The K2-18b Biosignature Fight

Transmission spectroscopy is a clean idea with a messy execution. When a planet crosses in front of its star, a sliver of starlight filters through the atmosphere on the way to us. Different molecules absorb at different wavelengths, so the light comes back with chemical fingerprints stamped into it. JWST's NIRISS SOSS and NIRSpec G395H instruments together cover 0.8–5.3 μm — the same dataset both the original team and the reanalysts worked from, taken under JWST GO Program 2722.

Here's the humbling part. Before JWST, Hubble reported water in this atmosphere at better than 3σ — three separate groups, at 3.6σ, 3.9σ, and 3.3σ. JWST later showed that feature was actually CH₄ wearing a water costume. The two molecules absorb in similar shapes across Hubble's narrower 1.1–1.7 μm bandpass, and the instrument's edge effects near 1.7 μm did the rest. The reanalysis paper is blunt about the lesson: this atmosphere has already fooled the field once, and the fooling had statistical significance attached to it.

Schmidt et al., published in The Astronomical Journal on 5 November 2025, did the thing the original paper didn't: they threw everything at it. Multiple independent reduction pipelines, multiple retrieval codes, and enough compute to make a supercomputer sweat.

• 60 different data treatments across three independent reduction pipelines — FIREFLy, exoTEDRF, and Eureka!
• Over 250 atmospheric retrievals using two independent retrieval codes, POSEIDON and BeAR
• The first-ever analysis of the second-order NIRISS SOSS data, which reaches bluer wavelengths (0.6–0.85 μm)
• Roughly 50 CPU-years of total compute time, with individual retrievals running 2–5 days on 24 CPUs each

The reason for all that machinery is the point of the paper. Performing several reductions on one dataset quantifies how choices like outlier rejection, background subtraction, and limb-darkening treatment ripple into the final spectrum. Running multiple retrieval codes shows how model-level assumptions — opacity sources, aerosol parameterization — propagate into what you think you've detected. Inferences that survive all of it are called robust. A molecular signal that shows up in only one reduction is a warning sign, not a discovery.

The verdict sorted cleanly. CH₄ survives at about 4σ across nearly every data combination, with most combinations favoring it above 4σ — robust by any reasonable standard, and consistent within 1σ of the original 5σ claim. The BeAR code independently returned 3.5σ to 4.5σ for methane, in good agreement with POSEIDON. CO₂ and DMS did not survive. The maximum Bayes factor for CO₂ across all 60 combinations tops out at 4.2 (about 2.3σ), and for DMS at 3.9 (about 2.2σ) — "weak evidence at best" on the Jeffreys' scale. Any marginal signal at all appears only under specific pipeline-and-resolution pairings: the CO₂ hint needs FIREFLy pixel-level NIRSpec data coupled to exoTEDRF NIRISS data, and the DMS hint only shows at NIRSpec resolutions of R≈100 or R≈200 from one pipeline, then disappears at higher resolution. A detection that only exists when you choose the right pipeline isn't a detection — it's a preference.

The biosignature was fragile from the start. In the original paper, the DMS inference sat only about 1σ from the best-fitting model — which is exactly why it read as "potential signs" rather than a detection.

DMS is also a genuinely hard molecule to trust in a retrieval. The state-of-the-art reference cross-section data (HITRAN) lacks pressure and temperature dependence, which can quietly bias any model that includes it. And even on its own terms, DMS isn't a clean biomarker: in the proposed metabolic pathway it functions as an energy source for methane-producing life, not a by-product of it — so it fails the definition. DMS has also since been detected on a comet, which establishes that it can form abiotically with no biology anywhere in the picture.

The physical ask is steep, too. To push DMS to detectable levels on a hycean K2-18b, the record estimates you'd need a biological flux roughly 20 times Earth's entire marine output. That's not a smoking gun; that's a planet-sized phytoplankton bloom you'd have to assume into existence.

