I spent 50 hours building a computational biology LLM Agent - Part 2
Reflecting on hours 10 to 50 of building Agent Phil, our in-house AI framework to make cell simulation accessible and more useful, I feel that feedback from my clients were crucial in prioritizing Phil’s development. Here are my reflections.
LY
by Laurence Yang
Client-Driven Development
Recent interactions with clients have highlighted some critical needs for small and mid-sized biotech companies:
What levers to pull?
An early-stage biotech startup emphasized the need for "useful recommendations on what levers to pullto achieve their customer's specified bioproduct."
How to quantify ROI?
Another client expressed difficulty in quantifying the return on investment (ROI) for innovative projects, particularly in AI and knowledge management.
Communicate value of simulation to non-experts
Two other clients reported challenges in communicating the value of computational simulations to non-technical customers.
Addressing Client Challenges
Here’s how I envision Agent Phil helping to address some recurring challenges expressed by clients:
Learning from past case studies
As our knowledge base of case studies increases, Phil’s ability to estimate KPI improvements for similar projects will become refined.
Forecasting multiple scenarios
Phil is developing the ability to forecast numerous scenarios in response to context-rich user prompts. Coupling this ability with a consultant’s expert judgment will yield unparalleled value to clients.
Translating simulation into actionable levers
LLMs excel in language.
Phil interfaces with multi-modal models capable of translating complex simulation results into simple language highlighting the key features and ‘actionable levers.’
Simulation results include data frames of network fluxes, plots of production envelopes, dynamic bioprocess simulations.
The ability to translate complex results into clear, concise language that non-experts can grasp readily is crucial for any team.
Especially for small and highly technical teams, as communication barriers must be overcome at two levels:
Within the organization between scientists and decision makers
When communicating the business impacts to customers
LLMs can help. Especially one possessing domain expertise and a knowledge base of relevant case studies like Phil.
My Prediction for 100 Hours Developing Agent Phil
Looking ahead to the 100-hour mark, I predict a significant shift in my approach to working with Agent Phil.
By that point, I anticipate spending more time on learning how to guide Phil's responses and thoughts to extract the most useful information.
The focus will likely transition from training the AI agent to training myself on learning how to interact and converse with Phil in ways that yield the most valuable insights.
This evolution in our interaction with AI agents like Phil represents an exciting frontier in computational biology. It suggests that as these tools become more sophisticated, our role as consultants, researchers, and developers will increasingly involve mastering the art of AI collaboration:
Knowing how to ask the right questions and interpret complex outputs in ways that drive innovation and solve real-world problems in biotechnology.
Subscribe to newsletter
I invite you to join me in my transparent journey in exploring the potential of AI-augmented computational biology, whether you're a biotech startup looking to optimize your processes or an established company seeking to leverage cutting-edge AI tools.