Fachsymposium "Artificial Intelligence for Life" 2024

Key Note: Prof. Dr. Karsten Borgwardt, Max-Planck-Institut
Machine Learning and the Future of Bioinformatics

Sebastian Burkhart, HSWT
Predicting Grain Yield with AI: Effects of Train-Split-Ratio (TSR) and Network Architectures

In this study, we explore the impact of the Train-Split-Ratio (TSR) on spectral grain yield prediction using AI, analyzing overall seven field trials from two locations over three and four years. We compare six algorithms across a TSR range from 5% to 95%, evaluating their performance based on a red edge vegetation index. Additionally, we investigate the optimization of neural network architectures to enhance prediction accuracy. Our findings provide insights into the optimal balance of training and test data and effective neural network designs for yield forecasting in agricultural trials.

Andreas Gilson, Fraunhofer IIS
FruitNeRF: Revolutionary Fruit Counting Combining the Power of NeRFs and Foundation Models

This talk will present FruitNeRF as part of the For5G project that has the overall goal of creating of end-to-end pipelines for digital twins in horticulture. FruitNeRF is a novel unified fruit counting framework that leverages state-of-the-art view synthesis methods to count any fruit type directly in 3D. Utilizing neural radiance fields and the foundation model SAM, it becomes possible to count arbitrary types of fruit based on an unordered set of 2D images without extensive manual labeling. The presented method also prevents typical pitfalls in fruit counting, like double counting or counting of fallen or background fruit and was evaluated on synthetic and real-world datasets.

Forecast-based energy management can play a large role in a smarter and more efficient use of renewable energies based on demand side management. Using approaches such as model predictive control, individual consumption devices can be shifted within operation constraints so that their electricity consumption optimally matches generation. In agriculture, large thermal storages make up a sizeable part of electricity consumption, and offer a potential use in the short term shifting of demand. Necessary for this are accurate models to forecast behaviour of such dynamic systems, so that minimal power demand and fulfilment of operation constraints can be ensured when computing optimal controls.
In this talk, different approaches to the modelling of thermal storages are presented, with a focus on the model training process through Parameter Identification. Model evaluation is conducted with the models’ use within a forecast-based Energy Management System in mind, and as an outlook, the operation of such a system is illustrated with preliminary results.

• Deep learning has become an indispensable tool for the biological sciences. However, numerous research articles indicate that reported results are often overly optimistic and not reproducible on independent data. This issue is particularly pressing in the life sciences, where developed methods ultimately aim to be applied to humans. One of the main reasons for the discrepancy between reported and actual performance is data leakage – the illicit sharing of information between the training and the test set.
  The field of protein-protein interaction (PPI) prediction is no exception to this issue. As it is not feasible to study all protein pairs exhaustively, many sequence-based approaches have been developed to predict PPIs as a binary classification task.
  We show that the phenomenal prediction accuracies reported for these methods are exclusively due to data leakage [1]. The leakage is introduced through an incorrect split of the input dataset, as well as dataset design. The issues stem from study bias in PPI networks, flawed generation of negative examples, an incorrect simulation of the application scenario in the test set, sequence similarities, and biased embeddings. Learned shortcuts artificially inflate the model performance.
  We supply a large, leakage-free gold standard dataset and show that a node degree-preserving randomization of the training data still makes accuracies around 90% possible. Likewise, leaking only one protein from the test set into the training data yields results around 90%. However, when all proteins from the test set are unseen and sequence similarity is minimized between the sets, performances drop to random. Our study highlights that shortcuts used in training machine learning obscure the true extent of the challenge and thus prevent new breakthroughs, especially for unseen or understudied proteins. Our study reveals over-optimism regarding the progress of sequence-based binary PPI prediction and shows that this challenge is far from being solved.

References
[1] Judith Bernett, David B Blumenthal, Markus List, Cracking the black box of deep sequence-based
protein–protein interaction prediction, Briefings in Bioinformatics, Volume 25, Issue 2, March 2024, bbae076,
https://doi.org/10.1093/bib/bbae076