Machine Learning Prediction of Antimicrobial Response in Pleurotus ostreatus Extracts Cultivated on Cassava Peel: A Proof-of-Concept Study

Imagine the world of tiny, invisible invaders—bacteria and fungi—that sometimes make us sick. For a long time, we've had special weapons called antibiotics to fight them, but these invaders are getting clever. They are learning how to dodge our weapons, a problem scientists call "antimicrobial resistance." It's like a game of hide-and-seek where the hiders are getting better at hiding, and we need to find new ways to catch them. One promising idea is to look at nature's own pharmacy, specifically at mushrooms. Some mushrooms, like the oyster mushroom (Pleurotus ostreatus), are famous for making chemicals that can stop these bad bugs in their tracks. Even cooler, these mushrooms can grow on things we usually throw away, like the peels from cassava roots. But here's the tricky part: testing every single mushroom batch against every single germ is slow, expensive, and takes a lot of hard work. Scientists are wondering if they can use a "crystal ball" made of math and computers—called Machine Learning—to guess which mushroom extracts will work best before they even start the messy lab tests.

This paper is a "proof-of-concept" study, which is a fancy way of saying, "Let's try a small test to see if this crystal ball idea is even possible." The researchers grew oyster mushrooms on cassava peels and then made two types of liquid extracts from them: one using alcohol (ethanol) and one using water. They tested these liquids against seven different types of germs to see how well they stopped the germs from growing. Instead of just looking at the results, they fed all the data into a computer program called a "Random Forest" model. Think of this model as a super-organized detective that looks at clues—like what kind of liquid was used to make the extract, what kind of germ it's fighting, and the chemical makeup of the mushroom batch—to predict the outcome.

The detective did a decent job at one specific task: predicting the size of the "zone of inhibition." Imagine drawing a circle on a petri dish where the germs can't grow; the bigger the circle, the better the mushroom extract works. The computer model could guess the size of this circle with moderate success, getting an accuracy score (R-square) of 0.68. It was pretty good at guessing that bigger circles would happen with certain conditions. The most important clue the detective found was the extraction solvent. Whether the extract was made with alcohol or water mattered way more than the tiny chemical differences between the mushroom batches. In fact, the specific chemical makeup of the mushrooms contributed very little to the prediction.

However, the detective hit a wall when trying to predict something else: the Minimum Inhibitory Concentration (MIC). This is a measure of how much of the extract is needed to stop the germ completely. When the computer tried to guess if the extract was strong enough to be a "winner" or a "loser," it performed poorly, getting the answer right only 43% of the time. It was basically guessing in the dark. This tells us that the clues the scientists had—like the batch chemistry and the type of germ—weren't enough to tell the difference between a strong extract and a weak one.

So, what's the verdict? The study suggests that machine learning can be a helpful sidekick for exploring mushroom medicines, but it's not a magic wand yet. The computer learned that the type of liquid used to make the extract is a huge deal, but it couldn't figure out the finer details of how strong the medicine would be. The authors are careful to say these results are just the beginning. Because they only tested a small number of samples (42 observations) and just three batches of mushrooms, the picture is still a bit blurry. To build a truly reliable crystal ball, they say we need much bigger datasets, testing many more types of mushrooms and germs, with even more detailed chemical measurements. For now, this study is a fun, promising first step that shows the potential of mixing biology with computer science, even if the computer still needs a lot more homework before it can replace the lab bench.