5 min Applications

Cradle makes biology programmable

Biotech player a "proud" Google Cloud customer

Cradle makes biology programmable

Cradle is harnessing AI to tackle biological challenges. CEO and co-founder Stef van Grieken has assembled a team with experience from Google, Deepmind, Uber, and the biotech sector. Leveraging Google Cloud, the company has already attracted notable clients.

We spoke with Stef van Grieken at the Google Cloud Summit in Amsterdam, where Google Cloud highlighted several customers. Van Grieken began as a programmer for Google Search in 2014, contributed to the traffic prediction feature in Google Maps, and helped design one of the TPUs (tensor processing units) used for AI workloads on Google Cloud Platform (GCP). Founded in 2021, Cradle is both a customer and a familiar face for Google, thanks to Van Grieken and the 40 employees who have backgrounds in Google Deepmind, YouTube, and other product teams.

Cradle’s mission: programming biology

Cradle serves R&D teams in labs worldwide, aiding them in finding the optimal protein sequences for products ranging from vaccines and cosmetics to enzymes for detergents. By manipulating DNA, Cradle tailors products for specific purposes, making them more effective, easier to store, or safer for the human body. While currently focused on designing optimal enzymes, Cradle also offers vaccines, peptides, and antibodies “in beta,” as noted on its website.

Cradle aims to make the biological world “programmable.” Van Grieken explains that biology essentially uses a non-human programming language: RNA acts as the application code executed by a cell, while DNA serves as the storage medium. However, the biological “compile time” is bound by natural laws and can’t be sped up with new chips or code refactoring. It takes about three months to “compile” a biological program due to various biological bottlenecks.

Some bottlenecks are unavoidable. Synthesizing DNA takes two weeks, and the length of synthesized DNA strands is limited, requiring assembly of DNA pieces (i.e.: “gluing” them together) to convey comprehensive instructions to cells. Finding the right candidate for the final product, E. coli in Cradle’s case, also poses limitations. “Cells don’t grow faster if you yell at them,” says Van Grieken.

E. coli cells need time to grow and are essential for converting DNA information into useful enzymes. The requirements for a drug are extensive: it must bind to certain cells, avoid others, trigger the right reaction, remain stable, be manufacturable in the first place, and be soluble in water. Testing the desired functionality at scale is challenging, often taking years before a product reaches the market. So, where can time be saved?

AI comes to the rescue

Van Grieken explains that traditional biotech research relies on randomness, with a success rate of about two percent for finding workable molecules. AI can increase this success rate by “two to ten times.” Searching for molecules randomly is computationally expensive and slow, but AI innovations significantly enhance the efficiency here.

Cradle’s AI expertise allows clients to advance their biotech projects by uploading lab data or starting from scratch and gaining insights into predicted performance before waiting three months for results. Companies like Johnson & Johnson, known for their COVID-19 vaccine, and Manus, which develops biological replacement ingredients, are already benefiting from these features.

AI models in biotech learn similarly to large language models (LLMs). Van Grieken likens the task to a “fill-in-the-blank” exercise, where the model earns points for correctly predicting amino acids.

Building AI from scratch

Cradle built its AI models from scratch using widely supported standards. While Van Grieken doesn’t detail the exact creation process, he mentions that using an Nvidia DGX system would have been too costly and ineffective for a startup. Scalability is crucial for a product that may become explosively popular, and continuous hardware upgrades are impractical. “You need a serious IT infrastructure to make this happen.”

Cradle turned to GCP, which offers scalable, state-of-the-art hardware. Van Grieken, a self-described “proud customer” of Google Cloud, has witnessed Google’s early AI stack development from the inside. With proprietary TPUs and mature tooling, he believes GCP has an edge over competitors.

There’s no risk of lock-in since much of the tech stack is standardized. “Most applications are Docker containers, and everyone uses Terraform for orchestration,” says Van Grieken. This standardization also keeps pricing competitive. “Google couldn’t suddenly make their TPUs six times more expensive.”

The importance of security

Developing specialized AI models requires robust security, especially when dealing with sensitive customer data. Van Grieken trusts Google Cloud’s security, helped in no small part by expertise from employees like Noé Lutz, formerly of Google’s Threat Analysis Group in Zurich. Google has a proven track record of defending against threats, adopting layered defense, and offering BeyondCorp Zero Trust Enterprise Security.

Google Cloud understands where threats originate, says Van Grieken. Internal security measures ensure that even Cradle employees cannot easily access customers’ intellectual property. Cradle’s IT environment uses encryption “all over the place”, zero-trust mechanisms wherever needed, and other security best practices. With highly experienced staff, maintaining a strong internal security culture is manageable. Van Grieken doesn’t need to worry about data security because Google Cloud has established a mature security infrastructure, not because it doesn’t matter.

Google Cloud is eager to showcase Cradle as a success story, and it’s no surprise to see why. Cradle’s rapid growth from an idea to a startup with prominent clients was facilitated by GCP’s infrastructure. The fast startup of cloud instances contrasts with the slower pace of biological processes. Optimizing other fronts is crucial when biology presents unavoidable bottlenecks. Cradle has mastered this optimization within its niche.

Also read: After the AI world, Databricks now wants to change the analytics market