LEAF-Link

In March 2024 NASA selected Lunar Effects on Agricultural Flora (LEAF) as one of the first instruments for astronauts to deploy on the lunar surface under Artemis. Led by Space Lab Technologies, it is the first experiment designed to observe plant photosynthesis, growth and systemic stress responses in space radiation and partial gravity.¹ Its growth chamber holds Arabidopsis thaliana, red and green varieties of Brassica rapa, and the duckweed Wolffia.²

LEAF-Link asked whether a plant habitat could act on stress before it becomes visible. The design has three parts: define plant stress states from gene expression; train a model that predicts those states from the chamber's ordinary sensor readings; and let the chamber controller respond when predicted stress rises. This study tested what public data allows: whether spaceflight gene expression gives a usable stress label, and how a chamber controller responds to a routine setpoint change, with and without a delay in its temperature reading.

1. Principal component analysis. Counts were transformed as log2(count + 1); the 5,000 most variable genes were kept and each was scaled to mean 0 and standard deviation 1. For each component, the share of its variance explained by strain and by spaceflight was computed as eta squared (between-group sum of squares over total sum of squares). The same steps were repeated within each strain.
2. Held-out classification. An L2-regularized logistic regression (C = 0.1) predicted spaceflight versus ground from expression alone, with the gene filter and scaling fitted inside each training fold. Each canister position contributed two samples, so both were held out together (28 folds). Significance came from 1,000 permutations that shuffled the flight and ground labels between canister positions within each strain.
3. Transfer and sensitivity. The classifier was trained on three strains and tested on the fourth, for each strain in turn. The principal component analysis and the held-out accuracy were repeated with 1,000, 2,000, 10,000 and all expressed genes (mean count of at least 10, 19,901 genes), and the classifier with C = 0.01 and 1.
4. Known spaceflight genes. The original study reported heat-shock genes up and peroxidase genes down in flight.⁴ For 14 heat-shock genes and the 57 expressed class III peroxidases, identified from UniProt reviewed entries,⁵ the change in flight was computed within each strain as the difference in mean log2 counts.
5. Growth-chamber control simulation. One temperature channel was modeled as a first-order system (time constant 180 s, 0.12 °C per W, heater limited to ±80 W). The setpoint stepped from 20 to 25 °C at 30 minutes, as habitats do between night and day. Three PID gain sets were compared on overshoot and on time to settle within 25 ± 0.3 °C, first with an instant temperature reading and then with the reading delayed by 15, 30, 60 and 120 seconds. An illustrative stress score, 1 − exp(−(0.35 × overshoot + 0.08 × minutes above 26 °C + 0.022 × settling minutes)), stood in for a biological model.

The principal component analysis and the control simulation without delay were written in April 2026. The classification, transfer and sensitivity checks, known-gene check and sensor-delay runs were added in October 2026, when the study was closed out.

The one sample labeled wrongly in the main test was a Ler-0 ground control. Accuracy was the same, 98.2%, for every gene filter and regularization setting tried, while the share of PC1 explained by strain varied with the number of genes kept (table below). Read depth (mean 38.7 million reads in flight samples, 43.0 million in ground) and ribosomal RNA share (1.1% and 1.0%) were similar between conditions.

1. Strain is the largest source of variation in this dataset. With the 5,000-gene filter, the first two components (16.4% and 14.4% of variance) follow strain at 81% and 94%; strain explained at least 77% of PC2 under every gene filter tested. This agrees with the original study's emphasis on natural variation.⁴
2. A spaceflight expression signature is present, robust and transferable within the species. Held-out accuracy was 98.2% (p = 0.001) and did not change across gene filters or regularization. A model that never saw a strain still labeled 49 of 56 samples correctly. Gene expression can therefore serve as the label in a supervised stress model across genotypes of one species, which was the first requirement of LEAF-Link. Transfer across species, which LEAF would need for Brassica rapa and Wolffia, was not tested.
3. The signature reflects known spaceflight biology. Heat-shock genes rose and class III peroxidases fell in flight in every strain, as the original study reported. This indicates that the separation reflects a biological response rather than only a technical difference between flight and ground samples.
4. The sensor-to-stress link was not tested. The OSD-37 files contain no chamber sensor records linked to samples. Testing that link needs a study that records sensor time series and gene expression for the same plants.
5. A stress-aware controller has to be tuned for sensor delay. Without delay, faster gains always scored better. With a reading delayed by 30 seconds or more, the aggressive gains overshot and oscillated, while the balanced gains stayed stable up to 60 seconds and failed at 120. The ranking reversed, so tuning must use the measured delay of the habitat's own sensors.

• One experiment, one species and 8-day-old seedlings. Flight and ground differ in several ways at once (microgravity, radiation, hardware handling, return), so "spaceflight" is a compound condition rather than a single stress.
• All tests used one experiment. The signature was not tested on a second, independent spaceflight study, so transfer across experiments and hardware is unknown.
• The two sample sets in the study (labeled R1 and R2) also contribute to the components that carry the spaceflight signal (31% of PC3, 19% of PC6). Both sets contain flight and ground samples in equal numbers, so they do not create the signal, but they add structured variation.
• How strongly strain dominates PC1 depends on the gene filter (37% to 98% across the filters tested); the dominance of strain on PC2 and the classification result did not.
• The known-gene check is descriptive. Genes in each family are co-regulated, so the counts are not independent tests.
• The chamber model is first-order with a fixed sensor delay, and the stress score is a placeholder rather than a measured plant response.