Fig. 1

Building a machine learning algorithm for the accurate prediction of novel functional lncRNAs. a Workflow for designing the INFLAMeR (Identifying Novel Functional LncRNAs with Advanced Machine learning Resources) algorithm and selecting targets. Upper: The algorithm was trained on previous high-throughput pooled CRISPR interference (CRISPRi) screening data from three previous high-throughput CRISPRi screens [24,25,26]; the algorithm was computed based on a total of 71 features comprising ENCODE ChIP-Seq transcription factor (TF) binding data for the regions surrounding lncRNA promoters [49,50,51] and genomic features. Lower: targets predicted to be functional by INFLAMeR in the K562 cell line (INFLAMeR score > 0.5) were selected for validation after excluding lncRNAs that were functionally characterized in previous studies, those with low expression, those not annotated in the Ensembl database, and those neighboring a protein-coding (PC) gene; thirty-nine lncRNAs were selected for validation. b INFLAMeR was built using an XGBoost classifier and its performance was calculated using a receiver operating characteristic (ROC) curve. Black ROC curve shows the mean classifier performance on the test set using three randomized seeds (red, green, and yellow curves). c Confusion matrix based on the test set. Percentages in the confusion matrix are row-normalized. d Local explanation summary of the impact of the top twenty features on the INFLAMeR score. Each dot represents one lncRNA. Red indicates a higher feature value (e.g., larger transcription start site (TSS) protein-coding (PC) distance or higher expression), and blue indicates a lower feature value (e.g., smaller TSS PC distance or lower expression). Higher (red) feature values with a positive SHAP value indicate a positive correlation, and lower (blue) feature values with a positive SHAP value indicate a negative correlation