Researchers have developed a machine learning method that draws on experiments in other cancer cells to improve predictions about which pairs of genes to investigate together.
The approach, called TL-iGeneCombo, addresses a bottleneck in cancer research: experiments that disable two genes at once can reveal useful interactions, but data for a cancer cell model are limited.
Published in Bioinformatics on October 10, the study used transfer learning to carry information between experimental datasets. The researchers reported a 19% improvement in synthetic lethality classification accuracy and a 29% increase in the Spearman correlation of predicted gene combination effect scores.
They then used findings from an initial experiment in the 22RV1 prostate cancer cell line to select gene pairs for a second experiment.
That follow-up exposed a distinction. It supported the method’s improved prediction of combined gene effects, but did not validate improved synthetic lethality prediction.
The work offers a way to prioritize laboratory experiments. It does not establish a cancer treatment or demonstrate benefits in patients. Its value lies in making better use of scarce experimental data before committing resources to testing.