Or is it a new hyped label for practises already existing for some time already? What’s its future?
In the last 5 years, AI has become omnipresent in all areas of life, especially since the launch of ChatGPT by OpenAI (source report). R&D is not escaping the trend. But is this a new thing in Science and R&D?
R&D and AI/ML
Looking under the hood at how AI (Artificial Intelligence), ML (Machine Learning) and LLM (Large Language Model) work, they all use large sets of data to detect and learn patterns in the existing dataset in order to predict outcomes or answer a question, that is similar to what was available before.
Scientists have used analytic methods and predictive tools for many years to help guide their experiments and designs. In small molecules R&D, I used AB-initio modelling in the 90’s to guide the design of molecules to dock in a particular protein binding pocket. It was not labelled AI then but many of the elements were similar if not identical.
What is new then?
The explosion of resources available:
- More computer resources with the omnipresence of Cloud computing. It takes a lot of CPUs to crunch very large amounts of data and perform the complicated calculations required. Nvidia in particular, has shown an exceptional growth in the life sciences sector.
- Many digital datasets since the digitalisation of life sciences is in full swing. Some of this data is made available for broader consumption either publicly or internally within a company. That said data may have been digitalised, but its quality may be poor and not suitable. There are still a lot of improvements needed to fully exploit data. I’ll come back to this.
- The emergence and rapid expansion of a new scientific discipline: computer science but more specifically Data science. There are now many data scientists, data analysts, and statisticians. It is now common to have specialised teams to “wrangle” data, analyse it and make sense of it, proposing different aspects that the bench scientists may have overlooked. Or even better bench scientists training in data science.
- Many algorithms: they come with the many data scientists. The number of companies offering AI services in life sciences has exploded in recent years and the market is set to continue to expand.
What is the future of AI in R&D?
I don’t have a crystal ball but I am sure that AI is here to stay. AI/ML is a great tool for R&D, allowing scientists to analyse large amount of data, identify trends that may not visible by the “naked” eye, predict and suggest what to do next in a particular circumstances.
The hype about AI will calm down the more people realise that AI and ML are tools that you need to learn to use rather than something that is going to provide the universal answer to life, the universe and everything.
What is slowing down the adoption of AI in R&D?
I can see 3 major challenges in no particular order:
- the quality and availability of the data to train the models.
- the cultural barrier for adopting AI
- learning how to formulate the questions to ask correctly.
In the Pistoia Alliance survey report, these 2 first items were clearly expressed by the respondents.
Data quality and availability can be solved with data governance, harmonisation and adoption of ontologies. This is especially important looking forward, for the data yet to be generated and captured. Data strategy and governance should be put in place and personnel trained to data good practises, to ensure the data produced from now is of good quality and can be used confidently and reused in the future. When it comes to existing data, it is more tricky due to the disparity in the data recording, the lack of context and description of the data.
The cultural barrier for the adoption of AI in R&D comes mainly with the lack of trust in my opinion. Scientists are inquisitive by nature. They will want to understand how the responses from AI are generated, what is the source data and what has been done to derive the conclusion. Therefore it is important that AI provider enable the scientists to see the original data source. The existence of “AI Hallucinations” where AI detects inexistent patterns, contributes to the trusts issues.
Like with other tools, you need to learn to use it to make good use of it. AI is not different. In addition to feed it quality data, you need to learn how to interrogate it properly. I am sure you have play with the open tools available and have noticed how you can obtain very different results when asking the same question but formulating it slightly differently.
So, is AI new in Science?
In conclusion, for me, in the scientific domain, what is known as AI today is the evolution and expansion of computing. In Sciences, it is a great opportunity to analyse and explore more and more efficiently and intelligently, provided that you learn to use it properly and the data is available and of good quality – think FAIR.
Leave a comment to discuss or even better: contact Nathalie to discuss you views and challenges.
Leave a Reply