AI has just surpassed millions of years of evolution. This will be a revolution in medicine.

 

Integra Therapeutics needed Large Language Models (LLMs)

to design novel, highly efficient gene-editing proteins that surpass the capabilities and limitations of naturally occurring ones. Traditional protein-engineering methods were too slow and restricted to existing natural proteins, hindering the development of advanced therapies.

The company's use of protein LLMs addressed several major challenges in developing its gene-writing platform, FiCAT:

The limitations of natural proteins

Integra's gene-editing platform, FiCAT, uses PiggyBac transposases—enzymes that can cut and paste DNA sequences—to insert new genes into a specific, safe location in the genome.

 

    Performance ceiling: Naturally occurring PiggyBac transposases, or variants modified through traditional engineering, could only be optimized to a certain point.

    Non-specific insertion: Conventional gene-editing vectors can insert genes randomly, which is a safety concern for therapeutic applications. To make its FiCAT platform more precise, Integra needed to engineer entirely new, programmable enzymes.

 

The challenge of de novo protein design

Before AI, designing new proteins "from scratch" was a major hurdle in computational biology.

 

    Vast design space: There is an almost infinite number of possible amino acid sequences for proteins. It is computationally impossible to explore and evaluate all these possibilities using traditional methods.

    Time and labor: Traditional approaches relied on laborious and time-consuming trial-and-error experimentation, often starting with existing natural proteins and making small, incremental changes.

 

How LLMs solved these problems for Integra

 

By training a protein LLM on vast databases of known protein sequences, Integra taught an AI the "grammar" or underlying principles of functional protein design.

 

This provided three major advantages:

 

    Accelerated discovery: Instead of working within the confines of natural evolution, the LLM could be used to generate entirely new, synthetic protein sequences from scratch. This ability significantly accelerated the discovery of novel proteins.

    Expanded functional diversity: The AI generated PiggyBac transposase variants with enhanced activity that outperformed the best versions found in nature. This expanded the potential of Integra's technology beyond what was naturally available.

    Improved therapeutic compatibility: The AI-designed transposases were created for enhanced compatibility with Integra's FiCAT platform, paving the way for more efficient manufacturing of engineered cell therapies.

 

This new way is met with excitement in Polish media. (Lithuanian journalists are now busy with one and only one question: Whom belongs Crimea? Who will answer this question according the propaganda of Lithuanian rulers, will become the Minister of Culture in Lithuania next week):

"Scientists have achieved a breakthrough by using artificial intelligence to design synthetic proteins. These outperform their natural counterparts. There's already talk of a 'paradigm shift' in genetic engineering."

 

Spanish researchers, harnessing the power of generative AI, have created synthetic genome-editing proteins whose activity and precision surpass their natural counterparts, shaped by millions of years of evolution. This extraordinary discovery has just been published in Nature Biotechnology. Experts believe this achievement paves the way for more effective and affordable gene therapies. This promises breakthroughs in the treatment of cancer and rare diseases, among other things.

"Molecular scissors." What can they do?

 

This is a moment that experts are unhesitatingly calling a paradigm shift in genetic engineering. For the first time in history, scientists have demonstrated that artificial intelligence is capable of not only mimicking nature but also creating "biological tools" superior to those developed through evolution. This breakthrough was achieved by researchers from Integra Therapeutics, who—in collaboration with Pompeu Fabra University in Spain and its Center for Genome Regulation (CRG)—used large language models (LLMs) to design entirely new, so-called hyperactive proteins. To visualize this discovery, imagine "molecular scissors" capable of cutting and pasting DNA fragments in human cells. These AI-created enzymes have demonstrated significantly greater efficiency and precision than their natural variants in laboratory studies. This solves one of the key problems that has so far limited the development and availability of advanced gene therapies.

 

Before the AI ​​could get to work, however, it needed data. The research team conducted an unprecedented computer-aided bioprospecting analysis, searching over 31,000 eukaryotic genomes. As a result, over 13,000 previously unknown sequences were discovered, and after verification in human cells, the 10 most active ones were selected, two of which matched the performance of versions previously optimized in laboratories.

 

This vast and unique dataset was used to train the AI ​​models. As Dr. Marc Güell, Scientific Director at Integra Therapeutics, notes, genAI was used for the first time to create "synthetic elements and extensions of nature."

 

The proteins designed by the algorithms not only retained their structural integrity but also proved more compatible with modern gene editing platforms. One variant demonstrated exceptionally strong activity in human T lymphocytes – cells crucial for the development of groundbreaking immuno-oncology therapies such as CAR-T.

 

To date, protein engineering has primarily involved the painstaking modification of existing, natural structures. Designing with the aid of AI allows for the creation of entirely new molecular tools, transcending the limitations imposed by evolution and endowing them with therapeutically desirable characteristics. This offers hope for more effective treatment, improved production, and reduced therapy costs.

 

A few weeks ago, Integra Therapeutics received nearly €11 million from the European Commission for research development. Importantly, more and more companies are considering this type of AI application – Profuent Bio, for example, is exploring this potential and is already achieving success with OpenCRISPR-1, a gene editor designed by AI (it demonstrates 95% fewer unintended side effects). The recent launch of Google DeepMind's AlphaProteo platform confirmed the growing importance of this trend. Analysts predict that the value of the AI-assisted protein design market will jump from its current level of $1.5 billion to $7 billion in 2033.”