Predictive resazurin phenotyping using VIMS oyster family data

resazurin

metabolism

cvirginica

Predictive phenotyping using resazurin curve characteristics in VIMS oyster family data

Author

Ariana Huffmyer

Published

July 23, 2026

I used AI to adapt existing code for the structure of this project and generate figures and used AI to write the base text below that I edited to describe the code and what was done in the analysis.

Resazurin Curve Feature Analysis

Overview

This analysis extracts quantitative curve features from resazurin metabolic assay data collected across oyster families from two phenotype groups: high-salinity (25 families, n = 638 individuals) and low-salinity (25 families, n = 617 individuals). Each individual was measured at 5 timepoints (0–4 h), yielding one metabolic activity curve per oyster. Twenty-five curve features were derived per individual — capturing magnitude (AUC, peak), rate (initial slope, Vmax), temporal structure (time to peak, inflection time), and depression-recovery dynamics — and compared across families within each phenotype group using one-way ANOVA ranked by effect size (η²), with Benjamini-Hochberg correction for multiple comparisons and Tukey pairwise post-hoc tests.

Data links

All data can be found in the VIMS Resazurin repository here.

Key Results

I added a directory to show diagnositic feature plots for each familiy here.

1. Family explains meaningful variation in metabolic curve shape

All top-ranked curve features showed statistically significant family effects (BH-adjusted p < 0.05 in all cases). Family identity explained 15–19% of individual variation in the best-distinguishing features (η² = 0.15–0.19), indicating that a substantial portion of metabolic output is heritable at the family level.

2. The most informative curve features differ between phenotype groups

The feature that best discriminated families was not the same between the two phenotype groups:

Phenotype	Best feature	η²	F (df = 24)	BH-adj p
High-salinity	`initial_slope`	0.154	4.65	< 0.001
Low-salinity	`auc_late`	0.193	5.90	< 0.001

High-salinity families were best differentiated by early-phase metabolic rate — how quickly resazurin reduction accelerated in the first hour. Low-salinity families were better differentiated by sustained late-phase output — total area under the curve in the second half of the assay. This pattern suggests the two phenotype groups may have distinct metabolic kinetics: high-salinity-origin oysters vary most in how fast they ramp up, while low-salinity-origin oysters vary most in how long they sustain high activity. The top five features for low-salinity families (auc_late, auc_total, delta_auc_late_minus_early, metabolic_scope, peak_value) uniformly describe metabolic magnitude and accumulation, whereas for high-salinity families the ranking is led by rate (initial_slope) alongside magnitude features. These results are relevant for this one experiment - we should not yet extrapolate these characteristics to these phenotype groups in general outside of the context of the individuals we measured here.

3. Individual families show consistent, distinctive metabolic fingerprints

The family mean heatmap reveals coherent family-level metabolic phenotypes that hold across multiple curve features simultaneously. In the high-salinity group, families 14 and 15 are consistently below-average across all top features (lower initial slope, lower metabolic scope, lower AUC), while families 24 and 43 are consistently above average. In the low-salinity group, family 78 stands out as markedly high across AUC-related features, while families 57 and 61 trend low. The consistency of these rankings across independent curve metrics strengthens the interpretation that these represent real family-level differences in aerobic capacity, not assay noise.

4. Most individuals follow a monotonic increase; trajectory class varies by family

The trajectory class plot shows that the dominant response pattern across both phenotype groups is monotonic increase (~60–90% of individuals per family) — resazurin reduction rises steadily throughout the assay, consistent with sustained aerobic metabolism. A minority of individuals show rise-then-depress trajectories (~10–45% per family depending on family), in which activity peaks and then declines, potentially reflecting stress-induced suppression of aerobic pathways. Notably, the proportion of rise-then-depress individuals varies substantially among families (e.g., family 6 in high-salinity and family 48 in low-salinity show ~45% rise-then-depress vs. families with near-zero proportions), and this qualitative shape difference is not captured by magnitude features alone. A small number of individuals show depress-then-recover trajectories.

