Drug discovery has always been a game of slow, expensive gates. From target identification to final optimization, each stage typically takes months, sometimes years. Neil Patil, co-founder of Chai Discovery, calls it a “very like waterfall model”—and for ambitious builders, that old way of doing things is dead weight.

Patil and co-founder Matt McPartlon are building Chai Discovery to flip that model on its head. Forget the linear, multi-year slog. Their vision is agile, iterative loops for designing new proteins and antibodies. Think of it less like a traditional pharma lab and more like a software team pushing daily builds.

Key Takeaways

  • Traditional drug discovery follows a "waterfall model" where each stage, from target to optimization, takes months or years and costs a fortune to even try new things, according to Chai Discovery's Neil Patil.
  • Chai Discovery uses AI models (like Chai 1, 2, and 3) to transform this linear process into an agile, iterative loop, generating "really promising candidates" much faster, akin to modern software development.
  • The core of this new method is a continuous feedback loop: lab results from initial candidates are used to condition and refine subsequent AI model runs, making the AI smarter with each cycle.
  • The ultimate goal is "one-shot" denovo design, producing therapeutic-grade molecules straight from the models, though Matt McPartlon admits this ambitious target faces "tons of roadblocks."
  • For founders, this means embracing the idea that products are often temporary bridges, not lasting monuments. Neil Patil notes that software built today might only last a year before needing replacement by the next generation of tools.

The Method: Drug Discovery's Agile Loop

Traditional drug discovery is notorious for its glacial pace. Imagine a waterfall, where each step must be completed before the next can begin. Patil explains, “This notion of target discovery and hit discovery and optimization where each of these has a gate and takes a few months to a few years is this very like waterfall model, right? Where the cost of trying things and getting things early is very expensive.”

Chai Discovery uses AI to break this inertia. Their models, Chai 1, 2, and 3, don't just predict; they design. The crucial shift is from prediction to design, then to iteration. When an AI model can spit out "really promising candidates" early in the process, you don't wait for a gate review. You test, learn, and feed those lab results back into the model.

“If you start to get in a regime where you can have models give you really promising candidates, you can start to make that look a lot more like a loop,” Patil says. “It's akin to like becoming more agile in software development.” This continuous feedback allows them to climb "levels of abstraction" in product development. They move from simple molecular inspection to orchestrating complex campaigns against entire biological pathways, refining their models with every data point.

The ultimate goal, according to McPartlon, is to “really produce drug-like molecules straight out of the models.” This isn't just about tweaking existing molecules; it's about true "one-shot" denovo design—creating brand new therapeutic-grade molecules from scratch, directly from AI.

Where This Breaks Down

While the vision is compelling, the path to "one-shot" design is steep. McPartlon freely admits, "of course this is going to be hard and like there are going to be like tons of roadblocks." Building the reinforcement learning (RL) stack necessary to teach models specific drug properties and then prompt them effectively is a massive technical hurdle. It requires vast amounts of data, immense compute power, and sophisticated experimental validation cycles.

The iterative mindset also requires a different approach to product development. Neil Patil reflected on this existential challenge: “man, all this stuff we're building in the product to like visualize molecules and do this like maybe I'm just going to have to throw it all away when like Matt ships like Chi 4, right?” This isn't a bug; it's a feature of building at the bleeding edge. Products become temporary bridges, designed for rapid deployment and even more rapid obsolescence. This can be jarring for teams accustomed to building software meant to last decades.

What to Do With This

Stop planning your core product or research as if it's a 20-year build. Instead, map out your current development process and identify the longest, most expensive "gate" or handoff point. How can you insert a tight, data-driven feedback loop this week that brings early results back to inform your next steps? Then, examine your most recent feature release. Was it built to last, or was it a crucial bridge to unlock the next, better thing? Plan to deprecate early and often; your best work may be the very thing that makes your current product obsolete.