Where genomics keeps breaking

The cost of running a single human genome dropped from roughly $100M in 2001 to under $200 today. Six orders of magnitude in just over twenty years. There is nothing else in biomedicine that has scaled like it. The sequencer is no longer the bottleneck for almost anything we want to do.

But that is just the sequencing. What nobody quotes is the cost of actually using one of those genomes for something downstream. Comparing two cohorts run on different chemistries. Re-running a variant call on a newer build of GRCh. Replicating a 2018 paper with a 2026 kit lot. Those costs have barely moved.

A few months ago I watched a four-day RNA-seq experiment fail to reproduce because a bead-cleanup ratio had drifted by 0.05x between two postdocs. The data, looked at on its own, was clean. The pipeline ran. The QC metrics passed. The pathway-level results shifted enough to flip a conclusion. It took two weeks of going through notebooks line by line to find the cause. The cause was three characters in a buffer table.

This is the part of genomics nobody writes about, because there is no clean story in it. The genome is not the problem. The genome is data. What we do not have a serious answer for is what produced the data. Buffer composition, incubation times, lot numbers, the exact timing on a magnetic stand, the room temperature when the library was made. Those things drift constantly. They are recorded in plain text in lab notebooks if at all. A lot of them never reach the paper because nobody thinks they will matter, until they do.

The next decade of work coming out of high-throughput sequencing will not be limited by how cheap a base pair gets. It will be limited by whether you can take a result and trace it back to the actual conditions that produced it. That is not a sequencing problem. It is a versioning problem.

A protocol is software. It has versions, dependencies, breaking changes, and a runtime behavior that drifts the moment you stop pinning its inputs. We treat experimental code that way. We do not treat the wet lab side that way, and the science is paying for it in a quiet, expensive, often invisible currency.

What gets interesting over the next few years is the moment a protocol becomes a first-class object the same way the sequencing read already is. When every reported result ships with the exact pinned protocol version that produced it. When comparing two studies starts with diffing their protocols the way we diff commit histories. When a question like “did the bead ratio change between cohorts” is something you can answer in thirty seconds instead of two weeks.

A lot of the most interesting work in single-cell, in spatial transcriptomics, in long-read assembly, is going to depend on this layer existing. We are starting to see the early version of it. It is not finished. I think we get there. I do not think we are there yet.