CATALYSTS forest, sandy and peat soils, as well as water bodies.
These habitats and access to areas such as deep-sea habitats( in collaboration with Biodiscovery) ensures high diversity. All of the collected materials comply with the Nagoya Protocol and do not entail any third-party access or benefitsharing obligations.
The collected samples are brought to the laboratory, where the entirety of the DNA contained in a sample( the metagenome) is isolated. Drawing on our decades of experience in the extraction of genetic material, sequencing, and bioinformatics, we digitise those environmental samples and add them to a growing digital database.
This represents the first module of MetXtra and is a vital resource for enzyme sequence information. Currently, over 99 % of the sequences in the MetXtra database are new and cannot be found in this form in any public database. Additionally, the database is five times more diverse than the public database UniProt. It enables mining for novel enzymes and also extends publicly known sequence space.
Figure 1- Modules of MetXtra platform & discovery filter
MetXtra design: Newto-nature enzymes
To further expand the enzyme sequence space, we go beyond nature. After the breakthrough developments of AI tools for enzyme structure prediction, opportunities for AI-driven enzyme design arose. We use these models to access a broader diversity of sequences and scaffolds, while also using our enzymology expertise to evaluate predictions.
One approach to generating new-to-nature enzymes is to use AI models to predict sequences that fold into a desired threedimensional structure. The structure of an enzyme defines its properties. In this way, enzyme properties can be predetermined. Enzymes designed using this method consistently show elevated stability, rendering them highly attractive for industrial applications.
A second approach enables the design of enzyme sequences that exhibit the desired biocatalytic activity, as defined by a specific enzyme class. Unlike the first method, this approach uses a conditional language model that learns from
© BRAIN Biotech AG available enzyme sequences and annotations showing the target activity, rather than relying on structural information.
As with all data-driven processes, the results are only as good as the data that can be used for training. We use the MetXtra database as an additional training dataset to finetune the generative AI, enabling it to learn more broadly and deeply. With this additional source of information, the module can make unique suggestions for enzymes that fit the application.
MetXtra discover: Selecting the right enzymes for application
The third MetXtra module has been developed to identify the most promising enzymes from the first two. Bringing together sequencebased analyses, AI predictions, and structural bioinformatics, over 100,000 initial enzyme candidates can be screened in silico.
For each challenge, a unique combination of in silico approaches is selected to ensure tailored identification of enzymes. The applied methods range from large multiple sequence alignments( MSAs) and sequence similarity networks( SSNs) to AI-driven property predictions and modelling of hundreds to thousands of structures. These methods also include docking studies and simulations that validate enzyme performance.
This multi-factorial, step-wise filtering system streamlines the selection of the most promising 20-200 candidates for laboratory evaluation( Figure 1). It is important to consider not only the desired target reaction, but also compatibility with process conditions and application constraints. Industrial conditions for enzyme applications often differ significantly from those in nature. This can include the utilisation of new substrates, high substrate loading, elevated temperature, pH, cosolvents and further parameters that vary depending on the application.
JAN / FEB 2026 SPECCHEMONLINE. COM
39