Automation Challenges in Next Generation Sequencing. Robin Coope Group Leader, Instrumentation BC Cancer Agency Genome Sciences Centre

Transcription

1 Automation Challenges in Next Generation Sequencing Robin Coope Group Leader, Instrumentation BC Cancer Agency Genome Sciences Centre

2 Outline The BCGSC s Scale of Operations Current Automation Projects and Lessons Challenges of Small Volume Pipetting Automating Fragment Analysis Calculations Automated Gel-Based Size Selection

3 BCGSC Monthly Total Libraries 14 HiSeqs, 3 MiSeqs, 8 Robots from 3 manufacturers ~ 50 Staff in high throughput lab groups.

4 BCGSC Monthly Pooled Library Output

5 12,500 Libraries per year QC Biospecimens QC Shearing and Amplicon Construction. QC RNA Isolation QC cdna Synthesis QC Normalization & Shearing QC Size Selection PCR (addition of indexes) QC End Repair A Tailing, Adapter Ligation Normalization and pooling Sequencing

6 Four Plate PCR-Free WGS A new protocol based on Illumina s but with more automation friendly labware Fully utilizes Biomek FX deck, grippers and plate stackers to enable library construction of four plates in six hours (w/o QC) One FX program with nine integrated stages

7 HL60 PCR Free

8 HL60 Normal (PCR) Library Construction

9 MultiMACS and Nimbus for mrna RNA isolation with MuliMACS allows Poly-A mrna capture while retaining the flow through which can be used to isolate mirna We are automating its operation with a Hamilton Nimbus. It has ul pipetting capacity and can work with the MultiMACS geometry.

10 Renormalization and Pooling Implementing a PE Janus for up-front quantification, normalization and rearrays A Span-8 has been acquired from our clinical partners for post library construction pooling and re-arrays

11 Hard Lessons High throughput production protocols need to be robust and easily maintainable. This means formal standards in program structure and liquid handling methods must apply. Pairing engineers with biologists to develop automation protocols with clear distinctions in mandate but great communication is most successful.

12 Where Liquid Handler Software Should Be Complexity of User Interface Most Commercial Liquid Handlers The Real Sweet Spot Power of Programming Environment

13 Pipetting or Things we discovered with difficulty that may have been obvious to others

14 QC Relies on Accurate Pipetting Liquid transfers below 2ul may depend on: Plate geometry Tip geometry Temperature Liquid properties Liquid surface angle Humidity

15 The Heartbreak of Backpressure Consider the following situation of successive transfers Aspirate 2ul

16 The Heartbreak of Backpressure Consider the following situation of successive transfers Dispense 2ul but with an added dispense motion to blow out the tips

17 The Heartbreak of Backpressure Consider the following situation of successive transfers In air, aspirate extra displacement for the blow out step But any leftover droplets may cause a vacuum to form in the tips

18 The Heartbreak of Backpressure Consider the following situation of successive transfers On next aspiration, the negative pressure from the liquid seal and pre-aspirate in some tips will cause over-aspiration of the sample

19 The Heartbreak of Backpressure 2 ul pipetted with a 1 ul blowout

20 The Heartbreak of Backpressure 2 ul pipetted with a 2 ul blowout

21 The Heartbreak of Backpressure 2 ul pipetted with a 4 ul blowout

22 The Heartbreak of Backpressure 2 ul pipetted with a 10 ul blowout

23 The Heartbreak of Surface Tension Using an Eosyn dye assay, we observed up to 20% variability in 2ul, transfers depending on the surface angle of the liquid. Successive single transfers were not consistent. Using the same protocol with the Artel MVS QC system, this effect was not observed. It was also not observed with DNA transfers.

24 intelligent Fluorescent Units

25 Illumina Clusters 20 MICRONS Higher molarity (number of fragments) = denser clusters Shorter fragments cluster more efficiently The optimum density is 4.6million clusters/tile with 100 tiles per lane. Too high density prevents base calling Runs cost ~$2500/lane

26 Molarity Estimation To get the molarity you have to get the mass and average size of DNA fragments. The mass is obtained by a fluorescent assay or qpcr Average size is generally determined by visual estimation from an Agilent Trace But this plot doesn t show molarity, it shows mass.

27 intelligent Fluorescent Units s-2 s-1 s s+1 s+2 FU x s s m s x The DNA moves with velocity ν(s) so: (ifu) Consider a capillary sections and DNA of a mass density of ρ(s). s is for size. MWDNAn x x v s s n(s) is the sample molarity s v s 1 dv 1 dv v s t ds t ds l so n s kfu s s v s dv ds and Mass Quant k 0 FU s v s s dv ds ds To use this you have to determine v(s) from DNA ladder.

