by use case, there is also an overarching need to build broader trust in AI, and the value it could deliver to the drug discovery field.
There are many paths by which this can be attained, with transparency being fundamental to help cut through the hype in the field today. Initiatives could include cataloguing the successes and failures of AI-derived assets, demonstrating tool performance and breadth (the CASP and CACHE competitions are already doing this), and transparently communicating the limitations of tools that are newly developed or available [27], [53]. Examples of the latter can be seen in OpenAI’s publication of the outcomes of red-team testing of GPT-4, or AlphaFold’s publication of a per-residue confidence score in its protein structure predictions [54], [55].
#artificialintelligence #deeplearning #machinelearning
Kudos to Neil Bence, Kai Wang, and the team for pushing the boundaries of research and innovation!