Visualisation and Prediction of Phosphate Deficiency via Auto‐Glowing Plant Bioluminescent Sensors Authors: Ching‐Wen Chiu, Ya‐Fen Zhan, Ting‐Yang Lu, Liang‐Jie Chiu, Ya‐Ru Li, Cheng‐Jiun Huang, Hao‐Jun Lu, Huai‐Ji Huang, Alexander S Mishin, Karen S Sarkisyan, Ting‐Ying Wu, Chiu‐Han Hsiao, Ming‐Jung Liu Published: 2026 ABSTRACT Summary statementThis study introduces BioLumiPi, a novel, low‐cost, and non‐invasive biosensor plant, engineered by combining a phosphate (Pi) deficiency‐induced promoter with a fungal bioluminescence system in tobacco. BioLumiPi plants emit light that accurately reflects endogenous Pi deficiency and responds specifically to Pi levels. Coupled with deep learning analysis, this technology offers a proof‐of‐concept strategy for real‐time visualisation and prediction of nutrient stress to optimise fertiliser application and advance sustainable agriculture. Soil nutrients, including phosphorus (P), are essential for plant growth and crop yield. Plants absorb inorganic phosphate (Pi) from the soil, but Pi availability is often limited due to low solubility, leading to widespread deficiency and heavy fertiliser use (Paz‐Ares et al. 2022). Monitoring plant endogenous nutrient status is important to optimise growth and prevent excess fertiliser. Current methods (Supplemental Table S1), including mechanical sensors and molecular reporters like GUS and GFP, have limitations such as being destructive or not reflecting in vivo physiological states (Burton et al. 2020). Visual sensors like anthocyanin lack specificity (Li et al. 2014; de Bang et al. 2021), while the RUBY system offers a noninvasive and cost‐effective alternative, though its ability to track dynamic changes of stress signals needs further study (He et al. 2020; Carreon‐Hidalgo et al. 2022). Developing specific and non‐destructive sensors for monitoring dynamic nutrient status in plants remains crucial for sustainable agricultural management. Here, we genetically developed tobacco sensor plants that emit auto‐luminescence in response to Pi deficiency, enabling the continuous and non‐invasive monitoring of plant nutrient status in vivo. The fungal bioluminescent systems was genetically introduced into plants (Mitiouchkina et al. 2020), including luciferin‐synthesis and recycle‐related genes driven by constitutive promoters and the luciferase (Luz) gene controlled by the promoters of 35S (referred to as “35S:Luz; a constitutive promoter as a positive control), “no promoter” (referred to as “Vector”; no promoter sequences as a negative control), and the tomato phosphate starvation‐induced (TPSI1) promoter (referred to as BioLuminescence Pi sensor [BioLumiPi]; being the sensor plant of detecting Pi deficiency) (Figure 1A; Figure S1A). The 1.5‐kb promoter region of the TPSI1 gene, encoding a non‐coding RNA in tomato, responds to phosphate deficiency within 1–5 days (Lin et al. 2016), driving luciferase expression in tobacco and tomato. The BioLuminescence Pi sensor plants (BioLumiPi) using fungal bioluminescent systems to detect Pi deficiency in vivo. (A) Schematic of designing phosphate deficiency‐sensitive, auto‐glowing plants by integrating fungal bioluminescence genes with Pi‐responsive regulation. The genes for luciferin synthesis, recycling (green boxes), and luciferase (Luz; orange box) were genetically inserted into tobacco. Luz expression was driven by different promoters: 35S::Luz, Pi‐deficiency inducible TPSI1 (BioLumiPi), or no promoter (Vector). (B) Images show auto‐luminescence of plants under various Pi conditions, captured by a DSLR camera. (C–F) The quantitative auto‐luminescence signals (C), normalised auto‐luminescence (D) and the normalised Luz gene expression (E) by contrasting Pi‐deficiency/Pi‐recovery to Pi‐sufficiency treatments, and cellular Pi levels (F), with significant differences (p < 0.05) between treatments for the plant lines in (B). Transgenic plant lines under Pi deficiency and then shifted to recovery are shown with filled bars, while those under Pi sufficiency are depicted with open bars, as indicated in (B). The orange, green, and grey bars represent the Pi deficiency, sufficiency, and recovery conditions, as indicated in (B). (G) The caffeic acid abundance for plants indicated in (B), with the treatments of Pi deficiency (0 μM) and sufficiency (500 μM) for ~17 days. Shown for the samples collected from six biological replicates. The statistically significant differences (p < 0.05) were determined using two‐way ANOVA. (H,I) The bioluminescence after the second Pi deficiency (H) and nutrient withdrawal of nitrate (‐N), phosphate (‐P) and potassium (‐K) (I). (J–L) Illustration of using BioLumiPi plants with image analysis and modelling to predict Pi deficiency (J), with model classification performance measured by AUROC curves (K) and the Grad‐CAM heatmap highlighting areas focused by the