I decided to test this because I have always wondered what happens to a trait that cuts so deeply into reproductive fitness. Suicide usually strikes young, before people have children. Evolution does not let that kind of trait drift forever. It pushes back. So I wanted to know: has suicidality been under negative selection?
Ancient DNA now makes that question testable. By combining ancient genomes with the new GWAS for suicidality (Colbert et al., 2025), I tracked polygenic scores (PGS) for the trait across more than twelve thousand years of human history.
The pattern that emerged was small but consistent: older genomes carry slightly higher suicidality PGS, implying a slow, millennia-scale decline. That is exactly what you would expect if a fitness-reducing behavior faced steady evolutionary pressure, generation after generation, shaving off the risk alleles a fraction of a percent at a time.
But selection did not act in isolation. Populations kept moving and mixing. Hunter-gatherers gave way to early farmers, who were later overrun or blended with Steppe pastoralists, followed by classical and medieval populations. Each migration carried not just new tools and genes but new psychological profiles.
And the geography still matters. Northern Europeans, especially those with strong Eastern Hunter-Gatherer (EHG) or Steppe ancestry, carry more suicidality-linked alleles on average. The same pattern that shows up in the ancient data echoes faintly in the modern world, where Scandinavia, the Baltics, and Eastern Europe have some of the highest suicide rates.
It is a sobering reminder that evolution shapes not only bones or skin but also minds, and sometimes leaves shadows that reach all the way into the present.
What I tested (and why)
A new multi-ancestry GWAS (Colbert et al., 2025) just mapped 77 genetic loci for suicidality, spanning suicidal ideation (SI), suicide attempt (SA), suicide death (SD), and a combined “suicidal behavior” (SB = attempt/death) across 54 cohorts (PGC, MVP, UK Biobank, etc.). Sample sizes are farily large: ~260k SI cases, ~65k SA cases, ~9k SD, and ~75k SB: not impressive big enough to build reasonably powered PGS for each phenotype. It’s a preprint (so not yet peer-reviewed), but it’s the best single resource we’ve had for this trait class so far.
From those GWAS summary stats I derived three ancient-DNA PGS: SA, SB, and SI, plus their simple standardized mean (Z). Then I asked:
Do older samples carry higher suicidality PGS? (If suicidality harms reproductive success because it concentrates in youth, then natural selection should push the polygenic mean down over time.)
Is the trend specific to a late-glacial vs Holocene shift? (I tested a hinge at 12,000 years BP.)
Are patterns robust once I adjust for latitude and coverage?
Do the three component traits (SA/SB/SI) agree?
What does the trend look like in 500-year bins for the Holocene (<12k BP)?
What I find
1) A slow, directional decline toward the present
I standardized each PGS to z-scores and modeled Z ~ time (centuries BP) + latitude + coverage. Across all dates, time has a positive slope (older = higher PGS), which reads as a decline toward moderns:
Mean suicidality PGS (Z): +0.0034 SD per 100 years (≈ +0.034 SD per millennium), highly significant.
SA / SB / SI each show similar positive slopes per century, all highly significant.
So the signal isn’t huge (PGS shifts almost never are), but it’s consistent across phenotypes and in the expected direction if suicidality has been selected against over very long timescales. (For broader context on genetic architecture and correlates of suicidal behavior, see Li et al. 2023. Nature)
2) A latitude gradient that survives controls
Even controlling for time and coverage, latitude is strongly positive: farther north → higher suicidality PGS (order of ~0.02 SD per degree, i.e., ~0.2 SD per 10°). That aligns with well-known geographic patterns in related traits and environment-by-gene correlations, though causality here is not straightforward.
3) Coverage doesn’t drive the trend
Coverage enters as a small/borderline covariate overall (significant for SA, near zero for SB/SI, borderline for Z). In other words, the time trend is not an artifact of read depth.
4)Do the three components agree?
Yes, but only modestly, which is exactly what I’d expect given they’re related but not identical phenotypes. The cross-trait PGS correlations in my ancient panel are:
PGS_SA PGS_SB PGS_SI
PGS_SA 1.000 0.673 0.315
PGS_SB 0.673 1.000 0.274
PGS_SI 0.315 0.274 1.000Visual checks you’ll see in the figures (for paid subscribers only)
Scatter + LOESS (all dates & ≤12k BP): the cloud tilts gently upward with age (remember the x-axis is Years BP, reversed, so “older” is left).
Hinge regression at 12k BP: I fit Z ~ Date + max(0, Date − 12,000) to test if the slope changes across the Late Pleistocene/Holocene boundary. The key parameter is Δ (post–pre) slope. (The p-value for Δ tells you if the Holocene trend differs from the earlier one; see the caption/table in the post.)
500-year bins (<12k BP): mean z(PGS) by 500-year bins with 95% CI—a tidy way to see the Holocene glide path without over-smoothing.
Trait overlay (≤12k BP): Z-standardized SA/SB/SI plotted together (no PGS_w), each with its own LOESS; they move together.
I also show a partial-residual plot: the effect of time on z(PGS_w) after removing the fitted effects of latitude and coverage. The residualized line still tilts upward with age, which reassures me the time signal is genuine.