AI in Life Sciences: just another tool or the ultimate game changer?

This article was originally published in Dutch by IT Daily.

Artificial intelligence (AI) is hot today, but it doesn't start and stop with applications like ChatGPT. Moreover: how should you handle AI in Life Sciences? Elisa Canzani, Data Science Lead at Cognizant, takes us on a journey to show that AI is not a one-size-fits-all solution, especially not in life sciences or other scientific branches. Together with Cognizant we explore the pitfalls of AI in life sciences and what is needed for a solid AI foundation.

A Diverse Range of LLMs

Many of the GenAI tools today work based on LLM’s (large language models). As an organization, the chances are slim that you will build your own LLM. It's better to look at a pre-trained model available for (commercial) use and which operates within your scope. OpenAI, Anthropic, Meta: the range of models is extensive. Take a look at the complete list here.

“You see the abbreviation LLM everywhere, but it is not always the best way to go for GenAI applications. A Small Language Model (SLM) is often an interesting option to fine-tune it yourself," says Canzani. "They are very reliable, efficient, and accessible for specific tasks."

Choose an (L)LM and Start Customizing

A popular approach to customise an LLM is prompt engineering. “The system prompt is the common way of imposing behavior”, and it ensures instructions are used while generating desired output to answer user prompts.

While prompts are effective but limited in the amount of instruction that can be given, a very powerful way to achieve customization is Retrieval-Augmented Generation (RAG). This combines the model with an external knowledge database. "When a user asks a question, the model searches the database for relevant information and processes it along with the question to provide a more accurate and specific answer." An important improvement in this is Graph RAG, which uses graph-based methods to optimize the search" says Canzani. This is crucial for sectors where precision and contextual relevance are essential, such as supply chain management or product development.

LLMs can also be fine-tuned: this is will not add knowledge, it rather adds new patterns of language, e.g. domain-specific terminology for healthcare or logistics. However, Canzani says that “fine-tuning is also a delicate task because we start changing the weights of our model and this can lead to what we call “catastrophic forgetting”.

Canzani: "These approaches make it possible to harness the power of an LLM as long there is an observability framework in place allowing organizations to retain control over the relevance, reliability, and accuracy of the generated information."

Hybrid AI Systems as a Basis for Better Decisions

LLMs can deliver impressive results, but they have limitations in domain-specific knowledge and logical deduction. Hybrid systems combine the strengths of generative AI with classical AI techniques, such as Bayesian Networks for causal inference. "This interplay makes it possible not only to answer questions or generate data but also to provide deep insights into cause-effect relationships. Think, for example, of problems in production processes or quality control."

An important element of hybrid systems is the concept of agentic solutions, where complex tasks are divided into smaller, more manageable subtasks. These subtasks are delegated to specialized agents, such as an SQL assistant for information retrieval or a root-cause analysis agent to identify the cause of, for example, a bottleneck in the supply chain which drastically increases the end-to-end lead time.

"Such an agentic system works as a network of collaborating experts, making it possible to answer very specific questions with accurate, contextual information," says Canzani. "This is especially useful in sectors like supply chain management, where multiple variables and dependencies influence the final results."

Another important aspect of hybrid systems is the integration of human feedback. Although agents and AI models can perform many tasks independently, human expertise remains crucial for tasks that require a high degree of interpretation or ethical considerations. Moreover, agentic systems continuously learn from user interactions and become increasingly better at delivering tailored solutions.

A Cognitive Framework That Provides Structure

Cognizant's cognitive framework for responsible Gen AI development offers a structured approach to building and implementing AI solutions in a business context. Cognizant's framework encourages an iterative approach. The process begins with mapping the user journey and desired functionalities, after which the AI solutions are continuously tested, adjusted, and scaled based on new user insights and changing business needs.

The development phase of the framework consists of four layers: the interaction layer, the generative layer, the computational layer, and the information layer. Each layer plays a specific role, from collecting user input to retrieving contextual data and generating decisions.

One of the unique aspects of this framework, according to Canzani, is the focus on dynamic context curation. The information layer must be continuously updated with the most current and relevant data, so decisions are based on accurate and up-to-date information.

Additionally, the framework proposes the LLM to act as decision-maker also for sub-task delegation to classical AI or other expert systems. The result: more reliability, precision, and efficiency, especially in domains like pharmaceutical production or supply chain management. An additional layer that crosses all the others is the observability layer, Canzani: “there is no AI system that can move to production without proper performance monitoring in place”.

Such a framework is what makes Cognizant's GenAI solution unique. It was ideated by the BeNeLux team and got the first place at the Cognizant Global Bluebolt GenAI Ideation Challenge for which more than 3000 ideas were submitted. Those who want to delve deeper into the possibilities can consult this video.

Practical Examples of GenAI

Talking about theory is important, but using GenAI in practice is even more important. During Canzani's presentation she points to unique opportunities to transform and accelerate processes in life sciences and manufacturing. A few examples:

• The discovery of new drugs by generating synthetic data, which can be used to model new complex bioprocesses and also accelerate tech transfer from R&D to manufacturing of digital twin applications.
• Designing molecules, RNA sequences, and protein structures, which is essential for R&D in pharmaceutical companies. GenAI significantly shortens the time needed for research and clinical trials, making it possible to develop solutions for critical health problems faster.
• In the manufacturing sector, generative AI is used to optimize production processes and improve quality control. Digital twin orchestrators use agentic AI to integrate data from the factory, enabling real-time analyses.
• AI-based assistants for SOPs (Standard Operating Procedures), which guide operators through complex tasks and ensure that procedures are consistently followed. These applications not only reduce operational time and increase production speeds but also ensure compliance with strict quality and regulatory requirements.

Looking at the complexity of supply chains in life sciences, GenAI can offer powerful solutions for challenges like delays and inefficiencies. Canzani showcases how a multi-agent solution can help to get supply chain data insights, identify root-cause of current total lead time, and accordingly suggest the right mode of transport to balance costs and delivery times. These applications not only offer operational improvements but also enable companies to make strategic decisions based on reliable and transparent data.

GenAI: More Than a Hype or Buzzword

"GenAI is often mentioned in the same breath as bias, hallucinations, and privacy risks. These concerns must be addressed with solid cognitive frameworks, data transparency, and ethical guidelines such as the EU AI Act," says Canzani.

We see how GenAI is not just a new tool but a game-changer. "As long as you combine this new technology with a solid foundation consisting of data, analytics, and other well-established AI tools," concludes Canzani.

Generative AI is not just a hype or buzzword, but a fundamental technology that can transform processes, provide new insights, and stimulate sustainable growth. The future lies in combining AI's technical power with human expertise to build solutions that are both innovative and responsible.