CellVGAE: an unsupervised scRNA-seq analysis workflow with graph attention networks.

MOTIVATION: Single-cell RNA sequencing allows high-resolution views of individual cells for libraries of up to millions of samples, thus motivating the use of deep learning for analysis. In this study, we introduce the use of graph neural networks for the unsupervised exploration of scRNA-seq data by developing a variational graph autoencoder architecture with graph attention layers that operates directly on the connectivity between cells, focusing on dimensionality reduction and clustering. With the help of several case studies, we show that our model, named CellVGAE, can be effectively used for exploratory analysis even on challenging datasets, by extracting meaningful features from the data and providing the means to visualize and interpret different aspects of the model. RESULTS: We show that CellVGAE is more interpretable than existing scRNA-seq variational architectures by analysing the graph attention coefficients. By drawing parallels with other scRNA-seq studies on interpretability, we assess the validity of the relationships modelled by attention, and furthermore, we show that CellVGAE can intrinsically capture information such as pseudotime and NF-ĸB activation dynamics, the latter being a property that is not generally shared by existing neural alternatives. We then evaluate the dimensionality reduction and clustering performance on 9 difficult and well-annotated datasets by comparing with three leading neural and non-neural techniques, concluding that CellVGAE outperforms competing methods. Finally, we report a decrease in training times of up to × 20 on a dataset of 1.3 million cells compared to existing deep learning architectures. AVAILABILITYAND IMPLEMENTATION: The CellVGAE code is available at https://github.com/davidbuterez/CellVGAE. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MF-PCBA: Multifidelity High-Throughput Screening Benchmarks for Drug Discovery and Machine Learning.

High-throughput screening (HTS), as one of the key techniques in drug discovery, is frequently used to identify promising drug candidates in a largely automated and cost-effective way. One of the necessary conditions for successful HTS campaigns is a large and diverse compound library, enabling hundreds of thousands of activity measurements per project. Such collections of data hold great promise for computational and experimental drug discovery efforts, especially when leveraged in combination with modern deep learning techniques, and can potentially lead to improved drug activity predictions and cheaper and more effective experimental design. However, existing collections of machine-learning-ready public datasets do not exploit the multiple data modalities present in real-world HTS projects. Thus, the largest fraction of experimental measurements, corresponding to hundreds of thousands of "noisy" activity values from primary screening, are effectively ignored in the majority of machine learning models of HTS data. To address these limitations, we introduce Multifidelity PubChem BioAssay (MF-PCBA), a curated collection of 60 datasets that includes two data modalities for each dataset, corresponding to primary and confirmatory screening, an aspect that we call multifidelity. Multifidelity data accurately reflect real-world HTS conventions and present a new, challenging task for machine learning: the integration of low- and high-fidelity measurements through molecular representation learning, taking into account the orders-of-magnitude difference in size between the primary and confirmatory screens. Here we detail the steps taken to assemble MF-PCBA in terms of data acquisition from PubChem and the filtering steps required to curate the raw data. We also provide an evaluation of a recent deep-learning-based method for multifidelity integration across the introduced datasets, demonstrating the benefit of leveraging all HTS modalities, and a discussion in terms of the roughness of the molecular activity landscape. In total, MF-PCBA contains over 16.6 million unique molecule-protein interactions. The datasets can be easily assembled by using the source code available at https://github.com/davidbuterez/mf-pcba.

Transfer learning with graph neural networks for improved molecular property prediction in the multi-fidelity setting.

We investigate the potential of graph neural networks for transfer learning and improving molecular property prediction on sparse and expensive to acquire high-fidelity data by leveraging low-fidelity measurements as an inexpensive proxy for a targeted property of interest. This problem arises in discovery processes that rely on screening funnels for trading off the overall costs against throughput and accuracy. Typically, individual stages in these processes are loosely connected and each one generates data at different scale and fidelity. We consider this setup holistically and demonstrate empirically that existing transfer learning techniques for graph neural networks are generally unable to harness the information from multi-fidelity cascades. Here, we propose several effective transfer learning strategies and study them in transductive and inductive settings. Our analysis involves a collection of more than 28 million unique experimental protein-ligand interactions across 37 targets from drug discovery by high-throughput screening and 12 quantum properties from the dataset QMugs. The results indicate that transfer learning can improve the performance on sparse tasks by up to eight times while using an order of magnitude less high-fidelity training data. Moreover, the proposed methods consistently outperform existing transfer learning strategies for graph-structured data on drug discovery and quantum mechanics datasets.

Modelling local and general quantum mechanical properties with attention-based pooling.

Atom-centred neural networks represent the state-of-the-art for approximating the quantum chemical properties of molecules, such as internal energies. While the design of machine learning architectures that respect chemical principles has continued to advance, the final atom pooling operation that is necessary to convert from atomic to molecular representations in most models remains relatively undeveloped. The most common choices, sum and average pooling, compute molecular representations that are naturally a good fit for many physical properties, while satisfying properties such as permutation invariance which are desirable from a geometric deep learning perspective. However, there are growing concerns that such simplistic functions might have limited representational power, while also being suboptimal for physical properties that are highly localised or intensive. Based on recent advances in graph representation learning, we investigate the use of a learnable pooling function that leverages an attention mechanism to model interactions between atom representations. The proposed pooling operation is a drop-in replacement requiring no changes to any of the other architectural components. Using SchNet and DimeNet++ as starting models, we demonstrate consistent uplifts in performance compared to sum and mean pooling and a recent physics-aware pooling operation designed specifically for orbital energies, on several datasets, properties, and levels of theory, with up to 85% improvements depending on the specific task.

Scaling up DNA digital data storage by efficiently predicting DNA hybridisation using deep learning.

Deoxyribonucleic acid (DNA) has shown great promise in enabling computational applications, most notably in the fields of DNA digital data storage and DNA computing. Information is encoded as DNA strands, which will naturally bind in solution, thus enabling search and pattern-matching capabilities. Being able to control and predict the process of DNA hybridisation is crucial for the ambitious future of Hybrid Molecular-Electronic Computing. Current tools are, however, limited in terms of throughput and applicability to large-scale problems. We present the first comprehensive study of machine learning methods applied to the task of predicting DNA hybridisation. For this purpose, we introduce an in silico-generated hybridisation dataset of over 2.5 million data points, enabling the use of deep learning. Depending on hardware, we achieve a reduction in inference time ranging from one to over two orders of magnitude compared to the state-of-the-art, while retaining high fidelity. We then discuss the integration of our methods in modern, scalable workflows.