Russian specialists have developed the SAASBench tool to assess the accuracy of AI models used in the development of antibody-based drugs. The system allows checking whether artificial intelligence can correctly identify the target protein of an antibody among several similar proteins and comparing the effectiveness of various computational models even before laboratory experiments begin. The development was created by specialists from Sberbank's Center for Practical Artificial Intelligence and the AIRI Institute, Sber's press service reported.
Antibodies are immune system proteins that recognize specific molecules called antigens. They are used to create modern drugs, including those for treating oncological, autoimmune, and infectious diseases.
An antibody can be thought of as a very precise "search engine" for a specific protein. If there are several similar proteins in the body, it must find the exact one needed. The new tool checks whether AI can correctly identify such a pair.
We decided to test artificial intelligence on a pragmatic question: will it find among several similar proteins the one that the antibody should work with? For this, we developed SAASBench – a tool for comparing AI models at the early stages of antibody creation. It does not replace the laboratory but helps weed out unreliable computational approaches.
One of the main indicators in such a selection is specificity, i.e., the ability of an antibody to bind precisely to the desired protein. This is fundamentally important. Even if an antibody binds well to the target molecule, it can simultaneously attach to other, similar proteins. This reduces the accuracy of diagnosis and treatment effectiveness, and the risk of side effects may increase.
SAASBench helps verify this aspect for AI-generated antibody sequences and understand which computational tools provide reliable results and which are best avoided.
Комментарии