Background Single-cell RNA sequencing (scRNA-seq) has emerged has a main strategy to study transcriptional activity at the cellular level. Clustering analysis is routinely performed on scRNA-seq data to explore, recognize or discover underlying cell identities. The high dimensionality of scRNA-seq data and its significant sparsity accentuated by frequent dropout events, introducing false zero count observations, make the clustering analysis computationally challenging. Even though multiple scRNA-seq clustering techniques have been proposed, there is no consensus on the best performing approach. On a parallel research track, self-supervised contrastive learning recently achieved state-of-the-art results on images clustering and, subsequently, image classification. Results We propose contrastive-sc, a new unsupervised learning method for scRNA-seq data that perform cell clustering. The method consists of two consecutive phases: first, an artificial neural network learns an embedding for each cell through a representation training phase. The embedding is then clustered in the second phase with a general clustering algorithm (i.e. KMeans or Leiden community detection). The proposed representation training phase is a new adaptation of the self-supervised contrastive learning framework, initially proposed for image processing, to scRNA-seq data. contrastive-sc has been compared with ten state-of-the-art techniques. A broad experimental study has been conducted on both simulated and real-world datasets, assessing multiple external and internal clustering performance metrics (i.e. ARI, NMI, Silhouette, Calinski scores). Our experimental analysis shows that constastive-sc compares favorably with state-of-the-art methods on both simulated and real-world datasets. Conclusion On average, our method identifies well-defined clusters in close agreement with ground truth annotations. Our method is computationally efficient, being fast to train and having a limited memory footprint. contrastive-sc maintains good performance when only a fraction of input cells is provided and is robust to changes in hyperparameters or network architecture. The decoupling between the creation of the embedding and the clustering phase allows the flexibility to choose a suitable clustering algorithm (i.e. KMeans when the number of expected clusters is known, Leiden otherwise) or to integrate the embedding with other existing techniques.

Motivation Single-cell RNA sequencing (scRNA-seq) provides transcriptomic profiling for individual cells, allowing researchers to study the heterogeneity of tissues, recognize rare cell identities and discover new cellular subtypes. Clustering analysis is usually used to predict cell class assignments and infer cell identities. However, the high sparsity of scRNA-seq data, accentuated by dropout events generates challenges that have motivated the development of numerous dedicated clustering methods. Nevertheless, there is still no consensus on the best performing method. Results graph-sc is a new method leveraging a graph autoencoder network to create embeddings for scRNA-seq cell data. While this work analyzes the performance of clustering the embeddings with various clustering algorithms, other downstream tasks can also be performed. A broad experimental study has been performed on both simulated and scRNA-seq datasets. The results indicate that although there is no consistently best method across all the analyzed datasets, graph-sc compares favorably to competing techniques across all types of datasets. Furthermore, the proposed method is stable across consecutive runs, robust to input down-sampling, generally insensitive to changes in the network architecture or training parameters and more computationally efficient than other competing methods based on neural networks. Modeling the data as a graph provides increased flexibility to define custom features characterizing the genes, the cells and their interactions. Moreover, external data (e.g. gene network) can easily be integrated into the graph and used seamlessly under the same optimization task. Availability and implementation https://github.com/ciortanmadalina/graph-sc. Supplementary information Supplementary data are available at Bioinformatics online.

The identification of transcription factor (TF) target genes is central in biology. A popular approach is based on the location by pattern-matching of potential cis-regulatory elements (CREs). During the last few years, tools integrating next-generation sequencing data have been developed to improve the performances of pattern-matching. However, such tools have not yet been comprehensively evaluated in plants. Hence, we developed a new streamlined method aiming at predicting CREs and target genes of plant TFs in specific organs or conditions. Our approach implements a supervised machine learning strategy, which allows to learn decision rule models using TF ChIP-chip/seq experimental data. Different layers of genomic features were integrated in predictive models: the position on the gene, the DNA-sequence conservation, the chromatin state, and various cis-regulatory element footprints. Among the tested features, the chromatin features were crucial for improving the accuracy of the method. Furthermore, we evaluated the transferability of predictive models across TFs, organs and species. Finally, we validated our method by correctly inferring the target genes of key TF controlling metabolite biosynthesis at the organ-level in Arabidopsis. We developed a tool -Wimtrap- to reproduce our approach in plant species and conditions/organs for which ChIP-chip/seq data are available. Wimtrap is a user-friendly R package that supports a R-shiny web interface and is provided with pre-built models that can be used to quickly get predictions of CREs and TF gene targets in different organs or conditions in Arabidopsis thaliana, Solanum lycopersicum, Oryza sativa, and Zea mays.

Single-cell RNA sequencing (scRNA-seq) produces transcriptomic profiling for individual cells. Due to the lack of cell-class annotations, scRNA-seq is routinely analyzed with unsupervised clustering methods. Because these methods are typically limited to producing clustering predictions (that is, assignment of cells to clusters of similar cells), numerous model agnostic differential expression (DE) libraries have been proposed to identify the genes expressed differently in the detected clusters, as needed in the downstream analysis. In parallel, the advancements in neural networks (NN) brought several model-specific explainability methods to identify salient features based on gradients, eliminating the need for external models.We propose a comprehensive study to compare the performance of dedicated DE methods, with that of explainability methods typically used in machine learning, both model agnostic (such as SHAP, permutation importance) and model-specific (such as NN gradient-based methods). The DE analysis is performed on the results of 3 state-of-the-art clustering methods based on NNs. Our results on 36 simulated datasets indicate that all analyzed DE methods have limited agreement between them and with ground-truth genes. The gradients method outperforms the traditional DE methods, which en-courages the development of NN-based clustering methods to provide an out-of-the-box DE capability. Employing DE methods on the input data preprocessed by clustering method outperforms the traditional approach of using the original count data, albeit still performing worse than gradient-based methods.

We propose a framework using contrastive learning as a pre-training task to perform image classification in the presence of noisy labels. Recent strategies, such as pseudo-labeling, sample selection with Gaussian Mixture models, and weighted supervised contrastive learning have, been combined into a fine-tuning phase following the pre-training. In this paper, we provide an extensive empirical study showing that a preliminary contrastive learning step brings a significant gain in performance when using different loss functions: non robust, robust, and early-learning regularized. Our experiments performed on standard benchmarks and real-world datasets demonstrate that: (i) the contrastive pre-training increases the robustness of any loss function to noisy labels and (ii) the additional fine-tuning phase can further improve accuracy, but at the cost of additional complexity.