imaging model (L). We screened and focused on T2 homozygous tobacco plants and successfully isolated multiple independent and stable transgenic lines of BioLumiPi, 35S::Luz, and Vector plants (See Supporting Methods; Figure S1B–D). These bioluminescent signals were autonomously emitted from plants and did not require the exogenous luciferin substrate. They can be captured by a low‐cost digital camera device (Figure 1B), by smartphone cameras such as iPhones (Figure S1I), and even visible to the naked eye in darkness. Below, one set of transgenic lines of 35S::Luz (#8), Vector (#1), and BioLumiPi (#31) was mainly included to represent the findings. The BioLumiPi plants exhibited significantly stronger auto‐luminescent signals upon Pi deficiency compared to the Pi sufficiency conditions (orange and blue arrows in Figure 1B–D; Figure S1E; Video S1). These signals were decreased during recovery conditions (i.e., phosphate‐sufficient treatment; 500 μM Pi; Figure 1B–D; Figure S1E,F). The mRNA abundance of the Luz gene in BioLumiPi plants showed a ~ 19.2‐fold increase under phosphate deficiency (green bar; Figure 1E; Figure S1H). The cellular Pi contents in BioLumiPi plants were low under deficiency conditions and increased during the recovery treatment (Figure 1F; Figure S1G). In contrast, the 35S:Luz and Vector plants did not show such a predominant difference in autoluminescence signals and Luz expressions upon Pi deficiency (Figure 1B–F; Figure S3D). The level of caffeic acid, the precursor of luciferin (Zheng et al. 2023), was similar across the three transgenic lines between Pi‐deficient and sufficient conditions (Figure 1G). To assess the robustness of BioLumiPi plants to Pi deficiency, we put the first‐round recovered plants into Pi deficiency and then recovery conditions. The analyses of auto‐luminescence signals, Luz mRNA expressions, and cellular Pi contents reflected the impacts of the second‐round Pi‐deficiency treatments (Figure 1H; Figure S1J–L), showing the repeatable responsiveness of BioLumiPi plants to external Pi dynamics. The BioLumiPi plants responded to Pi deficiency at both vegetative and reproductive stages. (Figure 1B,H; Figure S1M). Lastly, the Pi‐deficient induced auto‐luminescence signals were significantly observed at ~ 10 days whereas the plants became smaller and pale green at ~ 17 days (Figure 1B). This feature suggests the potential of the BioLumiPi plants as a readout of Pi‐deficient status before physiological damage. Together, these findings showed the responsiveness and robustness of auto‐luminescent biosensors in detecting endogenous Pi status in plants. In addition, the BioLumiPi plants provide high‐sensitivity signals recorded by low‐cost devices and have potential as an early indicator of Pi deficiency before physiological damage occurs. Note that the growth phenotypes, including leaf size and fresh weights from seedling to bolting stages, and flowering times were similar among wild‐type, 35S:Luz, Vector, and BioLumiPi plants (Figure S2). The correlation analyses of the global expression profiles among three distinct transgenic plants and wild‐type tobacco showed high similarity (Figure S3A). The Principal Component Analysis further showed that the phosphate treatments, but not the genetic background, were the primary factor in clustering samples (Figure S3B). The Gene Ontology analyses found genes induced by Pi deficiency were related to stress, abiotic stimulus, and phosphate starvation, a pattern found in both transgenic and wild‐type tobacco (Figure S3C). The expression patterns of hallmark Pi‐starvation genes (e.g., PHO1, SPX1‐4) were highly consistent across all lines (Figure S3E). These observations suggest that transgene introduction does not affect plant developmental processes and phosphate responses. We next asked, in addition to phosphorus, whether the BioLumiPi plants are responsive to other nutrients. Compared to the full nutrient solution (which includes nitrate, phosphate, and potassium), the BioLumiPi plants treated with the Pi‐depleted solution showed higher signals, consistent with their lower cellular Pi contents and increased gene expression of the Luz gene (Figure 1I; Figure S4). In contrast, the luminescent signals, cellular Pi contents and Luz mRNA expressions were comparable for BioLumiPi plants grown under nitrogen/potassium deficiency and full‐nutrient conditions, indicating that the BioLumiPi plants responded majorly to Pi deficiency. To further explore the responsiveness and sensitivity of BioLumiPi plants and the impact of culture medium in plant stress responses (see Supporting Methods), the plants were grown in soils (with limited Pi contents) and watered with nutrient solutions with five different Pi concentrations ranging from 0 to 500 μM (Figure S5A). The autoluminescent signals of BioLumiPi plants remained dim at 500 and 250 