AI-designed proteins test biosecurity safeguards

Recent updates to biosecurity screening software are enhancing its ability to detect potentially harmful proteins that could be synthesized using artificial intelligence.

This software plays a crucial role globally, overseeing the processes involved in artificially creating proteins to prevent malicious actors from generating hazardous substances like toxins. Researchers revealed in the October 2 issue of Science that minor alterations using AI on known toxins or viral proteins can sometimes circumvent these protective measures. However, by strengthening the existing screening protocols, the ability of these systems to identify risky AI-generated proteins can be significantly improved.

“Advancements in AI are driving innovations in the fields of biology and medicine,” stated Eric Horvitz, chief scientific officer at Microsoft, during a news conference on September 30. “However, along with this new capability comes the obligation for careful oversight and responsible management of risks.”

Proteins are fundamental to biological processes. They are involved in various cellular functions such as cell assembly and transportation of materials within the body. Thanks to AI, researchers are discovering methods to refine current proteins for specific purposes, design entirely new proteins, or even create novel organisms.

While AI can produce digital designs for proteins by identifying the necessary amino acids, it cannot fabricate physical proteins directly. Instead, DNA manufacturing services link the correct genetic sequences and distribute these synthetic genes to laboratories. Computational tools then screen these requests to ensure the resulting genes don’t yield dangerous proteins.

Horvitz and his team conducted simulations to evaluate biosecurity screening models, identifying vulnerabilities that might permit AI-generated proteins to bypass detection. They created approximately 76,000 designs for 72 hazardous proteins, including ricin, botulinum neurotoxin, and other proteins associated with viral infections.

Despite the biosecurity systems successfully identifying the DNA of nearly all proteins in their original states, numerous AI-modified versions evaded detection. Updates to the software improved its performance, enabling it to identify genes even after they had been parsed into smaller fragments. However, about 3% of variations were still overlooked.

All tests were performed on computers, meaning the team did not create any physical proteins in a laboratory environment, and it remains uncertain if the AI-altered versions maintained their biological functions.

In reality, the incidents of biosecurity screenings blocking requests for suspicious proteins are rare, remarked James Diggans, vice president of policy and biosecurity at Twist Bioscience, a San Francisco-based DNA synthesis firm, during the briefing.

While cybersecurity threats occur frequently, Diggans emphasized that “very few” individuals have attempted to manufacture malicious proteins. “While these systems provide a crucial defense against potential threats, we can take assurance from the fact that this scenario is not prevalent,” he noted.