Biosecurity challenges in the age of AI

The summary:

• AI is being used to design proteins—and these protein sequences could evade traditional biosecurity screening tools.
• Integrated DNA Technologies recently took part in a study to create new screening methods that can reliably identify harmful AI-designed protein sequences.
• The effort and results show the need to continuously monitor and evaluate biosecurity screening tools against new synthetic homologs and develop more sophisticated screening solutions.

Artificial intelligence (AI) is quickly working its way in every facet of our lives and science, and protein engineering is no exception. This technology offers immense promise for those seeking to use it to make breakthroughs in life sciences—but it may come at a cost. As AI tools become increasingly complex and sophisticated, the ability to use this technology to design and generate novel proteins that have potentially dangerous functions is growing. As a result, experts in the industry are calling for ever more robust biosecurity measures that can mitigate these risks. Let’s look at this in more depth.

AI used to boost protein engineering

AI tools are being used in several ways when it comes to protein engineering. One is in the prediction of what shape a protein would take based on its amino acid sequence. The use of these tools is being extended to learn from various data sources and then create better protein designs faster. Specifically, AI tools are being used to make proteins more stable, easier to produce, or more accurate at hitting their targets.

Enter protein engineering concerns. A critical control point in this process can be found in the screening of orders placed with nucleic acid synthesis providers—this is where the digital designs are translated into the physical instructions for creating proteins. The core challenge is in the reformulation of naturally occurring proteins of concern to create synthetic homologs—these are proteins with similar functions but much different sequences from their wild-type counterparts. These AI-redesigned sequences could evade traditional biosecurity screening tools, and this is where the threat comes in.

The challenge of synthetic homologs

Traditional nucleic acid biosecurity screening involves comparing the similarity of their sequences with databases of known hazardous proteins. However, with AI doing the design, these synthetic homologs can be engineered so these threats are much less obvious—meaning that a standard screen might miss them. This will require researchers to take a more sophisticated approach to biosecurity screening in order to detect potentially dangerous proteins even when their sequences aren’t anywhere close to known threats.

In a recent study, a group of researchers, including Adam Clore, the director of technology R&D at Integrated DNA Technologies (IDT), sought to evaluate the effectiveness of existing biosecurity screening tools in detecting these synthetic homologs. Initial testing revealed that some tools were not capable of reliably detecting AI-redesigned sequences. This highlights a concerning vulnerability in the biosecurity pipeline and underscores how urgently screening methodology improvements are needed.

“We found in a confidential process that several biosecurity screening software (tools), including screening methods in use at major nucleic acid suppliers, could not reliably detect such AI-reformulated toxins and viral proteins,” the authors wrote. “However, we showed that the screening machinery could be patched using insights from this effort to improve their ability to detect AI-designed variants—or synthetic homologs—of proteins of concern.”

The effort to enhance detection rates

In response to the identified vulnerabilities, the collaborative team worked to develop and deploy patches to improve the detection rates of biosecurity screening tools. Drawing from cybersecurity protocols, the researchers developed process patches that incorporated strategies meant to create tools that are better able to identify potentially dangerous synthetic homologs, including enhancements such as:

• Improved sequence alignment algorithms that compare query sequences against databases of known threats
• Function-based screening that considers predicted protein function—not just sequence similarity
• Machine learning that can identify the patterns and features associated with hazardous proteins even when there is high sequence similarity

The result? A significant improvement in detection rates, researchers reported—97% detection.

IDT’s role in improved biological screening

We are committed to leading the way with compliance with federal rules such as the BIOSECURE Act and in the development and adherence to biosecurity protocols. Existing IDT biosecurity measures include:

• Screening DNA orders and carefully vetting clients
• Limiting access to sensitive genetic materials, databases, and production facilities
• Adhering to national and international regulations governing the production, handling, and export of synthetic DNA
• Comprehensive employee training, physical security, data security, and incident response and containment

For this study, IDT and other DNA providers helped test AI-designed sequences to determine that some could escape detection in existing biological screening tools.

“This work shines light on an urgent need to develop community norms around AI biosecurity red teaming, coupled with rapid responses to biosecurity screening vulnerabilities as they are identified,” the study’s authors wrote. “Biosecurity screening would benefit enormously from development of supporting communities, institutions, standards, training, and practices like those found in the cybersecurity community. National governments have several critical roles to play, including incentivizing adherence to screening best practices, as the U.S. has done recently … Governments must also be proactive in funding studies, basic methods research, and ongoing capability assessments.”

To effectively counter the extant threats, the study recommends:

• Continuous monitoring and evaluation of biosecurity screening tools against new synthetic homologs
• Development of more sophisticated screening methods that can detect potentially dangerous proteins
• Collaboration and information sharing between researchers, biosecurity experts, and nucleic acid synthesis providers
• Ethical guidelines and responsible innovation, with a strong focus on biosecurity considerations

The future of protein engineering is exciting—and risky. By staying in front of the biosecurity challenges posed by AI, researchers can harness the power of AI to advance life sciences while also protecting us from potential threats.