What runs beneath the surface

Over the past few days we have taken this step out of the chain at Atlas. Not with a replacement product, but by using what was already there.

What the Chromosome Analysis Suite computes, it does not compute itself. Beneath the surface run the Analysis Power Tools – a command-line package from Thermo Fisher, freely available, built for Linux, documented down to the last parameter. The same holds for the Transcriptome Analysis Console and the Axiom Analysis Suite: they are interfaces around programs that can also be called directly.

This is neither a secret nor a hack. Thermo provides the tools itself, with a manual and example calls. They are rarely used, because the interface is what people know, and because nobody touches a working procedure as long as it runs.

An analysis pipeline for CytoScan HD that needs no graphical interface. A CEL file goes in; out come the CYCHP analysis file, a segment table filtered to the reporting thresholds, the quality metrics as structured data, two plots and a report in the Atlas design. Alongside them comes a manifest with checksums for every output, the versions of the program and the library files, and the genome build used.

A run takes just over two minutes per sample on an ordinary server. Five samples go through in a single call, overnight, with no one sitting beside it.

Two decisions in it carry more weight than the technology. First, the pipeline checks the quality metrics against the manufacturer’s thresholds before it outputs a segment table – if a sample falls short, the report says so and contains no segment list that nobody could interpret. Second, it knows the sex call and evaluates the sex chromosomes against it, instead of reporting a normal male X as a loss.

An analysis you build yourself is only worth as much as the evidence behind it. So we had the same samples computed twice: once through our pipeline and once through the Chromosome Analysis Suite, by a different person, without sight of the other’s numbers, against the same library and the same genome build.

The quality metrics agree to six decimal places. MAPD, SNPQC and Waviness – the same values, down to the last digit, for every sample.

In the segments, the two routes differ at 36 places, spread across all samples, out of several hundred segments in total. We looked at every one of them. They are segment boundaries, not diverging findings: where one analysis splits a region into two segments, the other merges it into one – the same statement at finer or coarser resolution. The differences cluster in the samples with the weakest quality metrics, which is to be expected: the noisier the signal, the more often two implementations of the same method decide borderline cases differently.

The comparison log with all 36 regions, the tolerances and the command to repeat the comparison sits in our repository. It is updated whenever the program version, the reference model, the genome build or the array type changes.

Three things remain for now, and we would rather name them ourselves.

The scanner is still controlled by the manufacturer’s software. It produces the CEL file and needs a Windows machine next to it. That is a driver question, not a law of nature: instrument control through an open interface, down to firmware level, already exists for many lab devices, and the scanner is the next layer we are looking at.

The arrays, the library files and the reference model still come from Thermo Fisher. What we have dropped is the dependence on their analysis interface, not on the manufacturer – and that is a distinction worth keeping sharp.

And incoming quality control is still measured by hand here. Concentration, purity, gel image: a person does the pipetting, and the values land as data fields on the order, from which the report writes itself. The measurement itself is a question of robotics: liquid handlers, plate readers and automated electrophoresis take it over as soon as the throughput justifies it.

The same route carries over to the other array types we run. Beneath the Transcriptome Analysis Console lie the Analysis Power Tools and the open Bioconductor path for expression data; beneath the Axiom Analysis Suite lies the genotyping program from the same package. Both are next for us, and we will prove both the same way: two independent runs, documented differences, a log in the repository.

The figures in this text come from our own validation work. We name no customer, show no finding and publish nothing from ongoing joint projects before the parties involved have released it. That limits what we can show, and we consider it the right price: whoever sends us samples should be able to count on their case staying between us.

Genomics is becoming a foundation of medicine, and its costs keep falling – for diagnostics as much as for gene and cell therapies and for drug discovery. What limits access today is rarely the chemistry. It is the points in between: the one computer, the one licence, the one person who has to click for a result to exist.

Every one of these points that disappears lowers the barrier for the next doctor and the next patient. This one has disappeared.