New Deep Learning Tool Predicts Bioactivity, Identifies Therapeutics
In the COVID era, computational biology is having a heyday – and machine learning is playing a massive role. With billions upon billions of compounds to search through for any given therapeutic application, strictly brute-force simulations are wildly unfeasible, necessitating more artificially intelligent methods of whittling down the options. Now, researchers from IRB Barcelona’s Structural Bioinformatics and Network Biology lab have developed a deep learning method that predicts the biological activity of any given molecule – even in the absence of experimental data.
The researchers, led by Patrick Aloy, are applying deep machine learning to a massive dataset: the Chemical Checker, which provides processed, harmonized, and integrated bioactivity data on 800,000 small molecules and is also produced by the Structural Bioinformatics and Network Biology lab. In total, any given molecule has 25 bioactivity “spaces,” but for most molecules, data on only a few are known – if that.
Using the new deep learning tool, that’s changing. The Chemical Checker database contains data on all 25 bioactivity spaces from each of those 800,000 molecules, and the tool, having been trained on that data, can predict all the bioactivity spaces of any molecules with incomplete bioactivity data. “The new tool … allows us to forecast the bioactivity spaces of new molecules, and this is crucial in the drug discovery process as we can select the most suitable candidates and discard those that, for one reason or another, would not work,” explained Aloy.
Of course, the prediction isn’t perfect, and assessing molecules with more available data will allow the tool to produce higher-confidence predictions. Some molecules, as well, prove simply more or less difficult for the tool to assess. “All models are wrong, but some are useful,” said Martino Bertoni, first author on the paper describing the research. “A measure of confidence allows us to better interpret the results and highlight which spaces of bioactivity of a molecule are accurate and in which ones an error rate can be contemplated.”
The researchers chose a challenging case for validation: a cancer-related transcription factor that was broadly considered an “undruggable” target. The tool identified 131 compounds that fit the target by predicting their bioactivity spaces, and their ability to degrade the target was experimentally confirmed.
The research described in this article was published as “Bioactivity descriptors for uncharacterized chemical compounds” in the June 2021 issue of Nature Communications. The article was written by Martino Bertoni, Miquel Duran-Frigola, Pau Badia-i-Mompel, Eduardo Pauls, Modesto Orozco-Ruiz, Oriol Guitart-Pla, Víctor Alcalde, Víctor M. Diaz, Antoni Berenguer-Llergo, Isabelle Brun-Heath, Núria Villegas, Antonio García de Herreros and Patrick Aloy. To read it, click here.