The Sixth Week

EXTRACTION

The petrous, the rock-hard part of the skull around your inner ear, is the densest bone in the body, which is why it survives when nothing else does, and why on day two of the sixth and final week of funding for this lab, I start hoping our sample from the lowland Peruvian site is contaminated.

I am the only paleogeneticist left running this extraction bench, which is what happens when your PI retires and the university only gives you six weeks to close out grants before they pull the utilities. Six weeks to publish something, anything, or lose the cold room keys. I have known this clock in my emails for two months and now I am counting it in sequencing run slots.

It should have been impossible. The burial is undisturbed, precontact, and in the sort of hot, wet soil that strips DNA to nothing. We did not expect endogenous reads. When the qPCR comes back with 0.4 percent, I sit on the floor behind the fume hood for five minutes because I do not trust my pipette.

The community that granted us permission for a single, non-destructive analysis is very clear: no re-sampling, full results returned in plain language, and no release without their first look. They have been burned before, so I cannot blame them. I have spent most of my career pulling DNA from teeth and petrosals to prove other people’s lineages. Being a diasporic lab tech who has never managed to prove her own is the sort of private fact you keep to yourself when you are asking someone else to trust you with their ancestor.

I keep a staff reference panel in a locked spreadsheet. Before I go to bed I run the numbers and catch myself hoping the contaminant is me.

LIBRARY PREP

I build the libraries myself. No one else left to do it.

Every prep gets four no-template controls. At two a.m. I quantify them and see a whisper of signal at one of the haplogroup-defining SNPs matching our candidate. I stop breathing. Contamination is not a theory, it is the default. You spend your life trying to keep it out and you spend just as much of it trying to prove it did not sneak in.

I rerun the spike. I recheck the index plates. Then I do the math the way you are taught to: expected cross-talk, index hopping rate for that pool, the number of molecules going in. It comes in below the limit of detection we set and locked months ago. Not zero. Below limit.

That is the worst number in this field. You cannot publish below limit. You also do not get to pretend it isn’t there.

I close the door and call the community’s education liaison, D., from a phone I am not supposed to use for case details. I break protocol and send them the raw damage plot. I do not send a draft, I do not send an abstract. I send one PDF with a plot, an axis, and a caption written as plainly as I can make it. D. texts back a single line: show me the fourth control. So I do.

I am not going to say what D. said in response. I will say they did not say yes and they did not say no. They ask me if I have the staff panel results yet.

INDEX

Sequencing is a hostage situation. You pay by flow cell, you buy time, and time is exactly what I do not have. The last slot runs out at the end of week six, on Friday at five. On Wednesday morning I have just enough yield to map the staff references against the candidate assembly.

Mapping to a modern reference is where most ghost populations are born. Human references are biased: they over-represent some ancestries, under-represent others, and when coverage is low, the aligner will happily place reads where they fit the model rather than where they came from. For a tropical sample this degraded, that bias would be violent. So we have spent years optimizing for it: a masked reference, strict mapping quality, trimming for deamination, and calling variants only in transversion-rich regions with high base quality. None of it erases the bias. It just tells you how big the error bar is.

I run the staff panel first.

I find myself doing a stupid thing: I am scanning for myself. I have never uploaded my genome to any of the public databases I cite. I keep it on an air-gapped drive, in case I ever have to rule myself out.

No match above threshold. Not mine. Not the techs. Not the retiring PI.

It does not make me feel better. The number sits there on its own.

ALIGNMENT

Coverage is brutal. We are looking at about threefold average, with huge drop-off in GC-rich regions. Most pipelines would have thrown it out. Ours does not get to throw it out.

I bin by read length and by the C to T misincorporation at the five prime end: the deamination clock. Authentic ancient DNA gets a ski slope: high deam at the ends, tapering in. If it is modern contamination from a wet lab, that slope is flat or absent.

Ours is mostly a slope.

But there is a small tail where it flattens unnaturally. I spend an hour trying to model it as DNA damage repair, another trying to model it as index hopping between closely related libraries. The difference is one or two reads. One or two reads would not kill a high-coverage genome. For this one, one or two reads could kill its claim.

I refuse to round it.

The population test is worse. With this little data, an unsupervised clustering will invent structure. Remove one SNP, add one, change the minimum allele frequency, and the ghost moves. I run ADMIXTURE with bootstraps until I can recite the seed numbers. The lineage does not slot into any published model for the Americas. It also does not slot cleanly into the proposed Beringian or East Asian clusters. It forms its own very small, very unstable component that vanishes if I force one less ancestral population.

A real lineage can do this at threefold. That does not mean this is one.

DAMAGE

I print the damage plots. I hate printing, but the community liaison asked for paper if the server was slow. I sit across from two of the community’s representatives in their library after close. No PI. Just me, a box of gloves, and three sheets of paper.

One of them, an elder who has been cataloguing their own oral histories for longer than I have been alive, points to the flat tail on the second plot and says, “That one looks like you are trying not to see something.”

I say, “That one looks like the thing I cannot prove is not there.”

Another disagrees. They want to know, correctly, who gets credit, who owns the sequence, and whether a rival lab with more money will strip-mine citations from a tiny community if we publish first and sloppily. Neither of them treats me as a villain. The bone is not mine, and nobody in this room needs to say so.

I go back to the lab and pull the single-use bag with the remnant petrous chunk. It weighs less than a gram now. It has already given us more than it owed any of us.

I cannot re-sample. I can try to pull more reads from the existing libraries, but they are exhausted. I can resequence with the last flow cell slot, which would give me maybe a quarter more coverage if I am lucky. Quarter coverage will not resolve the tail. It will just let me torture it differently.

EMBARGO

It is Thursday. Friday at five is the axe.

The lab’s draft abstract sits open. I have written the limitations section three times. The first was a paragraph. The second was a page. The third is two pages and still does not cover the tail.

Every grant report, every tenure-track panic, every diasporic kid who wanted her name on a tree is telling me to hit submit. I want it too. That is the problem.

I email the PI the limitations draft as-is. I do not attach the final tree figure. I ask for an embargo extension, knowing full well there is none.

Then I do something that is not in any protocol.

I make a folder on the community’s read-only upload drop. I title it with the site code and the accession stub we agreed. I do not put the manuscript in it. I put the following: the raw damage plots, with axes labeled in English and in the community’s working language as they asked; the contamination controls with their full quantification; the staff panel table; the bootstrap seeds for the ADMIXTURE runs; and the exact mapping parameters. I put a plain-language README, written in the first person, stating the uncertainty as bluntly as I know how: we may have found a deep lineage; we may have found a batch artifact; with the data we have, we cannot tell. No claim of first Americans, no phylogeny figure, no dates beyond the ones already published for the site.

I write the name the elder offered for the lineage in the README notes, exactly as they spelled it for me. I leave it there.

I do not hit submit on the paper.

I hit send on the upload.

The progress bar fills. It is still going as I log out of the sequencer, turn the monitor off, and sit down in front of the cold room with its one working light.

I can still run the last extraction tomorrow. I can sit here until Monday, when the keys go back, and not touch it.

The bar finishes.

The last light reflects off the smallest bag of bone, still sealed, waiting on the community.