Another interesting thread was the role of alternative architectures, where general-purpose LLMs may find their place in orchestrating workflows as part of autonomized agents. This could lead to a shift towards smaller, customized models, with an emphasis on routing and other less glamorous components, which was discussed in detail during the aforementioned Workshop.

While the limitations of GenAI are acknowledged, the community remains hopeful for ground-breaking use cases. For instance, in the field of GenChem, there is a desire to apply the same approach that emerged from AlphaFold for target structure prediction to small molecule prediction. However, the lack of evolved data in chemistry presents challenges that still need to be addressed.

Workshop attendees reached a consensus that LLMs contribute to efficiency improvements but not necessarily accuracy. They can assist in generating email blueprints or ontology candidates, yet human expert review remains essential to ensure accuracy; or indeed one can utilize computational representations of domain-specific knowledge to help ground, or steer, a generalized LLM in the right direction… enter ontologies.

Ontologies and Knowledge-graphs

Modeling knowledge in ontologies is not a new idea – the community has been capturing domain-specific knowledge in these standards for decades, and this is showing no sign of slowing!

The concept of a Digital Twin, which was expertly walked through by Caroline Chung, MD Anderson, at Bio-IT, shows wonderful potential within medicine and demands a mention. Replicating the idea demonstrated in aerospace in healthcare presents challenges due to the complexity of modeling multiple scales such as genome, cells, organs, and the entire human body, as well as the need for high-quality and real-time data. The application of ontologies and standards is crucial for progress in this area.

The importance of ontologies in generating and modeling knowledge graphs is well known. However, a common topic during events was the interplay between knowledge graphs and LLMs, particularly during a fantastic panel discussion entitled Pharma Knowledge Graphs and LLMs: Antagonistic or Synergistic, chaired by Tom Plasterer of AstraZeneca. The main interplay between the technologies appears to be twofold: i) generating relationships and KG content and ii) simplifying querying through NL-to-syntax conversion.

With the understanding that ontologies and knowledge graphs can be used to capture the knowns, an interesting thread discussing how these may be used in conjunction with the output of LLMs to help distinguish hallucinations and identify candidate novel hypothesis generation emerged. Obviously, one must acknowledge the associated risks with such approaches – ensuring that one can disprove and fail fast for those who are going nowhere.

A typical application of LLMs in the context of KG was lowering the barrier of entry, supporting the concept of data democratization. However, one obvious thing, or if it is not, it should be – we cannot lose focus on the data. There is no point in lowering the barrier of entry to rubbish data…

Data still rules supreme

The wonderful presentation by Dame Janet Thornton from the EBI at Pistoia highlighted the importance of quality data. Dame Janet touched on some of the limitations of AI within the Life Sciences, including data quality and a lack of standards (as well as confidentiality and ethical barriers). Her advice? …begin with data.

Furthermore, the importance of quality and FAIR data was covered in the Data Readiness for AI panel, led by Santha Ramakrishnan of Bayer. A direct quote from that discussion was a want to ‘increase the value we get out of data ‘and that, to do so, scientists should play a more central role in data management processes.

Image data, particularly in pathology, came up throughout the Bio-IT conference as well an increasing need to support multi-modal data, particularly text and images; others such as audio, video, 3D, and IoT received less attention.

As we navigate limitations and explore new frontiers, AI continues to shape the future of the Life Sciences industry, and the requirement for quality data to train these models persists. For areas with sparse regions, the concept of synthetic data persists as a noteworthy topic, utilizing LLMs to generate data for training smaller, more bespoke models for business tasks.

Lastly, the notion of a Data Product gained significant traction. These self-contained data solutions directly address business challenges and have the potential for monetization.

Launch of SciBite Chat

As part of Bio-IT World, we also launched SciBite Chat – our conversational chatbot that sits atop SciBite Search and enables users to have complex natural language-based conversations with their data.