Implications for Predictive Phenotyping

Family signal is detectable from a short assay. A 4-hour resazurin incubation produces enough curve structure to rank families by metabolic output with effect sizes (η² ~0.15–0.19).
Different features may be needed for different phenotype backgrounds. The divergence in best-discriminating features between high-salinity and low-salinity families suggests that screening programs should not apply a single universal metric. High-salinity families should be screened using early-rate features; low-salinity families using total or late-phase AUC.
Trajectory class adds complementary information. Families with a high proportion of rise-then-depress individuals may represent a metabolically distinct stress-response phenotype that is missed by simple AUC or slope summaries. Including trajectory class alongside magnitude features would give a richer picture of family-level capacity.
Consistent family rankings across multiple features increase confidence. The convergence of multiple independent metrics (slope, AUC, peak, metabolic scope) on the same high- and low-performing families reduces the chance that rankings reflect feature-specific artifacts.

Figures

curve_feature_family_effect_ranking.png
η² (family effect size) for the top 8 curve features per phenotype group; blue bars are BH-significant

curve_feature_best_family_boxplots.png
Distribution of the single best-ranked feature per phenotype group across all families

curve_feature_family_mean_heatmap.png
Z-scored family means for the top 5 features per phenotype group; red = above average, blue = below average

curve_feature_trajectory_classes.png
Proportion of individuals per family following each trajectory class (monotonic increase, rise-then-depress, depress-then-recover)

Resazurin Predictive Phenotyping

Overview

This analysis tested whether resazurin metabolic curve features measured on live oysters can predict family-level survival performance, and whether different metabolic signals predict survival under high-salinity vs low-salinity conditions.

Dataset: 48 families matched across two sources — resazurin curve features (~25 traits per individual, aggregated to family means) from output/resazurin-curves/curve_features.csv, and observed survival proportions under high-salinity and low-salinity challenge conditions from data/predicted_performance.xlsx. Families span two parentage groups: high-salinity origin (24 families) and low-salinity origin (24 families).

Approach: For each curve feature × parentage group combination, Spearman rank correlations were computed against both survival conditions in parallel. The top-performing features were validated with leave-one-family-out cross-validation (LOFO-CV), and a multi-feature resazurin index was built by z-scoring and sign-aligning the top four traits. All analyses were run separately for predicted high-salinity survival and low-salinity survival, with a direct comparison of predictive divergence between the two.

Key Results

1. Low-salinity survival is strongly and reliably predictable from resazurin curves

Multiple resazurin curve features show significant correlations with low-salinity survival, all concentrated in the low-salinity parentage group:

Feature	Spearman ρ	p	LOFO-CV ρ
`vmax`	+0.61	0.002	0.48
`auc_early`	+0.60	0.003	0.40
`initial_slope`	+0.57	0.004	0.35
`auc_total`	+0.57	0.004	—
`auc_late`	+0.53	0.008	—
`time_to_peak`	−0.50	0.013	—

The direction is consistent and biologically interpretable: families with higher metabolic capacity (faster initial rate, larger total and early-phase AUC, higher Vmax) survive better under low-salinity challenge. A multi-feature composite index built from the top four traits achieves a LOFO-CV Spearman ρ of 0.53, indicating that the resazurin metabolic profile genuinely recovers the low-salinity survival ranking for new, held-out families.

2. High-salinity survival is weakly and unreliably predictable

Only one feature reaches statistical significance for high-salinity survival: time_to_min_slope in the low-salinity parentage group (ρ = 0.42, p = 0.04, CV ρ = 0.27). No features in the high-salinity parentage group or the pooled dataset achieve p < 0.05. The multi-feature index built for high-salinity survival fails cross-validation entirely (LOFO-CV ρ = −0.26), meaning the in-sample relationship does not generalise to new families.

The weak signal that does exist points to timing and metabolic depression features rather than capacity: families that delay the onset of metabolic decline (time_to_min_slope) and show greater post-peak depression (metabolic_depression_index, depression_abs) tend to have marginally better high-salinity survival. This is a qualitatively different physiological story from what predicts low-salinity survival.

This reflects our previous analysis that correlated AUC with high and low salinity survival predictions.

