LeafLens — Physics-Informed ML for CRISPR Stomatal Engineering in Rice

A physics-informed machine-learning framework that predicts leaf gas exchange (carbon assimilation $A$ and stomatal conductance $g_s$) in CRISPR stomatal-engineered rice, built on a harmonized 165-measurement dataset compiled from three published studies and validated with genotype-grouped cross-validation.

Authors: P. Thrishna Sai, N. Jenika — Centre for Biotechnology, JNTUH-UCESTH, Hyderabad

---

What this project does

CRISPR editing of stomatal-patterning genes (OsEPF1, OsEPFL10, OsSTOMAGEN) can reduce the density of stomatal pores on a rice leaf, cutting water loss. The design question is how much reduction is appropriate: fewer pores save water but also restrict the CO₂ supply that drives photosynthesis.

LeafLens models that trade-off. It combines a light-response saturation prior from photosynthesis physiology with a ridge-regression residual learner:

y = y_max · PPFD/(PPFD + K_m) · S^β   +   f_ridge(S, PPFD, VPD, drought, T, CO₂)
        └────────── physics prior ──────────┘   └──────── learned residual ────────┘

where S is stomatal density relative to wild type.

Headline results (genotype-grouped CV, 20 randomized fold assignments)

Target	R²	Variance ceiling	% of achievable
Assimilation A	0.531 ± 0.166	0.830	64%
Conductance g_s	0.362 ± 0.152	0.501	72%

The physics prior earns its place: it improves R² by +0.116 ± 0.091 (A) and +0.042 ± 0.023 (g_s) over an otherwise identical model without it, winning in 20 of 20 paired repeats for both targets.

Robustness, measured rather than assumed: nested cross-validation (running the whole model selection inside the validation loop) puts selection-corrected performance at R² = 0.43 (A) and 0.17 (g_s); a permutation test rejects the null at p < 0.001 for both; and measured digitization error (0.41% of axis span) does not affect results until ~25× that level. The inner selection loop independently rediscovers the physics-informed ridge architecture in most folds. Three further checks: the 59 imputed rows do not inflate results (the model predicts them worse than measured rows, 0.03 vs 0.81); the physics functional form is immaterial (< 0.03 R² across three saturating forms); and all 165 rows fall within established physiological ranges for C3 rice.

Within the validated domain, a 30% stomatal-density reduction is predicted to retain 94.9% of wild-type assimilation while reducing conductance to 92.0% — water loss falls faster than carbon gain, which is the quantitative basis for stomatal engineering.

---

Three findings that constrain how far this can be pushed

These are reported as prominently as the performance numbers, because they bound it.

1. Intrinsic water-use efficiency (A/g_s) is not a learnable target here. An ANOVA intraclass-correlation analysis gives it an R² ceiling of 0.00 — its replicate noise is 2.7× its genotype effect. Modelling A and g_s separately works; modelling their ratio cannot. We recommend running this diagnostic before modelling any derived/ratio target.

2. Cross-study transfer is not demonstrated. Leave-one-study-out R² is strongly negative, and normalizing to within-study wild-type controls does not repair it (−0.28 for A, −0.04 for g_s). With three studies these estimates are unstable enough that no signed value should be trusted.

3. The decision-relevant range is nearly unmeasured. Across the whole published literature, stomatal-reduction levels jump from 28.5% to 58.0% — a 29.5-point gap — and the entire 30–70% band rests on a single genotype (5 of 165 measurements). Predictive reliability also declines with reduction severity (Spearman ρ = −0.50 for A, −0.58 for g_s).

---

Repository layout

scripts/
  parse_digitized_data.py     # build bio_master.csv from digitized figure data
  build_feautures.py          # feature engineering + climate PPFD conversion
  train_final_model.py        # FINAL model: physics prior + ridge, all validation
  run_supporting_analyses.py  # reproduces every remaining number in the paper
  run_robustness_checks.py    # digitization error, nested CV, permutation test
  make_final_figures.py       # all manuscript figures
  run_eda.py                  # exploratory plots + extrapolation diagnostic
  download_climate.py         # NASA POWER retrieval
  train_models.py             # legacy iWUE-target pipeline (superseded; see paper §4.7)
  optimize_sweep.py, run_shap.py, run_ablation.py   # legacy analyses
data/
  biological/raw/             # WebPlotDigitizer exports from source figures
  processed/training_data.csv # final 165-row feature matrix
  climate/                    # NASA POWER daily records, 5 Telangana districts
outputs/tables/               # all result tables (CSV/JSON)
outputs/figures/              # all generated figures
paper/
  main.tex                    # manuscript (Springer llncs proceedings class)
  LeafLens_paper.pdf          # compiled PDF
papers/                       # source publications (PDFs)
PROJECT_NOTES.md              # full methodology, corrections log, and rationale

Reproducing

python scripts/parse_digitized_data.py      # rebuild dataset from digitized figures
python scripts/build_feautures.py           # feature matrix + climate conversion
python scripts/train_final_model.py         # model, validation, ceiling analysis
python scripts/run_supporting_analyses.py   # leakage, transfer, coverage analyses
python scripts/run_robustness_checks.py     # digitization error, nested CV, permutation
python scripts/make_final_figures.py        # figures

Every quantitative claim in the manuscript is produced by one of these scripts and written to outputs/tables/.

Data provenance

All gas-exchange values are digitized from published figures using WebPlotDigitizer, or taken from values stated numerically in the source text. Each row carries an A_Source flag: digitized, digitized_genotype_mean, or estimated_regression (59 rows where the source study reported conductance but not paired assimilation under drought; estimated via a regression fitted on that study's own steady-state data, R²=0.931).

Sources: Caine et al. 2019 (New Phytologist, 10.1111/nph.15344) · Karavolias et al. 2023 (Plant Physiology, 10.1093/plphys/kiad183) · Karavolias et al. 2024 (Plant Biotechnology Journal, 10.1111/pbi.14464) · climate from NASA POWER.

Compiling the manuscript

paper/main.tex targets Springer's llncs (Lecture Notes in Computer Science) proceedings class. Unlike a Springer Nature journal class, llncs is a standard CTAN/TeX-Live package, so Overleaf and any normal LaTeX install supply it automatically — nothing extra needs to be downloaded or bundled.

Test-compiled with tectonic: 20 pages, 0 errors. One thing to know: \ackname, \discintname and the credits environment used in the acknowledgments/disclosure section were only added to llncs.cls in v2.25 (2026); if your installed copy predates that, main.tex carries a small compatibility shim (right after \documentclass) that defines them only if missing, so the file builds identically either way — do not remove it.

LeafLens_overleaf.zip (repo root) is a ready-to-upload package: main.tex plus the 9 referenced figures, nothing else. Overleaf → New Project → Upload Project → select it.

Status

This is a research prototype and a methods contribution, not a deployment-ready trait-design tool. See PROJECT_NOTES.md for the full correction history and the reasoning behind each modelling decision.