AI > Chemistry and pharma
pharma organisations to streamline development pipelines and potentially get drugs to market much quicker and for a reduced cost compared with traditional approaches. However, AI may be required for some of the most complex and unsolved diseases, such as Parkinson’s, lung cancer or multiple sclerosis. Causaly’s AI solution is specifically designed to untangle this complexity, empowering scientists to understand the underlying disease biology. Yiannis Kiachopoulos, co-founder and CEO of Causaly, said: “Knowledge is the lifeblood of research organisations, and we are committed to our mission to make it discoverable, working with our customers to make sense of their scientific data and apply insights to enable evidence-driven decisions. Recent advances in AI open completely new possibilities, and there is a great need for transparent AI systems that science leaders can trust.”
Preparing for the future To adopt AI, lab-based organisations need to prepare data sources and ensure that existing data sets are normalised curated and machine-readable so that they can be used for AI/ML projects in the future. AI should be considered a possible end goal for reaching a certain threshold of digital transformation. While it can be adapted for a single project, the work that goes into engineering existing data to fit with an AI framework can reduce productivity and ultimately reduce the benefits of using AI. Andrew Anderson, Vice President of
Innovation and Informatics Strategy at ACD/Labs, provides an overview of this process. ACD/Labs is helping customers to prepare for a future where data is routinely collected and organised based on data standards and machine-readable formats. This may be used for AI at some point in the future, but the very process of curating and engineering data should provide its own benefits to productivity and the discovery of new drugs. “What we’re doing right now is
observing, and helping folks prepare for that future,” said Anderson. “We often work in both the drug discovery and drug development areas in particular. So if you think about what is important in drug discovery, it’s among many, many things, identifying molecules which will have some sort of therapeutic intervention property against a disease or infectious disease target. Thinking about what folks would do in a lead optimisation paradigm presently, and how artificial intelligence and machine learning tools can make
that process more efficient, more reliable, or predictable.” “To train those predictive models,
you need data and lots of it,” Anderson continued. “What we’ve observed over the last 10 years is a lot of effort being focused around biological performance, taking assay data, performance data, molecular information, chemical structure information, for example, and building correlations between what has been observed experimentally. This is based on the identity of specific therapeutics and having these very large models with vast data collections and revealing and
‘We intend to create groundbreaking foundation models in biology and chemistry at a scale unlike anything that has ever been released in the biological space’
Chris Gibson, CEO, Recursion
generating latent relationships between structure and biological activity.” However, there are clearly other
variables to account for when it comes to the overall efficacy of a new compound. “Being able now to supplement biological potency prediction capabilities with off- target effects, making these molecules ultimately more selective for targets of interest,” stated Anderson. “So, you reduce, for example, toxicity.” “Because that data is heterogeneous,
there are a lot of different variables that you’re looking to describe with your various experiments, Anderson noted. “That heterogeneity requires a lot of thought into the normalisation of that type of data – especially when compared to like be traditional biological endpoints that folks have built models for over decades of high throughput screening and cell assays and the like. There’s a well-established data model for such data types but these
emerging new requirements or demands for data models for these off-target data types, require additional thinking. The role that we’ve played in this in this particular area is helping folks translate the data as it exists at its source, the disposition it exists in, and converting that and facilitating incorporation into these large model development activities,” Anderson said. Making the best use of all of this newly
created data requires organisations to carefully think about what data standards they need to adopt. ADF from Allotrope, UDF from Pistoia or AnIML are the most well-established. Still, each organisation must make their own decisions about which data standards are most appropriate based on their specific organisation’s use case and data types. Anderson said: “Talking to folks that look towards the future, not just myself, but management consultants or folks that are looking to help organisations build their labs of the future. Before they start moving data around and trying to normalise it or trying to shape it in a way that it’s digestible for these predictive unsupervised training applications, they suggest establishing a data standards practice within your own organisation.” “One particular standard that we’re looking at from our business is the ISA88 standard for experimental terminology,” Anderson said. “This standard helps to define experimentation variables.” Effectively this means providing standards and recommended practices for the design and specification of batch control systems as used in the process control industries. This would help scientists describe experimental steps using a machine-readable format. “What the ISA88 standard allows you
to do is if you have a software application that consumes your depiction of a laboratory experiment,” said Anderson. “It will convert that depiction to a set of structured text. In this case, it’s XML. Where you have tags associated with different operations like ‘heat’ or ‘stir’, you have different tags for material, and different tags for samples. so that when you present your data to a machine learning training set, it’s structured so that the machine can understand it.”
A new breed of AI Causal AI is an increasingly popular type of AI that aims to discover cause and effect relationships and can be used to augment
For more info about drug discovery, visit:
www.scientific-computing.com/energy-environment
SCIENTIFIC COMPUTING WORLD
Summer 2023 Scientific Computing World 19
g
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42
