MDB-47. DEEP LEARNING DECODES NEW SITES OF RNA TRANSLATION IN MEDULLOBLASTOMA

Abstract

Listen

Translate article

Abstract BACKGROUND Medulloblastoma, a malignant brain cancer predominantly affecting children, possesses a high propensity for metastasis and is frequently fatal. Through studies on RNA translation, a core biological process, we aim to uncover the disease biology of cancer in order to design therapeutic targets. Yet, the exploration of canonical protein-coding genes has thus far yielded few viable targets in medulloblastoma, where non-canonical open reading frames (ORFs) are increasingly being investigated for their regulatory role. METHODS To address the significant challenge in delineating novel translated ORFs in individual cells or tissues, we have designed RIBO-former, a best-of-its-class machine learning model that leverages ribosome profiling data to detect translated open reading frames. Through RIBO-former, we expanded the proteome of healthy prenatal (n=30) and adult (n=42) samples, which are notoriously hard to process due to low read depth and noise, in addition to medulloblastoma cancer tissue (n=8) and cell line (n=15) data. RESULTS Across all medulloblastoma cell lines, we determine an initial selection of 3,638 non-canonical ORFs. Unsupervised clustering of the ribosome profiling read counts within these non-canonical ORF regions group samples according to the described subtypes of medulloblastoma cancer, hinting at the significant role these ncORFs play in protein homeostasis. Looking into distinct expression profiles between cell lines with high and low MYC expression, we revealed 201 non-canonical ORFs that are differentially expressed. We evaluate our data through various analyses including testing of which ORFs have a viable sequence context for translation, achieved through another in-house tool (TIS Transformer). CONCLUSIONS By taking advantage of advances in machine learning, we were able to uncover new sites of translation in medulloblastoma that can aid the future discovery of therapeutic targets. As such, we were able to find 22 highly promising non-canonical ORFs that show unique expression profiles in relation to MYC in medulloblastoma.