Fig WGS lineage project · result note

Cessac and Smith are genetically indistinguishable

Two cultivars, two names, two very different prices — and, across 4,973,908 filtered SNPs in a 52-genome panel, one genome. Here is every number we have, including the ones that argue for caution.

Whole-genome shotgun sequencing · 52 fig accessions · 4,973,908 SNPs after filtering · 27 distinct genotypes · reference Ficus carica UNIPI FiCari 1.0 (GCA_009761775.1)

Why anyone asked

A young, expensive introduction standing next to an old, cheap one.

Smith has been in circulation for decades. The collector record traces it from Croatia to France, then to the Becnel family and into the United States, where Becnel Nursery popularized it across the humid South. It is valued for an unusual combination: high fruit quality while still performing in a climate that ruins most elite figs. It is widely available and inexpensive.

Cessac is much newer to the trade. It is reported as a discovery by Thierry Demarquest of Figues du Monde, at Cessac en Quercy in the Lot, France, and it reached collectors around 2020 amid heavy demand and high auction prices.

Growers who acquired both noticed they could not tell them apart. That is a question sequencing can actually answer — not “are these two fruits similar?”, which is subjective and varies by season and site, but “are these two plants the same clone?”, which is a fact about DNA.

What was sequenced

Deep whole-genome shotgun; both samples on one flow cell.

This is not a marker panel or a handful of microsatellites. Both accessions were sequenced as whole genomes with 2 × 151 bp paired reads, aligned to the published Ficus carica reference assembly (333.4 Mb), and called jointly with the other 50 accessions in the panel so that every sample is judged at exactly the same set of sites.

Those depths matter more than they look. At 58× and 56× these are the 5th deepest and 6th deepest libraries in the 52-sample panel, and among the most complete — only 0.24% and 0.24% of genotypes missing, against up to 35.6% for the shallowest gene-bank library in the cohort. It means this particular comparison is not living at the edge of what the data can support: it is the best-powered clone call in the panel, not the weakest.

Both were also prepared and sequenced in the same batch and on the same lane, so no library-prep or flow-cell difference separates them. And because both are this study’s own material, both went through the physical resampling check that rules out a mix-up at the bench.

The trick that makes this readable: a tree compared with itself

Zero differences is the wrong expectation. Here is the right one.

This is the part that trips up most DNA-identity claims, and it is worth understanding before looking at a single number below.

Figs are highly heterozygous outcrossers, and short-read sequencing is not perfect. If you sequence the same plant twice you do not get zero differences — you get thousands, from sequencing error, from alignment ambiguity in repetitive DNA, and from the ordinary noise of calling genotypes at seven million sites. So “they differ at some sites, therefore they are different cultivars” is not a valid inference. Neither is “the number looks small, therefore clone”. You have to know what the noise floor actually is.

So the panel contains a deliberate positive control: three separate leaves picked from one single fig tree, extracted, prepped and sequenced as three independent samples. Those three are the same plant, by construction. Whatever difference they show is the noise floor. Every claim below is calibrated against it.

If two differently-named accessions differ by no more than three leaves off one tree do, then the data cannot tell them apart — and no amount of extra sequencing depth would change that.

The panel carries the opposite reference point too: a pair of genuinely distinct cultivars (CDD Noir and CDD Ciutat), so every metric has both a floor and a ceiling. And it carries a third, subtler one — a confirmed first-degree pair, Purple Tiger (BM) and Black Madeira, which is very closely related but definitively not a clone. That pair is the hard case, and it is the whole reason for the window-by-window test further down.

Result 1 — the raw genome-wide distance

Smaller than one tree compared with itself.

Count every SNP position where the two samples differ at all, heterozygous sites included, and divide by the number of positions where both actually have a genotype. That last part is not a technicality: library completeness runs from 0.04% to 35.6% missing across this panel, so raw counts between samples of different completeness are not comparable. Everything here is per megabase of comparable sites.

16,127 per Mb, against 19,536–22,374 per Mb for three leaves off one tree. The pair is not merely near the floor; it sits below it, and it is the near-smallest value of all 1,326 pairs in the panel on both the raw count and the normalized rate. For scale, the closest genotype to either of them that is not the other is Burjassot Negre, at 257,141 per Mb — roughly 15.9× further away.

A caution, in the interest of not overselling it: this metric is inflated by fig heterozygosity and is genuinely reliable only for clustering. Sitting below the floor is a good sign, not proof. The next metric is the one that carries weight.

Result 2 — the clean metric, with the noise removed

One difference. Not a rounded zero — literally one site.

The strict test discards everything ambiguous. Keep only positions where both samples have between 15× and 150× coverage and where fewer than 5% of reads carry a minority allele — so both genotypes are clean, confident homozygotes — then count the positions where they are homozygous for different alleles. Those are differences that sequencing noise struggles to invent.

Read that single difference carefully, in both directions. It is not zero, and this document does not report it as zero. But 1 difference in 3,392,332 sites sits inside the range this filter produces on material that is definitely one plant: the same-tree control returns 0 across 3,069,595 comparable sites, and a known clone pair elsewhere in the panel returns 4. A somatic mutation arising between two vegetatively-propagated plants of one clone would look exactly like this.

Result 3 — nine measurements, one answer

Different statistics, different assumptions, different software.

A single metric agreeing with itself proves nothing. These are separate statistics computed by separate programs, most of them on an independently prepared marker set (194,363 LD-pruned SNPs rather than the full 4,973,908). They fail in different ways, so agreement across them is informative.

Three of those deserve a note.