3. High-salinity and low-salinity survival are predicted by fundamentally different features

The comparison scatter plot makes this divergence visually clear. Most features fall well off the diagonal (equal-prediction line), separating into two clusters:

Metabolic capacity features (vmax, auc_early, initial_slope, auc_total, auc_late, final_delta) cluster in the upper-centre of the plot: near-zero correlation with high-salinity survival but strongly positive correlation with low-salinity survival. These features predict one condition while being uninformative for the other.
Metabolic timing and depression features (time_to_min_slope, metabolic_depression_index, depression_abs) fall to the right of centre: positive correlation with high-salinity survival but weaker correlation with low-salinity survival. These are the only features with any above-diagonal signal for high-salinity performance.
High-salinity parentage families (orange circles) cluster tightly near the origin — their resazurin features are largely uninformative for either survival condition, at least at the 24-family scale analysed here.
resilience_ratio and time_to_peak are strong predictors of low-salinity survival in the negative direction (families that reach peak activity quickly and have lower resilience ratios survive better), and show negligible signal for high-salinity survival.

This pattern suggests that high-salinity and low-salinity survival draw on different physiological mechanisms that are captured by different aspects of the resazurin curve. Metabolic capacity (how much total aerobic activity a family can sustain during the assay) predicts performance under low-salinity stress; metabolic timing and depression dynamics (how long a family can sustain activity before declining) give a modest signal for high-salinity performance.

4. Predictive signal is concentrated in the low-salinity parentage group

Across both survival conditions, the low-salinity-origin families consistently produce the strongest predictive signal. The high-salinity-origin families show near-zero or negative correlations for both conditions (visible in the comparison scatter as orange circles near the origin), and no high-salinity parentage feature reaches significance. This may reflect greater phenotypic variation in metabolic rate among low-salinity-origin families, or a tighter coupling between metabolic phenotype and survival outcome in that cohort.

Figures

feature_phenotype_correlation_ranking.png
Spearman ρ for the top 8 features per parentage group and survival condition; blue = significant (p < 0.05). Low-salinity survival (bottom row) shows a cluster of significant positive predictors in the low-salinity parentage group; high-salinity survival (top row) shows only one.

feature_comparison_high_vs_low_salinity.png
Scatter of each feature’s ρ for high-salinity survival (x) vs low-salinity survival (y). Features above the dashed line predict low-salinity survival better; below predict high-salinity better. Capacity features (vmax, AUC) cluster upper-center; timing/depression features sit to the right.

best_single_predictor_scatter.png
(A) Best single predictor of high-salinity survival: time_to_min_slope in low-salinity families (ρ = 0.42, CV ρ = 0.27) — wide confidence interval, moderate scatter. (B) Best single predictor of low-salinity survival: vmax in low-salinity families (ρ = 0.61, CV ρ = 0.48) — tighter fit, clearer trend.

resazurin_index_scatter.png
Multi-feature composite index vs survival. (A) High-salinity: index fails cross-validation (CV ρ = −0.26); family 78 drives overfitting. (B) Low-salinity: index generalises well (CV ρ = 0.53); families 48 and 54 with high index values show highest low-salinity survival.

family_profile_heatmap.png
Z-scored family means for top 10 features, families ordered left-to-right from best to worst survival. (A) High-salinity: no clear colour gradient left-to-right, consistent with weak signal. (B) Low-salinity: visible gradient — left-side families (48, 54, 79) are redder on capacity features (vmax, auc_early, initial_slope), right-side families (57, 55, 84) are bluer.

Implications for Predictive Phenotyping

Resazurin is a viable screening tool for low-salinity survival. A 4-hour assay on low-salinity-origin families recovers the survival ranking with LOFO-CV ρ ≈ 0.50, which is meaningful given that the comparison families are different animals from the survival experiment. The key metric is early metabolic rate (Vmax or initial slope) — simple, fast, and non-lethal to measure.
High-salinity survival likely requires a different assay design. The near-zero predictive power for high-salinity survival suggests that the metabolic feature relevant to that condition — the timing of metabolic depression — may not be well-captured in a standard 4-hour resazurin incubation, or that survival under high-salinity is more strongly determined by genetic or developmental factors not reflected in resting metabolic phenotype. A longer or thermally challenged assay may improve signal.
A single universal resazurin metric cannot serve both conditions. The features that predict low-salinity survival (capacity) are orthogonal to those that weakly predict high-salinity survival (timing/depression). Screening programs should select the metric matched to the target condition.