Deep integrative omic auto-encoder analysis
||Deep integrative omic auto-encoder analysis|
||Mika Gustafsson <firstname.lastname@example.org>|
||2023-06-29 – 2024-01-01|
One of the research lines our group is currently working with concerns the implementation and optimization of deep learning approaches for the discovery of biological signals within the compressed representation of multi-omic data (Dwivedi et al., Nat Comm 2020). Due to the extensive sample sizes required for these purposes, from the tens of thousands to the millions of entries, plus the training times for the different architectures that can be tested, having a reliable access to high-level computational power is desired. Presently, we have been able to pre-process the inputs and deploy and test our workflows in the clusters Kebnekaise, Sigma and Tetralith, as well as use two local machines (4 x 2080 Ti GPUs) for early-stage pipeline development. However, the mentioned systems are severely limited in speed and memory, preventing the testing of models of higher complexity and increased sample size. Regarding this aspect, Berzelius is optimal due to both RAM and GPU availability. We would therefore expect a substantial speed up of the model training convergence and performance evaluation.