A drug may affect a tiny genetic message by turning its activity up or down—but finding out which direction it takes can require slow, expensive experiments. This study asks whether a computer can make that prediction first. Changes in these tiny genetic messages, called microRNAs, have been associated with many human diseases.

Drugs that regulate them could therefore help treat diseases caused by disturbed biological pathways. Small molecules have shown strong potential as drugs that regulate microRNA expression. But determining whether they raise or lower that expression is slow and costly by experiment, and practical computer help has been missing.

DeepsmirUD was developed to predict the regulatory effect of a small molecule on microRNA expression, specifically whether it raises or lowers that expression. The researchers used microRNA and cancer relationships to build a database of potential pharmaceutical drugs covering 1343 small molecules and 107 cancer diseases.

They also used DeepsmirUD to predict regulatory effects for a large number of high-confidence associated small-molecule and microRNA relationships for further study. Checking whether a small molecule and a microRNA bind is normally time-consuming and costly, because researchers cannot test every possible pairing.

Only a small fraction of all possible pairs have been experimentally verified. Earlier computer methods mainly asked whether a pairing existed. The harder question—whether the small molecule turns the microRNA up or down—remained largely unexplored.

That missing direction matters because it can speed the search for evidence about microRNAs involved in cancer and treatment. The system uses twelve deep-learning frameworks to predict how small molecules regulate microRNA expression, with models selected from curated biological relationships.

The models were selected and trained using known pairings together with physical and chemical information about them. One combined model performed best on experimentally resolved relationships, while another group gave the most stable predictions during long training.

Combining the models reduced the influence of their individual errors and improved performance further. The predicted increases and decreases were then connected with disease information to suggest possible drugs. The system was then tested on relationships involving either a new microRNA or a new small molecule.

Its performance was unsatisfactory, especially when the small molecule was new. Changing how the test examples were selected produced the same observation, so the weakness was not explained by that sampling choice.

In practice, nearly all of the deep-learning methods struggled with new small molecules, while their performance was better when the new element was a microRNA. The map suggests that small molecules are more often predicted to raise microRNA activity than lower it: around two-thirds of the indirectly linked relationships point upward.

That matters because reversing disease-related microRNA changes could reveal how drugs work and suggest new treatment uses. The system predicted regulatory effects for 224 indirectly linked small-molecule and microRNA relationships selected from pharmacological information.

Around two-thirds of those relationships were predicted to increase microRNA expression. This extends the search beyond relationships already directly linked, using indirect evidence to propose additional pairings for study. The researchers describe two cases in which drug-like small molecules identified by the analysis had anticancer effects supported by other studies.

One approved antibiotic had been found to inhibit breast-cancer growth and spread, and another compound had been verified to suppress a type of skin cancer. Many highly ranked links between individual small molecules and several cancers were also predicted correctly in line with earlier studies.

For example, an orange-peel compound has been reported to inhibit cervical, lung, and colon cancers, matching the analysis’s negative connectivity scores. The researchers point to a practical limitation: published pharmaceutical studies may contain more verified small-molecule and microRNA relationships than the current collection captures.

A larger publication-based database could improve the computer models. That database is identified as an important direction for future work. The same approach might also be extended to other kinds of genetic messages. The system can help suggest how small molecules may change these genetic messages, but its weakest point is unfamiliar drugs.

That makes it a guide for experiments and possible drug reuse, not a replacement for testing.