Journal: bioRxiv

Article Title: Context-Aware Transcript Quantification from Long Read RNA-Seq data with Bambu

doi: 10.1101/2022.11.14.516358

Figure Lengend Snippet: (a) PR curves for the performance of transcript discovery when using minimum, mean, and maximum to combine TPS across samples for the same read class on all HepG2 samples together without annotations for human chromosome 1. The performance of using the sum of read counts as the classifier across the samples is used as a comparison (b) A precision recall curve showing the performance of Bambu on PacBio data using either the PacBio trained model (purple), the pretrained model (green), or ranking read classes by gene proportion (orange) or read count (blue). The grey shaded area represents the mean +/- SE of the precision for each line. (c) A precision recall curve showing the performance of models pre-trained on human (purple), mouse (blue) or arabidopsis (green) data applied to another arabidopsis tissue. Additionally the performance of the sample trained model is included (orange). The grey shaded area represents the mean +/- SE of the precision for each line. (d) The precision and sensitivity of varying Bambu thresholds when looking at a subset of read classes divided into expression quantiles. The same model is applied to all subsets. The full data is coloured in red, read classes that have expressions ranging from 0 to the lower quartile are shaded in yellow, those ranging between the lower quartile and the median in green, between the median and upper quartile in blue and the upper quartile to the max in red. Each subset should represent approximately 25% of the read classes. The lowest expressed quartile is larger than the others due to a greater than 25% of read classes sharing a read count of 2. (e) The precision and sensitivity of varying Bambu thresholds when looking at a subset of read classes divided by the number of expressed isoforms (> 2 read count) their gene contains. The same model is applied to all subsets. The full data is coloured in red, the subset of read classes that are the only expressed isoforms in their gene are coloured purple, those which have two or more isoforms are shaded in teal, and those with five or more isoforms are green. (f) Precision and Recall curves of the classification using Bambu models trained using reference annotations missing a random fraction of annotations used from the human reference annotations (excluding chromosome 1). The models trained using these annotations, are used to classify read classes from chromosome 1. The Pretrained Model represents the in-built model in Bambu which is used when the annotations do not support training and Read Count classifies the read classes solely using read count alone. These were applied to all SG-NEx datasets. (g) We measured the difference in ROC AUC of trained and pretrained models in which the trained model was trained with missing reference annotations. Samples in which Bambu recommends using the pretrained model are coloured red (the recommended NDR was calculated as > 0.5), and samples where the trained model is used are coloured blue. (h) A box plot showing the distribution of NDR recommendations of SG-NEx samples when Bambu was run with different percentages of reference annotations. (i) The average sensitivity and precision of transcript discovery on core SG-NEx samples without annotations for human chromosome 1. Each tool is displayed at several different parameter thresholds: Bambu (blue) with NDR thresholds varying between 1 and 0.1, FLAIR (red), StringTie2 (green) and TALON (greygray) with read count/coverage thresholds varying between 2 and 10, with 4 additional thresholds for StringTie2 at 15, 20, 30 and 50. Horizontal error bars represent the mean +/- SD of the sensitivity and vertical error bars represent the mean +/- SD of the precision (j) The measured sensitivity and precision of transcript discovery when combining HepG2 samples, without annotations for human chromosome 1. Each tool is displayed at several different parameter thresholds: Bambu (blue) with NDR thresholds varied between 1 and 0.1, StringTie2 (green) and TALON (grey) with read count/coverage thresholds varying between 2 and 10, and 4 additional thresholds for StringTie2 at 15, 20, 30 and 50 (k) The average sensitivity and precision of transcript discovery on spike-in data without annotations for the spike-in chromosome. Each tool is displayed at several different parameter thresholds: Bambu (blue) with NDRthresholds varying between 1 and 0.1, StringTie2 (green), FLAIR (red) and TALON (grey) with read count/coverage thresholds varied between 2 and 10. Horizontal error bars represent the mean +/- SD of the sensitivity and vertical error bars represent the mean +/- SD of the precision (l) The average sensitivity and precision from transcript discovery outputs run on SG-NEx data with 50% of the spike-in annotations randomly removed. Each tool is displayed at several different parameter thresholds. Bambu (Blue) was run using novel discovery thresholds between 1 and 0.1. StringTie2 (green), FLAIR (red) and TALON (grey) were run with read count/coverage thresholds between 2 and 10. Error bars represent the standard error

Article Snippet: However, for cases where the pretrained model is not suitable due to species and technology differences, Bambu provides the option to train a new model on related data with comprehensive annotations that is more appropriate for the sample of interest (see Supplementary Text Section 5 for additional information).

Techniques: Comparison, Expressing

Journal: Sensors (Basel, Switzerland)

Article Title: SensorNet : An Adaptive Attention Convolutional Neural Network for Sensor Feature Learning

doi: 10.3390/s24113274

Figure Lengend Snippet: Performance of the pretrained SensorNet obtained from SLEEP-EDF-20.

Article Snippet: This experiment further demonstrates the superiority of the pretrained SensorNet in terms of its ability to transfer to a small sensor dataset regardless of input size.

Techniques:

Journal: International Journal of Computer Assisted Radiology and Surgery

Article Title: Attention-guided erasing for enhanced transfer learning in breast abnormality classification

doi: 10.1007/s11548-024-03317-6

Figure Lengend Snippet: Overview of the Attention-Guided Erasing (AGE) Methodology. a Self-Supervised Pretraining using DINO [11]: A teacher student ViT framework, leveraging a teacher-student ViT-S self-distillation framework. b AGE [13]: Attention head visualizations from the SSL pretrained teacher ViT-S are converted into binary masks to isolate key ROIs and then used to erase background regions. c Transfer Learning with AGE: AGE is used on the input images using each of the attention heads with a random probability during training. The attention head yielding the highest validation performance is selected for final AGE-based transfer learning

Article Snippet: Input image followed by six attention maps from each of the five pretrained DINO models associated with specific tasks: T1 (Breast Density in DM), T2 (Malignancy in CEM), T3 (Calcification ROI in DM), T4 (Malignancy ROI in CEM), and T5 (Mass ROI in DM).

Techniques: Distillation, Biomarker Discovery

Journal: International Journal of Computer Assisted Radiology and Surgery

Article Title: Attention-guided erasing for enhanced transfer learning in breast abnormality classification

doi: 10.1007/s11548-024-03317-6

Figure Lengend Snippet: Attention Head Visualizations. Input image followed by six attention maps from each of the five pretrained DINO models associated with specific tasks: T1 (Breast Density in DM), T2 (Malignancy in CEM), T3 (Calcification ROI in DM), T4 (Malignancy ROI in CEM), and T5 (Mass ROI in DM). The final selected attention heads used for transfer learning are highlighted in red

Article Snippet: Input image followed by six attention maps from each of the five pretrained DINO models associated with specific tasks: T1 (Breast Density in DM), T2 (Malignancy in CEM), T3 (Calcification ROI in DM), T4 (Malignancy ROI in CEM), and T5 (Mass ROI in DM).

Techniques: