Why Compugen chose K Pro to complement its in-house AI discovery platform 

Roy Granit leads Computational Discovery at Compugen, where his team hunts for novel cancer immunotherapy drug targets using Unigen, the company's in-house target discovery platform. When Compugen decided not to build its own generalist scientific AI agent, they turned to K Pro instead. Here's what he's learned since.

‍

You already have an internal AI tool. Why bring in K Pro too?

We have our own AI agent, called Genny, which is a nod to Compugen. But as a biotech, we have to balance where we put our resources and focus. Building and maintaining a general purpose horizontal AI agent that can agentify our workflows beyond our core work isn't the best use of our time. K Pro let us focus on our strength: the Unigen platform, our data, and our specific capabilities, instead of spending effort on something with broader use cases we chose to invest on the vertical AI agent and use K Pro for the rest.

‍

How have you found using K Pro?

I like it for deep-diving into a specific data set: understanding a specific gene or set of genes in a way that would otherwise take much longer, even with some coding. It's more seamless.

It's also displayed useful emergent agentic functionality. If I want to look at a property of the data that isn't labeled, K Pro will try to infer the missing label rather than stopping. That lets us explore data in more depth and from more angles. We've also used it to explore Owkin’s MOSAIC spatial data and other data types beyond our own.

I like to use it for open-ended exploration, when I have a specific question in mind. We're always asking ourselves questions about a target's biology, and that's where it's been most useful for going down that path. 

I’m looking forward to seeing how custom skills can help us shape K Pro’s answers to come even closer to the exact outputs we are looking for every time.

‍

How does it feel different from using something like Claude or GPT directly for this?

I believe K Pro is better suited to our domain. I think that comes down to the prompt engineering behind it, the thinking that's been built in to help the agent adapt to our needs. Take the missing-label example: Claude could think about that too, but it isn't instructed to, so I'd have to prompt for it myself. K Pro has already done some of that thinking for me, and it suggests the next step - "Would you like me to infer this label?" It saves the effort of prompt engineering, so I can spend that time on the actual science. There are also tools for spatial data and similar formats built in that Claude wouldn't deliver on its own.

K Pro also integrates pipelines - deterministic workflows, rather than LLMs which incorporate some randomness - which I like a lot because pipelines can be rigorously tested and verified. There's real value in structure and things that have been proven. AI is great because it's flexible, but it's not yet at the point where it beats a well-tested pipeline. It's good for something quick and dirty. But if you're trying to deliver a drug - which is what we've spent a long time doing - you need robust data, and AI on its own doesn't always QC things well. It can look at data from one angle and lead you to make mistakes.

‍

What matters most in an AI scientist or AI co-pilot?

It should act like a real co-pilot. Basically, something that helps your thinking and exploration, geared toward your actual goal, not something generic. It needs the right background to help you make the right decisions. And you need to be able to rely on it: it should be reasonably reproducible.

‍

Any highlights from working with the Owkin team?

It's been a genuinely positive, professional experience. One thing stood out: when the Owkin team onboarded some of our data into K Pro, they used it to recapitulate a published paper based on our own data. Seeing that K Pro could take our data and independently arrive at results matching real, published science was well appreciated, and it strengthened my confidence that K Pro can deliver genuine scientific value.

‍

Roy Granit is Head of Computational Discovery at Compugen.