I tend to be somewhat skeptical of artificial intelligence (AI). In the Chemistry World newsroom, we routinely see all sorts of wild and wonderful claims being made – in both academic and industrial contexts – about what AI is going to revolutionise next. Few of these hyperbolic assertions stand up to a great deal of scrutiny.
Chemistry and materials science are deeply rooted in the physical world, and there are significant holes in the knowledge frameworks and data required to build reliable and consistent models. But those holes are closing over time.
The models that do stand up tend to be those that have applied their algorithms to relatively well-defined problems, where there are high-quality datasets to learn from and straightforward ways to experimentally test, verify and refine the model and its outputs. The protein folding and structure prediction models recognised in the 2024 Nobel prize for chemistry, for example, built on huge libraries of crystal structure data and many years of algorithm development. Meanwhile, materials prediction start-up CuspAI (see below) has enlisted a cadre of partners to make and test the materials its models predict could address their problems.
Across industry, there is a wave of investment in AI ‘solutions’ to various problems. Some of this feels like it is being driven by a collective fear of missing out – ‘if our competitors are the ones that make this work, we will suddenly be at a big disadvantage, so we need to keep up’. Nebulous promises of cost-cutting (such as Dow’s justification for cutting thousands of jobs earlier in the year) are also very enticing in sectors that are under significant financial strain.
But it’s important to be realistic about what AI can, and can’t, achieve. Drug discovery offers a good example. There has been a huge increase in firms using AI to try and streamline the drug discovery process and hopefully bring down its astronomical failure rate. Some are augmenting existing medicinal chemistry programmes, others rely more heavily on computation.
These programmes may mean chemists spend less time making molecules that fail in preclinical testing, which is certainly beneficial. However, they have not yet had much impact on clinical trial success rates, which is the biggest (and most expensive) bottleneck in drug development. In the words of Chemistry World columnist Derek Lowe, the robots are simply queueing up to fail. But if those failures can be understood, codified and used to build more refined models, then the we should be able to make progress more quickly – it’s important to know what doesn’t work as well as what does.
I remain confident that carefully considered application of AI and machine learning – in combination with human chemical understanding and experimental verification – can have a significant impact on many of the gnarly problems that all the chemistry-using industries face. Hopefully those efforts will not be drowned out by burnished promises of quick fixes built on flimsy, opaque foundations.
No comments yet