ITMO University Unveils Benchmark for AI in Bioinformatics

Artificial intelligence can now generate molecules, but how can researchers tell whether an algorithm is getting them right? Scientists at ITMO University have developed a way to address that question.

A Test for the Invisible World

Aptamers are short DNA or RNA molecules that function like highly precise molecular locks. Their ability to bind selectively to targets makes aptamers valuable for developing biosensors, diagnostic tests and targeted drugs. Neural networks analyze an aptamer's sequence and the properties of its target to predict the strength and specificity of their binding. This allows researchers to eliminate less promising candidates early in the process.

For years, neural networks were trained on disparate datasets, making it extremely difficult to assess their real-world accuracy. AptaBench addresses this weakness with a reference dataset containing 6,289 experimentally validated interaction pairs. Algorithms can now be tested on their ability to predict the behavior of previously unseen aptamer families.

From Generation to Validation

To understand the scale of the shift, it helps to look back at the development of Russia's AI for Science field.

In 2021 and 2022, industry and government were still testing the waters: AI was being used for initial screening, while Russia's Ministry of Health was launching platforms for working with medical datasets.

By 2023 and 2024, Skoltech, ITMO University and startups such as Ligand Pro had moved on to sophisticated multitask learning and virtual screening, reducing the number of exploratory laboratory experiments.

The turning point came in 2025, when Innopolis University and VNIIA introduced a full-cycle AI platform for drug generation and optimization. But powerful generative models come with a downside: the more molecules a neural network generates, the greater the risk of a serious error. In biomedicine, an algorithmic mistake can mean millions of rubles lost and years of research wasted.

That makes 2026, marked by the release of AptaBench, a logical and critical next step. Russia has now moved beyond developing individual algorithms toward building a rigorous infrastructure for evaluating their quality.

Benefits for People and the Country

Benchmarks like these can directly affect how quickly new drugs and more precise diagnostic methods are developed. AI models rigorously evaluated with AptaBench could eliminate candidates unlikely to work before they reach the stage of costly in vitro experiments.

For Russia, this is also a matter of technological sovereignty. Having a domestic tool for standardized evaluation lays the foundation for independent computational biology. It creates a Russian validation pipeline: AI generates a molecule, AptaBench tests the prediction, a laboratory validates the most promising candidate, and the process can ultimately produce an innovative drug. This reduces reliance on foreign databases and evaluation methods.

A Global Standard and an Infrastructure of Trust

Does this matter for science beyond Russia? The reliability of AI in biology is a global challenge. The Russian-developed benchmark has the potential to become an international standard for evaluating models used to create new biomolecules. Exporting such IT tools, providing access to unique databases and collaborating with research hubs abroad are promising directions, although they would require independent international validation.

AptaBench is more than a specialized scientific tool. It is an important component of the infrastructure needed to build trust in artificial intelligence in a field where human health and lives are at stake. The era of medical “black boxes” is giving way to evidence-based algorithmic science, in which neural-network predictions can be subjected to rigorous mathematical and biological scrutiny.

Many interaction-prediction models are evaluated under relatively simple conditions, where the training and test data contain similar sequences and molecules. In real-world applications, however, we are often interested in a completely new molecule or aptamer, such as a sequence that has just been generated by another model. We wanted to understand how well the results of standard testing hold up under these conditions. It turned out that the difference can be quite substantial. That is why it is important to evaluate not only a model's accuracy but also the limits of its applicability

Maria Eremeeva

Researcher and Engineer at ITMO University's Center for AI in Chemistry and ChemBio Cluster