Study compares RF, MLP, and Boolean networks for plant gene regulation

Researchers led by Gonzalo A. Ruz compared continuous surrogate models (Random Forest regression and Multi-Layer Perceptron) against a discrete threshold Boolean network (TBN) for modeling gene regulation in Arabidopsis thaliana's induced systemic resistance (ISR). The study used eight defense-related genes measured over nine time points, evaluating models on rolling-origin one-step prediction, recursive multi-step rollout, and interpretability. In continuous domain predictions, RF outperformed MLP with a Mean Absolute Error (MAE) of 1.910 vs 2.089 and Root Mean Squared Error (RMSE) of 2.836 vs 3.106. However, in binary domain classification, TBN achieved the best one-step qualitative performance (binary accuracy 0.550, Hamming distance 3.600) versus RF (0.500, 4.000) and MLP (0.495, 4.040).

More strikingly, during recursive rollout (multi-step prediction), the TBN exactly reproduced the observed binarized trajectory, while MLP showed near-perfect fidelity (0.986 binary accuracy). RF accumulated significantly more deviation, with only 0.708 trajectory accuracy. This highlights a critical trade-off: models with superior local numerical accuracy (RF) can fail to capture global dynamical behavior, while simpler discrete models (TBN) achieve perfect qualitative alignment. The authors argue that continuous surrogates and Boolean networks should be viewed as complementary—continuous models for numerical precision, discrete models for capturing regulatory logic and long-term dynamics. The paper is to be published at IEEE CIBCB 2026 and underscores the importance of choosing modeling approaches based on specific biological questions.