PARNAS

PARNAS
Name	PARNAS
Type	Algorithmic framework
Developer	Various research groups
First release	2010s
Latest release	2020s
Written in	C++, Python, TensorFlow, PyTorch
Operating system	Cross-platform
License	Open-source / Proprietary variants

PARNAS PARNAS is an algorithmic framework used in computational biology, bioinformatics, and data analysis for representative selection and partitioning tasks. It has been developed in the context of phylogenetics, clustering studies, and sampling design, and has been applied alongside tools from machine learning, statistical inference, and high-performance computing. The framework has influenced implementations in academic projects affiliated with institutions such as Harvard University, Stanford University, Massachusetts Institute of Technology, University of Oxford, and University of Cambridge.

Introduction

PARNAS addresses problems related to selecting representative subsets from large datasets produced by platforms such as Illumina, Oxford Nanopore Technologies, and PacBio sequencers, and has been referenced in work involving model selection in Hidden Markov Model studies, Bayesian inference pipelines, and Maximum Likelihood phylogenetic reconstructions. Researchers working with datasets from projects like Human Genome Project, 1000 Genomes Project, The Cancer Genome Atlas, and UK Biobank have integrated PARNAS-style approaches with software including RAxML, IQ-TREE, BEAST, and FastTree.

History and development

Development of PARNAS occurred amid advances in computational phylogenetics and algorithmic selection methods emerging alongside algorithms such as k-means, k-medoids, facility location problem heuristics, and approximation schemes inspired by Greedy algorithm analyses. Early methodological roots trace to combinatorial optimization work at institutions like Princeton University and University of California, Berkeley, and were influenced by theoretical results from scholars affiliated with MIT and ETH Zurich. Implementations matured through collaborations with consortia such as Wellcome Sanger Institute, Broad Institute, and research projects funded by agencies like National Institutes of Health, National Science Foundation, and European Research Council.

Methodology and architecture

The methodology combines distance-based criteria from phylogenetic trees produced by programs such as FastTree and RAxML-NG with clustering objectives similar to those in Partitioning Around Medoids and facility location formulations. Architecturally, PARNAS implementations interoperate with libraries like SciPy, NumPy, Pandas, and deep learning frameworks such as TensorFlow and PyTorch when integrated into pipelines for Convolutional Neural Network feature extraction or Dimensionality reduction with t-SNE and UMAP. The architecture often leverages parallelism via MPI or OpenMP and containerization technologies like Docker and Singularity for reproducible workflows tied to platforms such as Galaxy (platform) and Nextflow.

Applications

PARNAS-style approaches have been applied to representative sampling in large-scale initiatives including GISAID sequence curation, outbreak investigation workflows exemplified by case studies involving SARS-CoV-2, Ebola virus epidemic, and Zika virus epidemic, and in biodiversity assessments linked to projects at Smithsonian Institution and Natural History Museum, London. Other applications include genotype imputation pipelines used in studies from Broad Institute and deCODE genetics, selection of prototypes for machine learning models in competitions like Kaggle, and integration into surveillance dashboards developed by organizations such as Centers for Disease Control and Prevention and World Health Organization.

Performance and evaluation

Performance evaluations compare PARNAS-based selection against baselines like k-means++, Greedy set cover, and exhaustive search on metrics used in studies from Nature Methods, Bioinformatics (journal), and PLOS Computational Biology. Benchmarks often involve datasets from GenBank, ENA (European Nucleotide Archive), or simulated data generated with tools like Seq-Gen and INDELible, measuring criteria such as representativeness, diversity, and computational cost on hardware ranging from academic clusters at Lawrence Berkeley National Laboratory to cloud platforms such as Amazon Web Services, Google Cloud Platform, and Microsoft Azure.

Limitations and challenges

Limitations include sensitivity to input distance metrics produced by tools like MAFFT and MUSCLE aligners, dependence on tree topology accuracy from programs such as IQ-TREE and PhyML, and challenges scaling to petabyte datasets generated in initiatives like Earth BioGenome Project. Additional challenges arise when integrating with privacy-preserving frameworks like GA4GH APIs and in handling biases documented in population studies such as those from All of Us Research Program and H3Africa.

Legal and ethical considerations

Legal and ethical considerations concern data sharing policies exemplified by agreements used by GISAID, consent frameworks seen in Human Genome Project follow-ups, and regulatory requirements handled by agencies including FDA and European Medicines Agency. Ethical debates overlap with topics in bioethics policy discussions at institutions such as UNESCO and Nuffield Council on Bioethics, especially regarding equitable representation in datasets from regions represented by World Bank and United Nations programs. Stakeholders include research funders like Wellcome Trust and publishers such as Nature Publishing Group and Oxford University Press that influence standards for reproducibility and data availability.

Category:Bioinformatics