Fig 1.
Phenotypes of rice seedlings grown under different P conditions.
(a) A representative image of rice seedlings grown in P100, P5 and P0.25 conditions, showing plant size, root length and leaf senescence. Scale bar = 5 cm. (b, d) Frequency distribution of leaf Pi content (nmol/mm2) (b) and biomass (shoot dry weight, g/plant) (d). The values in the X axis of (b) are marked with both linear and logarithmic scales. (c, e) Fold change comparison of leaf Pi content (c) and biomass (e) in P100 vs. P5 and P5 vs. P0.25 conditions. (f) Percentage of plants showing senescence in the oldest (L1), the second oldest (L2) and the third oldest (L3) leave. For each P condition, the number of plants analyzed was at least 1,500 plants. P100, P5 and P0.25 conditions contained 320, 16 and 0.8 μM NaH2PO4, respectively.
Fig 2.
Spectral reflectance of the leaves from different P conditions and classification of P deficiency levels using artificial neural network model.
(a) Average reflectance spectrum (380–790 nm). Shaded areas indicate the one-standard deviation range. (b) Overview of the model architectures for the standard convolutional neural network (outside of the dashed box) and the multi-task model (including the dashed box). See Methods for the detailed description of the architecture. (c) Confusion matrix of the best model (multi-task model with hidden dimension = 64) on the test set (n = 940). (d) Class activation maps (CAMs) for samples from each P deficiency class. CAM values conceptually represent the relative importance of the data from each wavelength on the predicted class (see Methods). Black trend lines indicate the running mean. Shaded areas indicate the one-standard deviation range.
Fig 3.
Correlation and regression analyses of Pi contents and reflectance ratio indices.
(a) Heatmap shows Spearman’s correlation between 217 reflectance ratio indices (RNIR /RVIS) and leaf Pi content determined from the same leaves. R740/R560 (r = 0.72) and R750/R700 (r = 0.70) (marked with yellow squares) were selected for a regression analysis with Pi contents (nmol/mm2) (b). Pi content and R750/R700 data was fitted by a non-linear regression model with an exponential decay function. The formula and R2 statistics are displayed in the graph. Each data point is from an individual plant (n = 172 accessions x 2 P treatments (P5 and P0.25) x 3 individual plants x 3 independent experiments). (c) Scatter plot showing the correlation between observed Pi contents and the predicted values made by an artificial neural network model on a held-out test set consisting of 569 plants. The corresponding R2 is 0.53. (d) Class activation map (CAM) showing the relative importance of the data from each wavelength on the predicted Pi content (see Methods). Black trend lines indicate the running mean. Shaded areas indicate the one-standard deviation range.
Table 1.
Descriptive statistics of Pi content, shoot biomass, PUtE and reflectance ratio phenotypic values from 172 accessions.
Fig 4.
Manhattan plots and quantile-quantile plots from GWAS of leaf Pi content and reflectance ratios.
(a) Manhattan plot for Pi_P5 (Pi content determined from the P5 treatment). (b) QQ plot for Pi_P5. (c) Manhattan plot for R750/R700 index of P0.25 treatment normalized by P100 treatment (d) QQ plot for the R750/R700 index. For Manhattan plots, the x-axis represents SNP positions across the entire rice genome by chromosome, and the y-axis is the -log10(p-value) of each SNP. Red lines indicate the threshold line at -log10(p-value) ≥ 6.35.
Table 2.
List of loci associated with Pi content and spectral reflectance indices.
The number of significant SNPs and -log10 (p-value) for each locus of the reflectance ratios R740/R560 and R750/R700 are listed and separated by semicolons.
Table 3.
List of putative candidate genes associated with Pi content and spectral reflectance indices.