Welcome to Make Sense of Science, where we present a publication in a short, concise form. Today, it's dynamicGP, an interpretable machine learning model for predicting how plant traits evolve over time, published in Nature Plants:

Predicting plant trait dynamics from genetic markers
David Hobby, Hao Tong, Marc Heuermann, Alain J. Mbebi, Roosa A. E. Laitinen, Matteo Dell’Acqua, Thomas Altmann & Zoran Nikoloski

https://helda.helsinki.fi/server/api/core/bitstreams/89abf6e5-7455-4a21-8a65-d36c3bcb5f27/content
Hobby, D., Tong, H., Heuermann, M., Mbebi, A. J., Laitinen, R. A., Dell’Acqua, M., ... & Nikoloski, Z. (2025). Predicting plant trait dynamics from genetic markers. Nature plants, 1-10.

Highlight

Plants don’t just grow—they develop in complex ways over time. Understanding how specific traits like size, shape, or color change day by day can help breeders select the best plants. But tracking this for every plant is time-consuming and costly. This study introduces a new method called dynamicGP (GP for genomic prediction), which combines genetic information with mathematical modeling to predict how plant traits change over time. With this tool, scientists can forecast the entire growth pattern of a plant from just its DNA and a few early measurements.

Methods Used

• High-throughput phenotyping (HTP): Automated imaging systems captures plant traits.
• Genomic Prediction (GP): Machine learning uses plant DNA (genetic markers) to predict traits.
• Dynamic Mode Decomposition (DMD): A math tool that simplifies complex time-based changes into patterns or “modes”.
• dynamicGP: A new combination of GP and DMD to forecast how plant traits evolve over time.

To build this new prediction model, the researchers combined two advanced tools. First, they used high-throughput phenotyping, which involves taking large numbers of automated images to record traits like plant height, leaf color, and shape across many timepoints. Second, they used genomic prediction, which links differences in DNA to differences in traits using machine learning. The novelty of their method lies in adding dynamic mode decomposition—a mathematical technique often used in physics and engineering to understand how systems evolve over time. This allowed them to capture not just what traits plants have, but how those traits change throughout the plant’s life. The combined model, called dynamicGP, was tested on maize and Arabidopsis plants to see how well it could predict trait development over several weeks.

 

Fig.: Forecasting traits like plant growth using DNA, time-series multi-source imaging, and dynamic modeling with dynamicGP. (After Hobby et al. (2025)¹, Jin et al. (2021)², Erichson et al. (2019)³)

Results and Discussion

The dynamicGP method was more accurate than standard prediction models that only use genetic information from one point in time. It performed well in forecasting how traits like plant color and shape evolved throughout development. The researchers tested two ways of using the model: one that updated predictions using new measurements at each time point (called “iterative”), and another that made all future predictions based on the first measurement alone (called “recursive”). While both approaches worked, the iterative version was generally more accurate. Traits that had more consistent genetic influence over time were predicted more reliably. For example, some color traits reached prediction accuracies above 80%, far outperforming traditional models. Even in Arabidopsis, which had more variation and lower trait stability, dynamicGP still outperformed the baseline approach in most cases.

Guidance for Breeders and Agriculture

• Fewer field visits: Breeders can now predict how a plant will develop with fewer measurements.
• Focus on stable traits: Traits that don’t change much in heritability over time are best suited for prediction.
• Better planning: Understand when and how specific traits emerge, aiding in selection decisions.

This new approach has clear benefits for plant breeding and agricultural decision-making. By using dynamicGP, breeders can make informed predictions about how plants will grow without needing to collect measurements at every stage. This reduces labor and cost, while still allowing for high-quality selection. The method is especially useful for traits that show consistent genetic patterns over time, making it easier to focus on selecting the best plants early in the breeding process.