holmes

An AWS Steps based service that does bioinformatics fingerprint checks, deployed as a CDK pipeline project.

Overview

The somalier tool is a useful tool for generating genomics fingerprint files - taking a BAM file and producing a much smaller representation of variance at set locations throughout.

These files can then be compared to each other and rated with a 'relatedness' score. Where genomics files are samples from the sample human, or close relatives - this score is high, and therefore the relatedness score can help guard against sample mix-ups - by uncovering where unexpected relationships exist between samples.

Holmes is a low-cost (activity only costing - almost no cost as rest) tool that can be interacted with via AWS API (Lambda and Steps) OR via Slack commands. It will perform a variety of somalier calls over a large database of fingerprints.

Developers

Before doing any development work - please see here for dev setup instructions.

Service (as API)

See here

Service (as Slack command)

See here

Deployment

The stack does not create the fingerprint bucket. Instead this should be created manually before installing Holmes.

(in the past this bucket was created in the Holmes stack but that prevented deleting the CDK entirely - so instead now the Holmes stack is entirely stateless and the bucket needs to be made separately)

Below is the previous definition. The only main thing of note is the two lifecycle rules that clean up data. To be honest, it doesn't particularly matter if they are not present, just the bucket will fill unnecessarily.

new Bucket(this, "FingerprintBucket", {
  bucketName: props.fingerprintBucketName,
  objectOwnership: ObjectOwnership.BUCKET_OWNER_ENFORCED,
  lifecycleRules: [
    // we give the test suites the ability to create folders like fingerprints-test-01231432/
    // and we will auto delete them later
    {
      prefix: "fingerprints-test",
      expiration: Duration.days(1),
    },
    // space for us to make temp file results from the DistributedMap
    {
      prefix: "temp",
      expiration: Duration.days(1),
    },
  ],
  removalPolicy: RemovalPolicy.RETAIN,
});

Costing

Estimates are available here. They have been shown in practice to be roughly correct.

Design

The service maintains an S3 bucket that stores fingerprint files (~200k per BAM) and then provides AWS Steps/Lambda functions that operate to run somalier over these files.

  graph TD;
      subgraph bams living externally in any location readable by the service
          bam1("gds://spot/bam1.bam")
          bam2("gds://otherspot/bam2.bam")
          bam3("s3://bucket/bam3.bam")
      end
      subgraph S3 fingerprints bucket
        subgraph S3 folder ABCDEF/
          f1("ABCDEF/encoded URL of BAM1")
          f2("ABCDEF/encoded URL of BAM2")
        end
      end
      bam1-->f1
      bam2-->f2
      bam2-->f3

The operations provided by the service are focussed around a) producing new fingerprints b) checking fingerprints against others

There is no other data store for the service - the existence of a fingerprint in S3 with a path matching the sites checksum and BAM URL (encoded) is the canonical definition that a BAM has been fingerprinted.

The check operation will always operate against all fingerprints that exist in the designated fingerprint folder.

Algorithm

The details of how somalier scores are used is documented here.

umccr/holmes