µM but became stronger at 50 µM, with similar magnitudes at 50, 10, and 0 µM, reflecting to the lower cellular Pi contents across the conditions with ≤ 50 μM Pi (Figure S5B,C). These results suggest that the endogenous Pi‐deficiency response and the TPSI1 promoter‐triggered luminescent signals of BioLumiPi plants were in a threshold‐dependent manner in response to Pi deficiency. An imaging system integrated with machine learning can remotely capture luminescence and detect stress signals from sensor plants (Figure 1J). To determine whether deep learning models, combined with the bioluminescence imaging system, can identify and predict Pi deficiency in sensor plants, we analyzed images of BioLumiPi plants using deep learning techniques to classify them as either Pi‐sufficient or Pi‐deficient (see Supporting Methods and Figures S5,6 for image dataset collection and model generation). Our analysis revealed that multiple Convolutional Neural Network (CNN)‐based classification methods performed well, achieving mean AUROC values ≥ 0.9 in distinguishing Pi‐deficient and Pi‐sufficient BioLumiPi plants (AUROC score: a single metric to summarise the model performance with the values 1 and 0.5 representing perfect and random prediction; Figure 1K). We further applied the Gradient‐weighted Class Activation Mapping (Grad‐CAM) analyses to visualise the image areas where the model found important for distinguishing Pi‐deficient and Pi‐sufficient BioLumiPi plants (Selvaraju et al. 2017). The regions highlighted by Grad‐CAM matched the plant‐image areas with higher auto‐luminescent signals (Figure 1L). Overall, our model effectively captured luminescence signals from stressed sensor plants, suggesting that image‐ and prediction model‐based stress detection with plant biosensors is a promising proof of concept. Light emission is key for plant sensors. Our bio‐luminescent sensors emitted more light under Pi deficiency but were dimmer than 35S::Luz lines (Figure 1B). Autoluminescence‐based methods depend on luciferase gene expression and substrate availability within cells. While cellular metabolism may influence the auto‐luminescent signal, our approach offers insights that can be validated with other methods. Additionally, emerging auto‐luminescence technologies, such as engineered luciferase variants that use the same substrate but emit light at different wavelengths, are expected to enhance the robustness of future sensors in terms of metabolism (Morozov et al. 2025). Besides, recent studies improved fungal bioluminescence stability and activity, and enhanced precursors like caffeic acid and hispidin, boosting plant autoluminescence (Ge et al. 2024; Shakhova et al. 2024). These strategies have the potential to boost Pi‐deficient plant biosensor brightness. In conclusion, Our findings showed that the self‐sustained, Pi‐deficiency‐responsive BioLumiPi plants can visualise and predict the in vivo phosphate status in plants. This serves as a proof‐of‐concept sensor for profiling endogenous stress signals in plants using a whole‐plant imaging platform with low‐cost consumer‐grade cameras. This strategy offers an affordable, highly sensitive system for real‐time monitoring, with potential uses in tracking plant signals across various developmental stages and environmental conditions at multiple scales (Mitiouchkina et al. 2020; Balakireva et al. 2026). Evaluating BioLumiPi sensor performance under fluctuating environmental conditions (e.g., varying light, temperature, and soil pH) is a critical next step for field‐scale deployment. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Supporting File 1 Supporting File 2 Supporting File 3 Supporting File 4 Acknowledgements We thank the Transgenic Plant Laboratory, Spectrometry Core Facility at Academia Sinica Biotechnology Center in Southern Taiwan and AS‐BCST Bioinformatics Core for agroinfiltration, LC‐MS/MS and computing services; Dr. Shu‐I Lin and Dr. Tzyy‐Jen Chiou for sharing the TPSI1 promoter construct and expertise on Pi deficiency treatment; Dr. Chia‐Yi Cheng, Dr. Kuo‐Chen Yeh, Dr. Hsin‐Hung Yeh, Professors Frank Yeong‐Sung Lin and Yennun Huang, and Mr. Yu‐Chun Lai for their intellectual input. This work was supported by grants RSF 25‐14‐00349 to Alexander Mishin. UKRI MRC MC‐A658‐5QEA0 and BBSRC ISPF APP33664 grants to Karen Sarkisyan, and AS‐TP‐110‐M07, AS‐IAIA‐114‐L05 and NSTC 113‐2326‐B‐001 grants to Ming‐Jung Liu. Figures 1B, 1J and S1I were created with BioRender.com. Data Availability Statement The data that supports the findings of this study are available in the supporting material of this article. References Associated Data Supplementary Materials Supporting File 1 Supporting File 2 Supporting File 3 Supporting File 4 Data Availability Statement The data that supports the findings of this study are available in the supporting material of this article.