The 1976 Viking landers conducted a handful of experiments that involved injecting a nutrient-laden solution into Martian soil, then measuring gases given off in response. Indeed, gases were observed from the regular soil, but not from soil that was first heated to 160 C (sterilized). That seemed intriguing to many scientists—but others noted that the same result could be obtained through (abiotic) chemical oxidation triggered by the application of water. If I understand the arguments, heating the soil would break down the presumed oxidizer in the soil so it would then react less or not at all to a new injection of moisture.
But lo and behold, the scientists who (still) insist that Viking found life have published a new paper: “Complexity Analysis of the Viking Labeled Release Experiments” by Bianciardi, Miller, Straat, and Levin. They’ve used “complexity variables” to characterize the time series data, then clustered them (with k-means clustering, k=2). Indeed, they found that presumed “active” samples (including some examples from Earth) clustered together while presumed “inactive” samples (including some controls from Earth) clustered in a different group.
Since my dissertation was on clustering, I thought I should take a look and see how this machine learning method was being used in this setting. And, well, I’m just not convinced. Yes, they do seem to have gotten two distinct populations. But they only used 15 samples (11 from Mars, 4 from Earth) and that hardly seems sufficient to characterize the range of behavior, nor are they all obviously comparable (one time series consists of “core temperature readings taken every minute from a rat in constant darkness”; how is this related to possible bacterial activity in soil? Is darkness relevant? What about a rat in daylight, or a diurnal cycle?). The authors have agreed that more data would be better. I think more data, and thoughtfully chosen, would be essential.
My other reservation is about the “complexity variables” that were used. These are presented with no justification or discussion:
- LZ complexity
- Hurst exponent
- Largest Lyapunov exponent
- Correlation dimension
- BDS statistic
- Correlation time
Especially since these generated the 7D space in which the clustering happened, it’d be nice to have some intuition about why these might relate to life. There are some brief comments about life being “ordered” and of “high complexity” (and I’ve worked on this subject myself!) but I’m not convinced that the distinction they found is truly meaningful.
I don’t want to be unscientifically biased or negative. The results as presented in the paper do seem to show a quantitative separation between active and inactive samples. But this should be conducted with hundreds or thousands of samples from the Earth at the very least, where we have tons of examples of life-bearing soils as well as artificial or sterilized samples. These could fill out the feature space and properly position the Viking observations in more context.
Of course, it would also be useful to get more Martian samples!