
Blue Carbon: A Water Body as a Carbon and ESG Asset
What blue carbon is, how water bodies store carbon and what an ESG report gives. AVELIFE roadmap: pilot, measurement, methodology.
Artificial intelligence appears more and more often in reports on drug development. In antiviral research it can indeed cut down on manual work: quickly matching large sets of compounds, viral sequences and biological assay results. But an algorithm does not turn a prediction into a finished drug. The most accurate way to describe its role is as a tool for generating and prioritizing hypotheses that still have to pass laboratory, preclinical and clinical testing.
At an early stage, a research team may have hundreds of thousands or millions of molecules available. Testing them all experimentally is expensive and slow. Machine learning models rank the library by likely activity, toxicity or similarity to known active compounds. As a result, a smaller but more meaningful set goes on to the primary assay.
A well-documented example is the work of NCATS researchers. A model trained on biological activity profiles proposed 311 candidates against SARS-CoV-2; in a live-virus assay, activity was confirmed for about 32% of the selected compounds. That is a strong result for the prioritization stage, but the authors call them potential leads, not drugs: most needed further work on mechanism, selectivity and development. Details and limitations are in the original NCATS publication.

Another direction is combining chemical features with data on viral genomes. In a 2025 study, models predicted the selective activity of compounds, after which 346 predictions were tested experimentally. In the entry assay, 24 of 256 compounds proved active, and in the RNA-dependent RNA polymerase assay, 47 of 128. The difference between 9.4% and 37% shows something important: the result depends not only on the algorithm but also on the task, the quality of the training data and the specific biological assay. The full design is described in the open-access original article in Communications Chemistry.
Structural models also help find pockets in viral proteins, predict how a molecule fits into them and compare conserved regions across related viruses. But a computed “good fit” does not guarantee that the molecule will enter the cell, remain stable, reach the required concentration and not harm the body.
In another computational-experimental project, a model screened about 700 thousand compounds and passed the 100 top-ranked to the laboratory. In Caco-2 cells, 19 showed activity, and eight showed dose-dependent activity in Vero cells. The researchers stressed that these are in vitro results; they do not prove efficacy or safety in humans. This very move from a large library to dozens of verified signals is the real, measurable benefit of the algorithm. Source: the original study on deep-learning profiling.

| Stage | Useful role of the algorithm | What needs to be confirmed |
|---|---|---|
| Target selection | Finding patterns in genomic, structural and CRISPR data | Whether the target is essential for replication and accessible to intervention |
| Virtual screening | Ranking large libraries | Activity, selectivity and reproducibility in a biological assay |
| Lead optimization | Predicting properties and suggesting structural changes | Pharmacokinetics, metabolism, toxicology and synthesizability |
| Resistance prediction | Assessing the consequences of possible mutations | Viral fitness, clinical sequences and phenotypic assays |
A 2026 review groups the uses of AI into four areas: target selection, repurposing, design of new molecules and resistance prediction. At the same time, the authors stress that for many studies the evidence stops at the in silico or in vitro level, and there are few in vivo data. This is an important caveat for any loud claim about an “AI-discovered drug”. See the review in Drug Discovery Today.
A model inherits the limitations of its data. Results from different laboratories may be incompatible; “inactive” compounds were often tested in a different cell line or under a different protocol. If closely related chemical structures dominate the dataset, the prediction will work worse outside familiar space. Additional risks come from data leakage between training and test sets, opaque ranking and optimizing only one metric.
Even a correct binding prediction does not cover a drug’s whole journey. Solubility, stability, tissue penetration, interactions with other molecules, genotoxicity, effects on organs and the risk of rapidly emerging resistance must all be checked separately. In the NIH Antiviral Summit report, AI is treated as a tool for the early stages, but alongside synthesis, cytotoxicity tests and confirmation of viral suppression.
The useful questions are simple: was there an independent test set; how many predictions were actually synthesized or purchased; what share was confirmed; is there a dose-response; was toxicity measured; was the result repeated in another cell model; are the code, data and negative results available. For a mature project, what matters is not the size of the virtual library but the number of reproducible experimental leads and the quality of the transition between stages.
NIH created nine AViDD centers as a multidisciplinary system in which computational approaches work together with structural biology, medicinal chemistry and virology. The aim of the program is to build series of leads and candidates against virus families with pandemic potential, not to replace experiment with prediction. The official description is available in the NIH announcement on AViDD.
AI speeds up the search where a huge space of hypotheses has to be put in order and the best experiments chosen. It does not yet remove a single critical barrier between a prediction and a safe medicine. The most convincing result is not a picture of a molecule on a screen but a transparent chain from data to independent laboratory confirmation, preclinical assessment and properly planned clinical trials.
No. It can suggest or rank molecules, but activity, mechanism, safety and clinical benefit are established only by successive experiments.
A prediction exists in a computational model. A lead has confirmed biological activity and sufficient properties to start optimization, but it is not yet a finished drug.
There is no single universal metric. What matters is independent validation, the share of experimentally confirmed predictions, the diversity of the structures found and reproducibility across several relevant assays.
It can help with analysis and planning, but it does not replace preclinical studies or the phases of clinical assessment of safety and efficacy.

What blue carbon is, how water bodies store carbon and what an ESG report gives. AVELIFE roadmap: pilot, measurement, methodology.

How silica sorption, biochar and bacterial strains purify water: sorbent, biosurfactant, sediment back to soil. AVELIFE bioactivator technology.

AVELIFE water body restoration complex: survey, water treatment, sediment work and an ESG report. Kyiv, 2025. Request a pilot.