The Great Challenge

Client is a biologist who came to Anfluss with a twofold ambition: to develop an AI so knowledgeable in microbiology that it can answer any questions about cell cultvation and simulate bioreactor experiments computationally for estimated measures, before the researcher would set up and run said experiments, thus saving time and resources.

Every year, failed bioreaction experiments costs millions of dollars along with precious human time, effort, and motivation. Furthermore, when a lab experiment manages to yield essential proteins, one must face decisions to scale up (producing more products in a bigger, industrial-scale reactor) or scale out (producing multiple variations with large amounts of similarly scaled and precisely controlled parameters). Each option in turn presents its own set of biological, mathematic, and computational problems. If only an AI could instruct and inform lab researcher and bioengineers, simulate experiments and detect potential issues before resources are spent, and assist in bioengineering computation and decision-making.

Client has shared the idea with a few potential investors and prospective partners, who consider the vision too ambitious, especially as many people associate the concept of “AI” either to the recently popular Large Language Models (LLMs) or to a yet-to-come general intelligence (AGI). Regarding the client’s first vision, LLMs can pull relevant knowledge from embedded documents, as in a GraphRAG process, to emulate a knowledgeable biologist (provided a rich library of research paper and books, aka. knowledge base). However, regarding the second vision, no current LLM model can run a full-on simulation of a bioreaction. If an LLM is prompted to do so, it would create codes to run simple simulations with limited parameters and vast margins of errors. Not to mention, in a bioreactor lab workflow, it would not be efficient for an LLM to create a different simulator every time a biologist asks “Can you simulate…”

In accordance with our mission to support socially beneficial startups, and with an attitude to be the client’s longterm partner, whose success is our success, Anfluss engages with the client to:

• Define the problem as multi-faceted and thus applicable to a multimodal strategy
• Design a minimum viable product

Solutions

Modeling such complex processes as cell cultivations require synergized knowledge of scientific computing, biology, mathematics, and data. Anfluss neither dismissed the client’s ambitions nor attempted to build an impossible “biologist AI.” Instead, we sought to define and design, to keep clients on track towards a minimal viable product with flexibility to improve and grow as AI technology progresses. But first, the team needed to educate ourselves on biology. During the first few weeks of getting to know the client, we studied papers on cell cultivation computation, met with experts, and even visited biology labs to familiarize ourselves with the process.

We defined a product strategy that incorporates multiple data science process, not one ambitious “AI”:

• GraphRAG for documentation management and knowledge query
• Database to college and manage lab experiments data
• Simulation models that combines mechanical and statistical methods

This strategy allowed client to focus on building and delivering a viable web app for demonstration and prospect onboarding. With this initial web app, which has the capability to store and manage experiment data in a more organized and analytically useful way than disparage Excel files, the client was able to onboard initial adopters who need to manage their lab’s workflows and documentations. Given limited data, the simulation model does not yield accurate preditions. However, as the startup collects experimental data according to agreements with its initial adopters, it will in time build strong datasets for statistical modelling of specific cell lines and proteins. These data will also come in handy when foundation models improve iteratively into the future.

Anfluss Team also supported clients in coding aspects of the platform, particularly the bioreactor simulation model for demonstration purposes.

Results

With this strategy, client created a minimal viable web app, onboarded initial customers, and successfully partnered with experienced software engineers and product managers who are equally passionate and technically proficient. Client has shipped and scaled with speed based on this product strategy. Anfluss is honored to have been a major supporter of client in the ideation and prototyping phase of client’s scientific-entrepreneurial endeavor.