Prediction of T-cell receptor (TCR) interactions with MHC-peptide complexes remains highly challenging. This challenge is primarily due to three dominant factors: data accuracy, data scarceness, and problem complexity. Here, we showcase that “shallow” convolutional neural network (CNN) architectures are adequate to deal with the problem complexity imposed by the length variations of TCRs. We demonstrate that current public bulk CDR3β-pMHC binding data overall is of low quality and that the development of accurate prediction models is contingent on paired α/β TCR sequence data corresponding to at least 150 distinct pairs for each investigated pMHC. In comparison, models trained on CDR3α or CDR3β data alone demonstrated a variable and pMHC specific relative performance drop. Together these findings support that T-cell specificity is predictable given the availability of accurate and sufficient paired TCR sequence data. NetTCR-2.0 is publicly available at https://services.healthtech.dtu.dk/service.php?NetTCR-2.0.

Autoencoders are commonly used in representation learning. They consist of an encoder and a decoder, which provide a straightforward method to map n-dimensional data in input space to a lower m-dimensional representation space and back. The decoder itself defines an m-dimensional manifold in input space. Inspired by manifold learning, we showed that the decoder can be trained on its own by learning the representations of the training samples along with the decoder weights using gradient descent. A sum-of-squares loss then corresponds to optimizing the manifold to have the smallest Euclidean distance to the training samples, and similarly for other loss functions. We derived expressions for the number of samples needed to specify the encoder and decoder and showed that the decoder generally requires much fewer training samples to be well-specified compared to the encoder. We discuss the training of autoencoders in this perspective and relate it to previous work in the field that uses noisy training examples and other types of regularization. On the natural image data sets MNIST and CIFAR10, we demonstrated that the decoder is much better suited to learn a low-dimensional representation, especially when trained on small data sets. Using simulated gene regulatory data, we further showed that the decoder alone leads to better generalization and meaningful representations. Our approach of training the decoder alone facilitates representation learning even on small data sets and can lead to improved training of autoencoders. We hope that the simple analyses presented will also contribute to an improved conceptual understanding of representation learning.

Differential gene expression analysis of bulk RNA sequencing data plays a major role in the diagnosis, prognosis, and understanding of disease. Such analyses are often challenging due to a lack of good controls and the heterogeneous nature of the samples. Here, we present a deep generative model that can replace control samples. The model is trained on RNA-seq data from healthy tissues and learns a low-dimensional representation that clusters tissues very well without supervision. When applied to cancer samples, the model accurately identifies representations close to the tissue of origin. We interpret these inferred representations as the closest normal to the disease samples and use the resulting count distributions to perform differential expression analysis ofsinglecancer sampleswithoutcontrol samples. In a detailed analysis of breast cancer, we demonstrate how our approach finds subtype-specific cancer driver and marker genes with high specificity and greatly outperforms the state-of-the-art method in detecting differentially expressed genes, DESeq2. We further show that the significant genes found using the model are highly enriched within cancer-specific driver genes across different cancer types. Our results show that thein silicoclosest normal provides a more favorable comparison than control samples.

Motivation Learning low-dimensional representations of single-cell transcriptomics has become instrumental to its downstream analysis. The state of the art is currently represented by neural network models such as variational autoencoders (VAEs) which use a variational approximation of the likelihood for inference. Results We here present the Deep Generative Decoder (DGD), a simple generative model that computes model parameters and representations directly via maximum a posteriori (MAP) estimation. The DGD handles complex parameterized latent distributions naturally unlike VAEs which typically use a fixed Gaussian distribution, because of the complexity of adding other types. We first show its general functionality on a commonly used benchmark set, Fashion-MNIST. Secondly, we apply the model to multiple single-cell data sets. Here the DGD learns low-dimensional, meaningful and well-structured latent representations with sub-clustering beyond the provided labels. The advantages of this approach are its simplicity and its capability to provide representations of much smaller dimensionality than a comparable VAE. Availability and implementation The code is made available in this GitHub repository. scDGD is available as a python package at https://github.com/Center-for-Health-Data-Science/scDGD. Supplementary information Supplementary data are available at Bioinformatics online.

A deep generative model is characterized by a representation space, its distribution, and a neural network mapping the representation to a distribution over vectors in feature space. Common methods such as variational autoencoders (VAEs) apply variational inference for training the neural network, but optimizing these models is often non-trivial. The encoder adds to the complexity of the model and introduces an amortization gap and the quality of the variational approximation is usually unknown. Additionally, the balance of the loss terms of the objective function heavily influences performance. Therefore, we argue that it is worthwhile to investigate a much simpler approximation which finds representations and their distribution by maximizing the model likelihood via back-propagation. In this approach, there is no encoder, and we therefore call it a Deep Generative Decoder (DGD). Using the CIFAR10 data set, we show that the DGD is easier and faster to optimize than the VAE, achieves more consistent low reconstruction errors of test data, and alleviates the problem of balancing the reconstruction and distribution loss terms. Although the model in its simple form cannot compete with stateof-the-art image generation approaches, it obtains better image generation scores than the variational approach on the CIFAR10 data. We demonstrate on MNIST data how the use of a Gaussian mixture with priors can lead to a clear separation of classes in a 2D representation space, and how the DGD can be used with labels to obtain a supervised representation.