Problem

Synthetic lethality (SL) is defined as a gene combination effect (GCE) of two genes where loss-of-function of both genes, rather than one, causes cell death. While carrying strong therapeutic potential in the treatment of cancer, SL interactions are context specific and sparse, thereby limiting in silico SL identifications to within same contexts. There is a need for better SL prediction models for an individual sample.

Solution

To leverage all SL datasets including existing CRISPR-based Double Knockout (CDKO) data and maximize information gain, researchers at OSU have introduced transfer learning to SL domain, namely TL-iGeneCombo and designed a new gene pair viability measure that improves on current SL scoring algorithms by being simple to measure simply named GCE using log fold change. Five transfer learning methods were developed and implemented, namely finetuning, freeze the first layer, freeze high parameter, Shapley and their ensemble.TL-iGeneCombo on both SL classification and GCE regression tasks was applied across four scenarios: datasets with (1) same genes and different cells, (2) same cell and different genes, (3) similar study mechanisms, and (4) big data to small data. An ensemble of four transfer learning methods was used for each scenario. Over 8,900 SL and 8,600 GCE TL scenarios, on average, TL-iGeneCombo improved SL classification task by 19% accuracy and GCE regression task by 29% Spearman correlation. The most effective TL approach was same genes and different cells, where SL prediction accuracy and GCE Spearman correlation improved by 20.9% and 38% respectively. The proposed transfer learning approach, TL-iGeneCombo, is highly informative and beneficial for selecting genes and cells in new CDKO experiments.

Applications

• Predict synthetic lethality of an individual sample as a companion diagnostic to improve cancer treatment outcome
• Optimize gene pair selection for CRISPR-based double knockout experiment and maximize the chance in identifying synthetic lethal gene pairs.
• Predict synthetic lethality for developing new, better, precision cancer treatment

Advantages

• Transfer learning improves prediction accuracy of synthetic lethality of an individual sample
• Transfer learning uses existing CDKO data to generate info to improve gene pair and cell selection for future CDKO experiments and biological validation

Seeking opportunities for out-licensing and collaboration