1 of 11

Transcriptomics Atlas Use-Case

Piotr Kica, Kamil Burkiewicz, Jan Przybyszewski, Sabina Lichołai, Maciej Malawski

Sano Centre for Computational Medicine

Krakow, Poland

EXTREME NEAR-DATA PROCESSING PLATFORM

2 of 11

Description of the use-case

Transcriptomics Atlas Processing high amount of RNA-seq with a resource-intensive pipeline to create the Atlas useful for scientists and researchers.

Federated Learning Training a joint machine learning model within a distributed workflow without training data transmission, which increases privacy and GDPR compliance.

Serverless processing - DataPlug integration Transcriptomics pipeline for producing a statistical model of transcriptome with serverless functions and DataPlug integration

Transcriptomics Atlas Pipeline

3 of 11

Why extreme data?

  • The Transcriptomics Atlas pipeline will process over 130TB of FASTQ data related to 19 human tissues.

​

  • The final Atlas can be extended with additional human tissues further increasing the processed FASTQ data amount.

​

  • Existing human genome variation datasets that we focus on in FL experiments at petabytes of data. This volume is still growing and will grow ever faster with the advent of new, high-throughput data collection methods.

Transcriptomics Atlas architecture

4 of 11

Datasets

  • NCBI Sequence Reads Archive repository.  We select a dataset of RNA-sequences of human tissues. www.ncbi.nlm.nih.gov

​

​

​

  • For FL experiments, we generate a substantial dataset based on the sim1000G tool. The tool is based on the 1000 Genomes Project database.

​

Transcriptomics - Sano

5 of 11

Components integrated

  • Scone - allows to securely execute the FL workflow using untrusted infrastructure using Trusted Execution Environments.

​

  • Lithops – enables running part of Transcriptomics Atlas in serverless environment using AWS Lambda.

​

  • DataPlug - partitions input files for transcriptomics computations to achieve higher parallelism.

​

Transcriptomics - Sano

Lithops and DataPlug integration into serverless pipeline

6 of 11

Description of the use-case

  • We developed a Federated Learning framework designed to carry out experiments on transcriptomic datasets
  • The framework is based on Flower FL library and uses PyTorch for machine learning models training implementation
  • We integrated the framework with HPC, Microsoft Azure and Kubernetes execution environments
  • The framework will be integrated with SCONE to improve security

Transcriptomics - Sano

7 of 11

KPIs

KPI-1: Significant performance improvements

KPI-3: Demonstrated resource auto-scaling for batch processing.

​

KPI

Optimization

Impact

KPI-1

STAR index

generated on newer release

~12x faster STAR 

alignment time

~8.4x reduced costs

KPI-1

STAR Early Stopping feature

Significant STAR pipeline 

throughput increase: 19.5%

KPI-3

Cost-efficient 

instance type

~15% cheaper and ~30% faster than alternative type

KPI-3

STAR Index on NFS

Fast and scalable

 solution for index distribution problem

KPI-3

Spot instances usage

Reduced compute

 costs by > 50%

STAR alignment execution time reduction due to early stopping feature

Transcriptomics - Sano

8 of 11

KPIs

KPI-4: High levels of data security and confidential computing validated using TEEs and federated learning in adversarial security experiments.

KPI-5: Demonstrated simplicity and productivity of the software platform, validated with real user communities in International Health Data Spaces.

KPI

Optimization

Impact

KPI-1

Data transfer reduction thanks to FL nature

Data transfer reduction of 80% compared to Centralized Learning (CL) approach

KPI-1

Improved data ingestion rate in FL workflow due to parallel processing

ETL phase of the FL workflow executes 75% faster compared to CL approach

KPI-4

Increased regulatory compliance and data security due to FL nature

GDPR compliance due to data minimization principle in FL – high practicality

KPI-5

Increased privacy due to FL nature

​

No sensitive data transmitted even if existing in training datasets - increased privacy

STAR alignment execution time reduction due to index generated on newer genome release

9 of 11

Large scale experiment

Experiment setup:

  • 7126 files (18TB of SRA and 130TB of FASTQ)
  • Max 50 instances of r7a.2xlarge type
  • Index - release 111

Optimizations:

  • Spot instances -> 50% cost reduction
  • Early Stopping -> 23% total STAR alignment reduction
  • Newer index release -> 12 times faster STAR alignment time and 8.4 times reduced costs

Outcome:

  • Successfully processed 99% of input dataset�in 8820 core-hours
  • 0.066$ per FASTQ file
  • Further optimizations possible

Timeline of a large scale experiment.

10 of 11

KPIs

KPI

Optimization

Impact

KPI-1

Serverless functions performance improvements

~10x faster Salmon pseudoalignment 

Transcriptomics - Sano

Speedup achieved with parallel execution

KPI-1: Significant performance improvements

KPI-3: Demonstrated resource auto-scaling for batch processing.

​

11 of 11

Transcriptomics - summary

  • Pipeline implemented and evaluated
  • FL framework available
  • Large computing campaign processed
  • Successful experiments with Lithops and DataPlug
  • „Novel Approaches Toward Scalable Composable Workflows in Hyper-Heterogeneous Computing Environments” - Bader J., ... Kica P., Lichołai S., Malawski M., ... , Mandal A. - WORKS 2023 at SC23, Denver, USA.
  • „Optimizing STAR Aligner for High Throughput Computing in the Cloud” - Kica P., Lichołai S., Orzechowski M., Malawski M. – poster accepted at IEEE Cluster 2024 conference, Kobe, Japan.

​

Transcriptomics - Sano

https://github.com/neardata-eu/transcriptomics-atlas-sano