ADAM: Genomics Formats and Processing Patterns for Cloud Scale Computing

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission. Executive Summary Current genomics data formats and processing pipelines are not designed to scale well to large datasets. The current Sequence/Binary Alignmen-t/Map (SAM/BAM) formats were intended for single node processing [18]. There have been attempts to adapt BAM to distributed computing environments, but they see limited scalability past eight nodes [22]. Additionally, due to the lack of an explicit data schema, there are well known incompatibilities between libraries that implement SAM/BAM/Variant Call Format (VCF) data access. To address these problems, we introduce ADAM, a set of formats, APIs, and processing stage implementations for genomic data. ADAM is fully open source under the Apache 2 license, and is implemented on top of Avro and Parquet [5, 26] for data storage. Our reference pipeline is implemented on top of Spark, a high performance in-memory map-reduce system [32]. This combination provides the following advantages: 1) Avro provides explicit data schema access in C/C++/C#, Java/Scala, Python, php, and Ruby; 2) Parquet allows access by database systems like Impala and Shark; and 3) Spark improves performance through in-memory caching and reducing disk I/O. In addition to improving the format's cross-platform portability, these changes lead to significant performance improvements. On a single node, we are able to speedup sort and duplicate marking by 2×. More importantly, on a 250 Gigabyte (GB) high (60×) coverage human genome, this system achieves a 50× speedup on a 100 node computing cluster (see Table 1), fulfilling the promise of scala-bility of ADAM. The ADAM format provides explicit schemas for read and reference oriented (pileup) sequence data, variants, and genotypes. As the schemas are implemented in Apache Avro—a cross-platform/language serialization format—they eliminate the need for the development of language-specific libraries for format decoding/encoding, which eliminates the possibility of library incompatibilities. A key feature of ADAM is that any application that implements the ADAM schema is compatible with ADAM. This is important, as it prevents applications from being locked into a specific tool or pattern. The ADAM stack is inspired by the " narrow …