Biology subjects

Caragea, D.

Publications and source records attributed to Caragea, D..

4 recordsLinked to original sources

bioRxiv2025

Deep Learning for Sorghum Yield Forecasting using Uncrewed Aerial Systems and Lab-Derived Imagery

The AI revolution, advanced Graphics Processing Units (GPUs), and open-source platforms have enabled Machine Learning (ML) and Deep Learning (DL) algorithms to rapidly and accurately extract phenotypic features from Uncrewed Aerial System (UAS)-derived imagery. Such advancement leads to phenotypic digitization and sorghum yield forecasting. Yield analytics are critical for breeding programs to assess the genetics and breeding potential of genotypes to enhance cultivar development. This trial followed a three-replicated Randomized Complete Block Design (RCBD) with 36 diverse sorghum genotypes in 2023 at Ashland Bottoms, Kansas. The field images were captured 6 meters above using a DJI M300 drone equipped with the P1 sensor at 90{degrees}nadir and 45{degrees} oblique angles. This research trained YOLO and the Faster R-CNN (Detectron2) models to harness yield attributes from UAS field and lab images. The YOLO models outperformed the Faster R-CNN model in detecting sorghum panicles, achieving a mean average precision at 50% IoU (mAP@0.50) ranging from 0.92 to 0.98, compared to 0.61 to 0.89. Panicle detection from field imagery correlated at 0.86 with ground truth. Lab imagery analyses measured panicle size, seed counts, and seed area with correlation coefficients of 0.71, 0.95, and 0.25, respectively. Three machine learning models: Support Vector Regression (SVR), Decision Tree Regression (DTR), and Random Forest Regression (RFR) are used to predict yield with correlation coefficients of 0.58, 0.76, and 0.70, respectively. We observed that YOLO models are well-suited for extracting yield-attributing traits from images, which are then incorporated into ML regression models to improve yield prediction performance.

plant biology↗

bioRxiv2024

CLR_ESP: Improved enzyme-substrate pair prediction using contrastive learning

To reduce the cost of experimental characterization of the potential substrates for enzymes, machine learning prediction model offers an alternative solution. Pretrained language models, as powerful approaches for protein and molecule representation, have been employed in the development of enzyme-substrate prediction models, achieving promising performance. In addition to continuing improvements in language models, effectively fusing encoders to handle multimodal prediction tasks is critical for further enhancing model performance using available representation methods. Here, we present FusionESP, a multimodal architecture that integrates protein and chemistry language models with a newly designed contrastive learning strategy for predicting enzyme-substrate pairs. Our best model achieved state-of-the-art performance with an accuracy of 94.77% on independent test data and exhibited better generalization capacity while requiring fewer computational resources and training data, compared to previous studies of finetuned encoder or employing more encoders. It also confirmed our hypothesis that embeddings of positive pairs are closer to each other in high-dimension space, while negative pairs exhibit the opposite trend. The proposed architecture is expected to be further applied to enhance performance in additional multimodality prediction tasks in biology. A user-friendly web server of FusionESP is established and freely accessible at https://rqkjkgpsyu.us-east-1.awsapprunner.com/.

bioinformatics↗

bioRxiv2023

Using recurrent neural networks to detect supernumerary chromosomes in fungal strains causing blast diseases

The genomes of the fungus Magnaporthe oryzae that causes blast diseases on diverse grass species, including major crop plants, have indispensable core-chromosomes and may contain one or more additional supernumerary chromosomes, also known as mini-chromosomes. The mini-chromosome is speculated to play a role in fungal biology, provide effector gene mobility, and may transfer between strains. To understand and study the biological function of mini-chromosomes, it is crucial to be able to identify whether a given strain of M. oryzae possesses a mini-chromosome. In this study, we applied recurrent neural network models, more specifically, Bidirectional Long Short-Term Models (Bi-LSTM), for classifying DNA sequences as core-or mini-chromosomes. The models were trained with sequences from multiple available core- and mini-chromosome assemblies. The trained model was then used to predict the presence of the mini-chromosome in a global collection of M. oryzae isolates using short-read DNA sequences. The model predicted that the mini-chromosome was prevalent in M. oryzae isolates, including those isolated from rice, wheat, Lolium and many other grass species. Interestingly, 23 recent wheat strains collected since 2005 all carried the mini-chromosome, but none of nine early strains collected before 1991 had the mini-chromosome, indicating the preferential selection for strains carrying the mini-chromosome in recent years. Based on the limited sample size, we found the presence of the mini-chromosome in isolates of pathotype Eleusine was not as high as isolates of other pathotypes. The deep learning model was also used to identify assembled sequence contigs that were derived from the mini-chromosome and partial regions on core-chromosomes potentially translocated from a mini-chromosome. In summary, our study has developed a reliable method for categorizing DNA sequences and showcases an application of recurrent neural networks in the field of predictive genomics.

genomics↗

bioRxiv2018

Root Anatomy based on Root Cross-Section Image Analysis with Deep Learning

Aboveground plant efficiency has improved significantly in recent years, and the improvement has led to a steady increase in global food production. The improvement of belowground plant efficiency has potential to further increase food production. However, belowground plant roots are harder to study, due to inherent challenges presented by root phenotyping. Several tools for identifying root anatomical features in root cross-section images have been proposed. However, the existing tools are not fully automated and require significant human effort to produce accurate results. To address this limitation, we use a fully automated approach, specifically, the Faster Region-based Convolutional Neural Network (Faster R-CNN), to identify anatomical traits in root cross-section images. By training Faster R-CNN models on root cross-section images, we can detect objects such as root, stele and late metaxylem, and predict rectangular bounding boxes around such objects. Subsequently, the bounding boxes can be used to estimate the root diameter, stele diameter, late metaxylem number, and average diameter. Experimental evaluation using standard object detection metrics, such as intersection-over-union and mean average precision, has shown that the Faster R-CNN models trained on rice root cross-section images can accurately detect root, stele and late metaxylem objects. Furthermore, the results have shown that the measurements estimated based on predicted bounding boxes have small root mean square error when compared with the corresponding ground truth values, suggesting that Faster R-CNN can be used to accurately detect anatomical features. A webserver for performing root anatomy using the Faster R-CNN models trained on rice images is available at https://rootanatomy.org, together with a link to a GitHub repository that contains a copy of the Faster R-CNN code. The labeled images used for training and evaluating the Faster R-CNN models are also available from the GitHub repository.

bioinformatics↗