2020 STAT115 Lect1.2 Big Data Challenge

Main Ideas, Concepts, and Lessons

1) Growth of sequencing scale and automation (2001 → present)

• Around 2001: Sequencing was large-scale, labor-intensive, and highly structured.
  • A genome center could have multiple rooms:
    • Sample-prep room: ~35 people, preparing samples for 3–4 weeks to get DNA into a sequencing-ready form.
    • Sequencing room: about 74 large sequencing machines.
  • Operations / logistics:
    • ~10 people operated the machines.
    • Each instrument could run 15–40 runs per day.
  • Data output (per day):
    • Each instrument produced about 1–2 megabytes/day (as stated in subtitles).
    • The entire room could produce about 120 megabytes/day.
• Key improvement: Modern sequencing is highly automated and dramatically better, but the major revolution occurred in 2007 with massively parallel sequencing.

2) Modern sequencing workflow: faster sample prep + long-running machines

• 2007+ workflow changes:
  • Sample preparation: reduced to something like a station-based setup, allowing one person to prepare samples in under a day.
  • Sequencing machines:
    • The same operator can load samples and then leave.
    • Machines run for 3–5 days.
    • Each instrument can generate roughly ~0.5 gigabytes/day per machine (as described).
  • The process is described as a flow path:
    • load DNA on one part of the system,
    • DNA flows through,
    • sequence data comes out directly from the instrument.

3) “Mini machines” vs sequencing facilities: different scales, output, and reads

• Individual lab / mini machines (example given):
  • Can run at least a full day.
  • Produce on the order of millions of reads per day (e.g., “4 million” or “25 million reads/day,” as stated).
  • Reads are typically paired-end with ~150 base pairs per read (paired structure emphasized).
• Sequencing facility / core facility (bigger scale):
  • Machines run 1–2 days.
  • Example output:
    • ~300 gigabytes in 1–2 days
    • about 1 billion reads from one run
    • paired-end, ~150 base pair reads
  • Mention of large institutions having such core facilities (e.g., Harvard-related and cancer institute-related facilities).

4) Very large “mega” sequencing runs (flow cells, capacities, and genome throughput)

• The “mega” sequencer is described as being machine-sized like a copier, but with extremely high throughput.
• Continuous operation idea:
  • It can run two runs at a time by alternating flow cells, keeping the machine running much of the time.
• Flow cell / read-length flexibility:
  • Output changes depending on sequencing length settings (shorter vs longer read lengths).
• Data scale (as stated):
  • In dual flow-cell mode, up to ~6000 gigabytes of data per certain unit time/run configuration (subtitles mention “as high as six thousand gigabytes of data”).
• Examples connecting read length to time and output:
  • Sequencing length (e.g., ~160 bp) relates to runtime:
    • mentioned: ~24 hours for 160 bp
    • longer lengths (e.g., 250 bp) imply longer runtime (e.g., ~60 hours mentioned)
  • Each run produces on the order of tens of gigabytes (e.g., ~40 GB stated in one example).
• Throughput claim:
  • A single run can cover:
    • about 48 whole genomes
    • and ~500 exomes (as stated in subtitles).

5) Dramatic drop in cost and the role of computing/algorithms

• Human Genome Project (early era):
  • Initially cost ~$30 million over 13 years (as stated).
• Modern era improvements:
  • With enough automation and the high-throughput shift:
    • can redo the project for about ~$100 million (as stated for a hypothetical redo).
  • 2007: high-throughput sequencing availability → huge price drop.
• Current pricing (as stated):
  • Pricing is said to hover around ~$1000 per human genome.
• But computational burden remains:
  • Sequencing generates terabytes of raw data, requiring:
    • major storage
    • heavy processing
    • efficient algorithms to interpret results quickly.
• Algorithmic progress using deep learning (2018 example):
  • A deep neural network approach was developed to call mutations / germline variants from raw data.
  • Motivation:
    • Individual genomes are ~99.99% similar
    • so it’s better to store/capture only differences rather than everything.
  • Outcome:
    • quickly identifies variant locations without needing to store all raw sequence detail.

6) Real-world applications: personal health and precision medicine

• Angelina Jolie (germline risk → preventive action):
  • Chose double mastectomy due to inherited risk indicated by family history and genetic testing.
  • Demonstrates how genetic risk information can lead to preventative procedures.
• Consumer genetic testing (23andMe example):
  • 23andMe: send DNA via mouth swab
  • Cost around $99 (sometimes on sale)
  • Provides health/disease risk predictions (as referenced in subtitles).
  • Potential downstream actions:
    • increased screening
    • preventive procedures (for high-risk individuals)
    • lifestyle changes (e.g., diet/alcohol/exercise adjustments).
• Cancer and somatic mutation targeting (precision medicine):
  • Example institution: Dana-Farber Cancer Institute.
  • For diagnosed cancer patients:
    • DNA sequencing is performed on tumor samples
    • typically not whole genome
    • instead, sequence a few hundred genes (targeted panels)
    • use detected mutations to choose targeted therapies
  • Matching mutation → drug targeting (examples mentioned in subtitles):
    • If relevant mutations like an “…EGFR mutation” are present, use an EGFR inhibitor
    • Specific mutations correspond to specific targeted drugs (some drug name text was garbled in the source, but the intent was mutation-specific therapy).
• Lesson / application takeaway:
  • Genomic sequencing supports both:
    • research and
    • clinical decision-making and
    • consumer health risk assessment
  • This creates a “huge data challenge,” motivating the course.

---

Methodology / Instructions Presented (Workflow Summary)

While no formal step-by-step “how-to” list is provided, the subtitles describe a sequencing workflow at a high level:

• Sample preparation
  • Historically: weeks of prep with many staff
  • Modern systems: prep in <1 day
• Load samples into sequencing system
  • Place prepared DNA into the sequencer and start the run
• Run sequencing for extended duration
  • Modern large instruments run for 3–5 days (in one described mode)
• Generate sequence data
  • Data output is produced directly by the sequencing hardware across long runs
• Downstream compute / interpret variants
  • Use algorithms (including deep neural networks) to call mutations/variants
• Clinical use for treatment decisions
  • In cancer care:
    • sequence tumor DNA using targeted gene panels (not full genome)
    • identify mutations
    • choose drugs targeted to those mutations

---

Speakers or Sources Featured (As Named in Subtitles)

• Harvard (references to Harvard main campus / Harvard Medical School)
• Dana-Farber Cancer Institute
• 23andMe
• Angelina Jolie
• Human Genome Project (project as a source/event; not an individual)
• Moore’s law (concept cited)