The emerging potential of interactive virtual reality in drug discovery.

INTRODUCTION: The potential of virtual reality (VR) to contribute to drug design and development has been recognized for many years. A recent advance is to use VR not only to visualize and interact with molecules, but also to interact with molecular dynamics simulations 'on the fly' (interactive molecular dynamics in VR, IMD-VR), which is useful for flexible docking and examining binding processes and conformational changes. AREAS COVERED: The authors use the term 'interactive VR' to refer to software where interactivity is an inherent part of the user VR experience e.g. in making structural modifications or interacting with a physically rigorous molecular dynamics (MD) simulation, as opposed to using VR controllers to rotate and translate the molecule for enhanced visualization. Here, they describe these methods and their application to problems relevant to drug discovery, highlighting the possibilities that they offer in this arena. EXPERT OPINION: The ease of viewing and manipulating molecular structures and dynamics, using accessible VR hardware, and the ability to modify structures on the fly (e.g. adding or deleting atoms) - and for groups of researchers to work together in the same virtual environment - makes modern interactive VR a valuable tool to add to the armory of drug design and development methods.

Are there any 'object detectors' in the hidden layers of CNNs trained to identify objects or scenes?

Various methods of measuring unit selectivity have been developed with the aim of better understanding how neural networks work. But the different measures provide divergent estimates of selectivity, and this has led to different conclusions regarding the conditions in which selective object representations are learned and the functional relevance of these representations. In an attempt to better characterize object selectivity, we undertake a comparison of various selectivity measures on a large set of units in AlexNet, including localist selectivity, precision, class-conditional mean activity selectivity (CCMAS), the human interpretation of activation maximization (AM) images, and standard signal-detection measures. We find that the different measures provide different estimates of object selectivity, with precision and CCMAS measures providing misleadingly high estimates. Indeed, the most selective units had a poor hit-rate or a high false-alarm rate (or both) in object classification, making them poor object detectors. We fail to find any units that are even remotely as selective as the 'grandmother cell' units reported in recurrent neural networks. In order to generalize these results, we compared selectivity measures on units in VGG-16 and GoogLeNet trained on the ImageNet or Places-365 datasets that have been described as 'object detectors'. Again, we find poor hit-rates and high false-alarm rates for object classification. We conclude that signal-detection measures provide a better assessment of single-unit selectivity compared to common alternative approaches, and that deep convolutional networks of image classification do not learn object detectors in their hidden layers.

Researchers Keep Rejecting Grandmother Cells after Running the Wrong Experiments: The Issue Is How Familiar Stimuli Are Identified.

There is widespread agreement in neuroscience and psychology that the visual system identifies objects and faces based on a pattern of activation over many neurons, each neuron being involved in representing many different categories. The hypothesis that the visual system includes finely tuned neurons for specific objects or faces for the sake of identification, so-called "grandmother cells", is widely rejected. Here it is argued that the rejection of grandmother cells is premature. Grandmother cells constitute a hypothesis of how familiar visual categories are identified, but the primary evidence against this hypothesis comes from studies that have failed to observe neurons that selectively respond to unfamiliar stimuli. These findings are reviewed and it is shown that they are irrelevant. Neuroscientists need to better understand existing models of face and object identification that include grandmother cells and then compare the selectivity of these units with single neurons responding to stimuli that can be identified.

Neuromorphic computation with spiking memristors: habituation, experimental instantiation of logic gates and a novel sequence-sensitive perceptron model.

Memristors have been compared to neurons and synapses, suggesting they would be good for neuromorphic computing. A change in voltage across a memristor causes a current spike which imparts a short-term memory to a memristor, allowing for through-time computation, which can do arithmetical operations and sequential logic, or model short-time habituation to a stimulus. Using simple physical rules, simple logic gates such as XOR, and novel, more complex, gates such as the arithmetic full adder (AFA) can be instantiated in sol-gel TiO2 plastic memristors. The adder makes use of the memristor's short-term memory to add together three binary values and outputs the sum, the carry digit and even the order they were input in, allowing for logically (but not physically reversible) computation. Only a single memristor is required to instantiate each gate, as additional input/output ports can be replaced with extra time-steps allowing a single memristor to do a hitherto unexpectedly large amount of computation, which may mitigate the memristor's slow operation speed and may relate to how neurons do a similarly large computation with slow operation speeds. These logic gates can be understood by modelling the memristors as a novel type of perceptron: one which is sensitive to input order. The memristor's short-term memory can change the input weights applied to later inputs, and thus the memristor gates cannot be accurately described by a single perceptron, requiring either a network of time-invariant perceptrons, or a sequence-sensitive self-reprogrammable perceptron. Thus, the AFA is best described as a sequence-sensitive perceptron that sorts binary inputs into classes corresponding to the arithmetical sum of the inputs. Co-development of memristor hardware alongside software (sequence-sensitive perceptron) models in trained neural networks would allow the porting of modern deep-neural networks architecture to low-power hardware neural net chips.

Combining random walk and regression models to understand solvation in multi-component solvent systems.

Polysaccharides, such as cellulose, are often processed by dissolution in solvent mixtures, e.g. an ionic liquid (IL) combined with a dipolar aprotic co-solvent (CS) that the polymer does not dissolve in. A multi-walker, discrete-time, discrete-space 1-dimensional random walk can be applied to model solvation of a polymer in a multi-component solvent mixture. The number of IL pairs in a solvent mixture and the number of solvent shells formable, x, is associated with n, the model time-step, and N, the number of random walkers. The mean number of distinct sites visited is proportional to the amount of polymer soluble in a solution. By also fitting a polynomial regression model to the data, we can associate the random walk terms with chemical interactions between components and probe where the system deviates from a 1-D random walk. The 'frustration' between solvents shells is given as ln x in the random walk model and as a negative IL:IL interaction term in the regression model. This frustration appears in regime II of the random walk model (high volume fractions of IL) where walkers interfere with each other, and the system tends to its limiting behaviour. In the low concentration regime, (regime I) the solvent shells do not interact, and the system depends only on IL and CS terms. In both models (and both regimes), the system is almost entirely controlled by the volume available to solvation shells, and thus is a counting/space-filling problem, where the molar volume of the CS is important. Small deviations are observed when there is an IL-CS interaction. The use of two models, built on separate approaches, confirm these findings, demonstrating that this is a real effect and offering a route to identifying such systems. Specifically, the majority of CSs - such as dimethylformide - follow the random walk model, whilst 1-methylimidazole, dimethyl sulfoxide, 1,3-dimethyl-2-imidazolidinone and tetramethylurea offer a CS-mediated improvement and propylene carbonate results in a CS-mediated hindrance. It is shown here that systems, which are very complex at a molecular level, may, nonetheless, be effectively modelled as a simple random walk in phase-space. The 1-D random walk model allows prediction of the ability of solvent mixtures to dissolve cellulose based on only two dissolution measurements (one in neat IL) and molar volume.