Vertex 制造系统工程师谈生物制造数字化架构须先于分析：捕获工艺意图以防技术转移中的知识泄漏

By Tyler Menichiello, Chief Editor, Bioprocess Online

I’ve covered digitalization and institutional AI adoption before, and I’ve stressed the importance of preserving tacit knowledge during tech transfers. But at this year’s BioProcess International conference in Boston, Vertex Pharmaceuticals’ Martin Kockx connected these concepts masterfully in his talk on digital architecture.

Kockx, a principal manufacturing systems engineer at Vertex, gave a presentation titled “Transforming Biomanufacturing: Why Architecture Must Precede Analytics,” where he explained the importance of capturing process-specific context or the intent and rationale behind developmental decisions, and doing so in a way that’s accessible to AI systems, scientists, and engineers.

He argued that preserving this context and reasoning at each stage of development not only protects against “knowledge leaks” during tech transfers, which significantly limits ambiguity and backtracking in later stages, it also strengthens the foundation for AI tools. As Kockx said during his presentation, AI systems built on broken architecture just fail faster and more expensively.

I was fortunate enough to catch up with him after his presentation to discuss the main takeaways for our Bioprocess Online audience.

The following transcript has been edited for clarity and readability.

You gave a talk on why the underlying architecture needs to precede analytics, specifically as it relates to AI and data systems. Can you summarize your talk and some of the takeaways you hope people left with?

Kockx: I think at the core of it, it’s about focusing on the intent. We go through a very lengthy process from discovery all the way through commercialization, and at every step of the way, there's usually a handoff. You might have handoffs between teams within a specific domain, but you certainly have massive handoffs between domains. You're going from discovery into early development; into process development for preclinical; then to clinical; and then hopefully, into commercial.

For commercial, you may go straight to a CMO, or it might go internally; we've had instances where they go back and forth. At every one of these handoffs, there's a massive transfer of knowledge and an opportunity for knowledge leaks.

If you think about it, all the tacit knowledge that went into a particular part of the process is not something you're going to write up in a paper that someone's going to read. It comes condensed and prepackaged for what you think they're going to need to know at the next stage. And if you get that perfect, great, but I find it is rarely perfect.

So, what do you do if you need to go back and figure out something that was decided or found during a previous stage of the process? Are you going back to ask the engineers? What if they left? Or are you going to dig through a bunch of files? How long is that going to take?

AI tools are going to make that a lot easier, but it’s still a backtrack.

So, as much as you can, tie all that context together and bring that intent across. That's a lot of what I don't see, the reasoning behind why we make decisions. Why did we pick this type of bioreactor instead of another type? Why are we doing electroporation instead of mechanoporation?

All of these are questions you might have down the line. Let's say you have a process that's really difficult and it's already been pushed out into production and you say, “Wow, this piece of machinery is really not suited for manufacturing.” Well, I can go find another piece of equipment that’s suited for it, but maybe they decided why they couldn’t for some reason three, four, or five steps before, but I can't find that rationale easily.

This halts and adds friction downstream from a continuous improvement and operational excellence perspective, which I think could be greatly helped by capturing intent where it's generated and making sure it follows the product or the process along the way.

Does any of this context or intent get captured currently?

Kockx: It’s captured, but is it captured in a place that's accessible? We talk about FAIR with AI (Findable, Accessible, Interoperable, and Reusable); if I can't find it, then it doesn't exist. It's not useful to us.

So, while it might be written down in a notebook or somewhere digitally, if it's not tied to something, I'm going to have a hard time finding it; the AI tools we want to build are going to have an even harder time finding it. So, the accessibility has to be there.

 Regarding the intent and context we talk about all the time in manufacturing: A lot of times, we stop at, “What's the plant model context? How does this fit into a plant model?” But there's also all the metadata. We think of metadata in real-world terms as the stuff attached to photos — where did it happen and all these types of things — and everyone gets upset at the use of their metadata, but we need it in manufacturing. What are all those reasons that operator tied to it? So, a lot of it is adding that context, capturing that intent, and making it accessible.

What practical piece of wisdom or advice would you offer to our audience if they wanted to take a step towards improving the way this data is captured, to include the intent and context at each phase?

Kockx: It’s going to run the gamut. It's hard to pinpoint, because you might have early research-stage engineers at a startup with very limited resources when it comes to digital and tech, or you might have big behemoth companies that have full stacks.

With AI coming along, if it doesn't already exist for you or is accessible for you, then talk to your IT professionals, talk to your digital people about getting that access or at least making it self-serve. Because nowadays, with some of this AI capability, you can build something that'll allow you to capture intent. It doesn't have to be a multimillion-dollar electronic lab notebook project. It doesn't have to be a full LIMS system. If you can use these tools to make your job easier, while also building in that intent capture at the same time, that decision rationale, then it's already digitally native, and then it's on your digital leaders to make sure it's intertwined, traceable, and findable.

What do you hope the audience takes away from the upcoming panel you’re participating in?

Kockx: The panel kind of goes a level deeper into what I talked about today, where I was talking about an entire architectural concept. Part of that is your data. At BioPhorum, we published a paper on AI data readiness as a risk-based framework. It goes heavily into the detail of the characteristics that make data AI-ready.

What is the framework for you to assess data readiness when you’re building an AI agent, an internal application, or even evaluating an application from another vendor? How can you ensure the data it's going to access is in a safe state? We talked a lot about keeping data clean and contextualized, but I think we have 29 or 39 characteristics spanning everything that would be important for data. How do you ensure it's there and ready so that you can use it reliably, responsibly, and leverage those cool tools?

What are you most excited about being here at BioProcess International? What seems to be the general sentiment you’re seeing? What are this year’s key themes thus far?

Kockx: Another shameless plug: I'm also heavily involved with ISPE, and one of the things we tout as the benefits of ISPE that I think is similar here is it doesn't matter what your role is, you can learn something new.

I'm probably going to catch a talk on ADCs and some other things later. I'm not going to ever build that; I'm not a process development scientist. I'm not down at that level, but I think it helps us as engineers and automation people, even in other areas, to understand what's coming.

Even something really high-level, like the talk on new modalities and technologies and what's coming down the pipeline, does that affect my job directly? Probably not. But is my company going to possibly have that come through their pipeline, and is it going to affect how we do digital for it? Is it some type of equipment that we're going to need to know how to get data out of and contextualize? Maybe.

I love the fact that I can go to any topic, and there's a lot of breadth of topic here. Plus, the showroom floor has all sorts of fun.