GDGT distribution in tropical soils and its potential as a terrestrial paleothermometer revealed by Bayesian deep-learning models

Branched and isoprenoidal glycerol dialkyl glycerol tetraethers (br- and isoGDGTs) are membrane lipids produced by bacteria and archaea, respectively. These lipids form the basis of several frequently used paleoclimatic proxies. For example, the degree of methylation of brGDGTs (MBT'5Me) preserved in mineral soils (as well as peats and lakes) is one of the most important terrestrial paleothermometers, but features substantial variability that is so far insufficiently constrained. The distribution of isoGDGTs in mineral soils has received less attention and applications have focused on the use of the relative abundance of the isoGDGT crenarchaeol versus brGDGTs (BIT index) as an indicator of aridity. To expand our knowledge of the factors that can impact the br- and isoGDGT distribution in mineral soils, including the MBT'5Me index, and to improve isoGDGT-based precipitation reconstructions, we surveyed the GDGT distribution in a large collection of mineral surface soils (n = 229) and soil profiles (n = 22) across tropical South America. We find that the MBT'5Me index is significantly higher in grassland compared to forest soils, even among sites with the same mean annual air temperature. This is likely a result of a lack of shading in grasslands, leading to warmer soils. We also find a relationship between MBT'5Me and soil pH in tropical soils. Together with existing data from arid areas in mid-latitudes, we confirm the relationship between the BIT-index and aridity, but also find that the isoGDGT distribution alone is aridity dependent. The combined use of the BIT-index and isoGDGTs can strengthen reconstructions of past precipitation in terrestrial archives. In terms of site-specific variations, we find that the variability in BIT and MBT'5Me is larger at sites that show on average lower BIT and MBT'5Me values. In combination with modelling results, we suggest that this pattern arises from the mathematical formulation of these proxies that amplifies variability for intermediate values and mutes it for values close to saturation (value of 1). Soil profiles show relatively little variation with depth for the brGDGT indices. On the other hand, the isoGDGT distribution changes significantly with depth as does the relative abundance of br- versus isoGDGTs. This pattern is especially pronounced for the isoGDGTIsomerIndex where deeper soil horizons show a near absence of isoGDGT isomers. This might be driven by archaeal community changes in different soil horizons, potentially driven by the difference between aerobic and anaerobic archaeal communities. Finally, we use our extensive new dataset and Bayesian neural networks (BNNs) to establish new brGDGT-based temperature models. We provide a tropical soil calibration that removes the pH dependence of tropical soils (n = 404; RMSE = 2.0 degrees C) and global peat and soil models calibrated against the temperature of the months above freezing (n = 1740; RMSE = 2.4) and mean annual air temperature (n = 1740; RMSE = 3.6). All models correct for the bias found in arid samples. We also successfully test the new