The mid-infrared didn't rescue it. During revision, Madhusudhan et al. (2025) reported independent evidence for DMS and/or dimethyl disulfide (DMDS) at about 3σ from MIRI LRS data. Then the counter-analyses arrived. A separate 2025 paper, Stevenson et al. — titled, with no ambiguity, K2-18b Does Not Meet The Standards of Evidence For Life — ran the MIRI data through independent reductions and found the transit spectrum is highly susceptible to unresolved instrumental systematics. Different wavelength binning schemes produced, in their words, "a potpourri of planet spectra." Under their preferred binning, 87.5% of retrievals do not favor DMS or DMDS. Their read: red noise, not an astrophysical signal. Schmidt et al. reach the same conclusion from the other direction — extrapolating the mid-IR DMDS model backward predicts strong absorption near 2.3 and 3.3 μm that simply isn't in the near-infrared data, a mismatch driven by the MIRI data preferring roughly 350 K while the near-IR prefers roughly 240 K.

Strip out the biosignature and you're left with the older, deeper question: is K2-18b a hycean world with a liquid-water ocean under a thin hydrogen sky, or a plain mini-Neptune with a deep gassy envelope and no surface to speak of?

The interior models lean hard toward mini-Neptune. To keep water liquid, the surface has to stay below water's critical point — 647 K and 220 bar — which caps the hydrogen/helium layer at no more than 0.03% of the planet's mass. Schmidt et al.'s posterior gives that configuration a probability of just 0.59%. Worse, matching the observed radius with such a thin envelope forces a water fraction of at least 70%, which is improbable from a formation standpoint: the solar water-to-rock ratio sits near 54%, and giant impacts can only deplete water, not add it. It's not physically impossible — it's a very specific fine-tuning to bet on.

The mini-Neptune alternative doesn't need any of that. A model at 100× solar metallicity, oxygen-poor and nitrogen-depleted by a factor of about 5000 — consistent with nitrogen dissolving into a basal magma ocean — reproduces both the retrieved CH₄ abundance and the missing ammonia, no ocean and no life required. The retrieved methane abundance itself, around 10%, actually resembles the deep methane abundance measured on Neptune (4% ± 1%).

And the photochemistry cuts against hycean directly. In a thin, CH₄-rich atmosphere, methane should convert to CO₂ on a geologically short timescale of roughly 10–100 Myr. So observing lots of CH₄ and no CO₂ is genuinely hard to square with the hycean picture — even an inhabited one with a large methane flux would build up CO₂ and CO to levels at or above the measured upper limits. The record's summary line is that the fine-tuning a hycean K2-18b would require is "an unlikely product of nature."

Underneath the astronomy, this is really a methods argument — and it's the part working scientists have been screenshotting. The reanalysis makes a sharp claim about how the field converts statistics into confidence.

The standard move is a Bayesian model comparison: run the retrieval with a molecule, run it without, take the Bayes factor, and translate it into an "equivalent detection significance." Schmidt et al. argue that translation routinely oversells. A Bayes factor of 3 gets reported as 2.1σ under the common scheme — but a Bayes factor of 3 is only 3-to-1 odds. A claimed "3σ detection" corresponds to a Bayes factor of just 21. Using a more conservative mapping that equates the false-positive rate with the p-value, the numbers come out roughly one sigma lower across the board.

Run through that lens, the original CO₂ "3σ detection" is better described as a 2σ result, for which "detection" is too strong a word. The reanalysts reserve "robust" for signals that hold across multiple reductions and multiple retrieval codes — a bar CH₄ clears and CO₂ and DMS do not.

This isn't a closed book — it's a paused one. Program GO-2372 (PI: Renyu Hu) is collecting four additional NIRSpec G395H transits and two NIRSpec G235H transits, sharpening precision across the 1.7–5.2 μm range for CH₄, CO₂, DMS, CO, H₂O, NH₃, and HCN.

Here's the twist that keeps everyone honest. Better data could actually turn up CO₂ — but at an abundance below the current upper limit, which would be consistent with a plain mini-Neptune rather than evidence for an ocean. A CO₂ signal near 1%, on the other hand, would be much harder to explain without a hycean surface. The same measurement points in opposite directions depending on where it lands.

What would genuinely reopen the biosignature question is narrower: a robust DMS detection above 3σ that holds across multiple independent reduction pipelines and retrieval codes, not just one binning scheme. No such detection exists in the available data right now. The reanalysis is careful to note that absence of evidence isn't evidence of absence — its own mini-Neptune models still predict some CO₂. The data will decide. It hasn't yet.