KING φ and IBD PI_HAT are the panel maxima. φ = 0.4427 and PI_HAT = 0.847 are the highest values of any pair in the 52-sample panel — higher than three leaves off one tree score against each other (0.4270–0.4340 and 0.831–0.848). That is not a paradox: the control’s three libraries are shallower and noisier than these two, and noise pushes an identity statistic down. Two very deep libraries of one genotype are the cleanest thing in the dataset.

NgsRelate R0 is exactly 0, and KING IBS0 is exactly 0 across 193,435 markers. Both count positions where one sample is homozygous for one allele and the other is homozygous for the other — the one pattern that clonal propagation cannot produce. Here there are none at all. For contrast, the closest non-partner relationship either member has is to Burjassot Negre, at R0 = 0.00457.

The clustering algorithm merges them without being asked. poppr collapses the 52 samples to 27 distinct multilocus genotypes using a threshold derived from the data rather than chosen by hand; Cessac and Smith come back as one genotype (#13). Under textbook idealizations — every marker independent, random mating — the chance that two genuinely different fig plants would match across 194,363 markers computes to around 10^−53,380. Real fig populations violate both assumptions, so read that as “astronomically small”, not as a calibrated probability.

Result 4 — walking the chromosomes, window by window

The test that separates “clone” from “extremely close relative”.

Everything above is a genome-wide average, and averages have one specific blind spot that matters enormously here.

Consider a fig and its own seedling offspring. They share exactly one haplotype at every single locus — so they too have essentially no opposite homozygotes, and R0 and IBS0 sit near zero for them as well. A clone and a parent–offspring pair can look alike on those statistics. What distinguishes them is not the average but the pattern along the chromosome: a clone shares both copies everywhere, whereas a relative shares one copy in long contiguous blocks and neither copy in others, because inheritance arrives in chromosomal chunks.

So the analysis was redone in 271 separate 1 Mb windows walking along the 13 chromosome-scale scaffolds. Each window is scored on its own as sharing both haplotypes, one, or neither.

The column that actually carries the argument is the second numeric one: the longest consecutive stretch of windows that are not fully shared. Genuine shared ancestry is contiguous along a chromosome; genotyping noise is scattered, so run length discriminates where a window count does not. Cessac ↔ Smith has a longest run of 0. The confirmed first-degree pair has 60. The distinct pair has 51. Even the known-same-cultivar replicate — one cultivar held in two different gene banks and sequenced twice by two institutions — shows a run of 2, because cross-archive batch differences add scatter that this pair simply does not have.

The test also validates itself, which is why the other four rows are shown: it correctly resolves a known first-degree relative as not a clone, and a known distinct pair as not a clone. A method that called everything clonal would be worthless.

Result 5 — a completely different variant caller

Because the answer should not depend on one piece of software.

Every number above ultimately comes from one genotype caller (bcftools mpileup, a statistical model). If that caller had a systematic quirk it could in principle manufacture agreement. So both samples were re-called from the same aligned reads with DeepVariant, a deep-learning caller that shares no algorithm with the first — it classifies pileup images with a convolutional network instead of fitting a likelihood model. The same-tree control and a distinct pair went through the identical merge, because these percentages are only comparable within one merge.

0.037% against the same-tree control’s 0.032% — the floor to within 16% — and 135× below the distinct pair. This is also the best-powered comparison in the cross-check, at 680,332 high-confidence comparable sites, more than any other pair, precisely because both libraries are so deep. Its overall genotype-difference rate (0.878%) is actually lower than the same-tree control’s (1.404%).

Only three of the eight clone groups in this panel have independent-caller confirmation. This is one of them.

How this happens without anybody lying

Both names run through France.

It would be easy, and wrong, to read a result like this as an accusation. The provenance records supply a mechanism that requires no error by anyone.

Smith’s recorded route runs Croatia → France → the Becnel family → the United States. Cessac is reported as a French discovery, from Cessac en Quercy in the Lot. Both routes pass through France. A clone exported to the United States under one name could perfectly well have remained in French circulation the whole time, been re-encountered decades later as an unnamed local tree, and re-entered the collector trade under the name of the place where it was found. Every person in that chain behaves honestly, and the outcome is still two names for one genome.

Nor is this rare. Fig cultivar synonymy is pervasive and already documented at whole-genome resolution — Spanish ‘Smyrna’ = Turkish ‘Sarılop’; a Spanish ex-situ collection of 209 accessions turned out to hold only 98 unique genotypes. Roughly half the names in a well-curated gene bank can be duplicates. This panel is another instance: 52 accessions collapse to 27 genotypes, in eight clone groups.

The collector literature got there first

Which makes it independent support, not an echo.

This is worth stating plainly, because it is the strongest non-genetic evidence available. Both editions of the collector cultivar encyclopedia already describe Cessac as “often compared with Smith”, name Texas BA-1 as a third name in the same comparison, and note that no published leaf or tree trait reliably separates them. The original edition goes as far as reporting side-by-side collector comparisons as “nearly identical, if not identical” and lists “possible redundancy with Smith” as a weakness; the revised edition pulls back to “a fruit-quality and provenance comparison, not a forced synonym”.

The horticultural record reached that suspicion by growing the plants, with no access to any DNA. The sequencing did not tell it what to think, and it did not tell the sequencing what to look for. Two independent lines of evidence converging is worth considerably more than either one alone.

What this does not say

Please read this section before quoting the result.

Or compressed to one line: the evidence that these two accessions are one clonal genotype is about as strong as this kind of data gets, and that is a different claim from any claim about names, motives, or garden performance.

Methods, in enough detail to check

One request

If you grow both of these and have kept records — ripening dates, splitting, hang time, taste notes, and especially anything that looks like a consistent difference between them — that information is genuinely valuable, and it is exactly the kind of evidence this analysis cannot generate. A shared genome plus a real phenotypic difference is a more interesting result than a shared genome alone.