28 An Empirical Method 1. Calculate the area under each ladder band peak 2. Divide each peak s corresponding mass concentration by its area and generate a correction factor curve 3. Divide the FU signal of the sample by the correction curve to generate a corrected mass concentration curve 4. Convert to molarity by dividing by size

29 The Two Methods Agree The two methods of calculation have a 94% correlation

30 Complementary Errors 1. The x-axis is linear in time not size, so larger sizes are more bunched together. 2. Linearizing the X-axis increases the weight of larger sizes. 3. Converting to molarity increases the weight of smaller sizes. 4. These visual errors cancel for well shaped peaks.

31 An Extreme Example

32 Correlation to Aligned Sequence Top: Actual library fragment sizes found from paired end sequencing and alignment of an amplicon pool. Middle: A plot of the same pool showing mass vs. linearized size. Bottom: A plot of the same pool showing molarity calculated by ifu using the empirical method. For n=8 amplicons, correlation was 60% for Agilent and 70% for ifu. For n=51 WGS samples, correlation was >90% for both

33 OLC (pm) Weighting for Cluster Efficiency Optimal Loading Concentration (OLC) vs. Size y = ln(x) R² = Size [bp] The relationship of mean size to clustering efficiency is shown in the empirically determined curve above. This gives a function OLC(s) which is then used to find the optimal loading concentration using the following formula. n(s) is the molarity as a function of size: c 0 n s 0 OLC n s s ds ds

34 Clusters Per Tile Weighting for Cluster Efficiency 6.50E+06 Cluster Density 6.00E E E+06 Cluster Density ifu Prediction Values Target Cluster Density 4.50E E E+06 Amplicon Samples This retrospective prediction assumes a linear relationship between loading concentration and cluster density. The ifu with OLC(s) weighting reduces the cluster density error by an average of 60% for these (n=28) amplicons.

35 Clustering efficiency (K/mm2 per pm) Clustering efficiency (K/mm2 per pm) qpcr vs. Intercalating Dye MiSeq library clustering efficiency - using Qubit quant All types R² = Average library size (bp) MiSeq library clustering efficiency - using qpcr quant CTL WGS PCR-free WGS WGBS RNA-Seq Amplicon Direct-amplicon Linear (All types) R² = Average library size (bp) All types CTL WGS PCR-free WGS WGBS RNA-Seq Amplicon Direct-amplicon Linear (All types)

36 Conclusions For tight libraries visual estimation of mean insert size gives good results and cluster errors are more likely to be from mass quantification errors than insert size estimation errors. For wider library distributions like amplicons, ifu improves the correlation to the aligned sequence by ~13% over the mass plot. For cluster generation, our prediction method shows a reduction in cluster density error by 60% in this set (n=28) of amplicons.

37 Gel Based Size Selection

38 Why Gel Based Size Selection To purify small or large fragments, like mirna or kilobase+ fragments To mitigate PCR induced biases in sensitive libraries such as for mrna To rescue libraries where shearing is off target and beads may miss the peak of peaks

39 Barracuda The GSC has built four 96 channel size selection robots and run over 10,000 samples since In Q3-Q4 of 2013, Coastal Genomics will release The Ranger based on this technology.

40 60mm Hole-to-Hole Size Selection

41 mirna Size Selection N.B. All of the mirna sequences in The Cancer Genome Atlas Project were size selected on Barracuda

42 Mid Length Range Two size selections of NGS libraries prior to adapter ligation Before size selection After size selection Target: bp Target: bp

43 Amplicons Amplicon before and after A 1kb target before and after

44 10kb Target from HL60 A 9-13kb Target Region This is an Agilent artifact!

45 10kb Target From Sheared gdna

46 18kb Target from Ladder A: 1kb Extended ladder B: 2.5kb Molecular Ruler C: A+B A2,B2 Target 18-50kb A3,B3 Target 25-50kb

47 Barracuda Control Software Barracuda uses a single pipettor so each sample is loaded and runs independently.

48 Performance Recovery efficiency is 90% for 10kb bands. Time: Barracuda, with a single channel can take up to 7.5 hours for a full plate Coastal Genomics Ranger will have 12 channels so expect full plates in ~1hr depending on protocol.

49 Instrumentation Lok Tin Lam Steven Pleasance Duane Smailus Philip Tsao Acknowlegements Quality Assurance Miruna Bala Susan Wagner Andrew Mungall and the Library Construction Team Coastal Genomics Jared Slobodan Matt Nesbitt Tech D, past and present Martin Hirst Michelle Moksa Yongjun Zhao Pawan Pandoh Cluster Density Estimation Rod Docking Lateef Yang Richard Corbett Aly Karsan Richard Moore Michael Mayo and The Sequencing Team

50 Quality Control in NGS You want to have the right amount of DNA going into the protocol and know it s not degraded You want to know how much and what size DNA is coming out of the protocol, so you can get the right cluster density. In theory this is simple but sample diversity and length dependent protocol steps (shearing, bead cleans) makes it complex

51 Declaration Robin is a co-inventor of the size selection technology described here and a board member of Coastal Genomics.