Regularity Normalization: Neuroscience-Inspired Unsupervised Attention across Neural Network Layers.

Inspired by the adaptation phenomenon of neuronal firing, we propose the regularity normalization (RN) as an unsupervised attention mechanism (UAM) which computes the statistical regularity in the implicit space of neural networks under the Minimum Description Length (MDL) principle. Treating the neural network optimization process as a partially observable model selection problem, the regularity normalization constrains the implicit space by a normalization factor, the universal code length. We compute this universal code incrementally across neural network layers and demonstrate the flexibility to include data priors such as top-down attention and other oracle information. Empirically, our approach outperforms existing normalization methods in tackling limited, imbalanced and non-stationary input distribution in image classification, classic control, procedurally-generated reinforcement learning, generative modeling, handwriting generation and question answering tasks with various neural network architectures. Lastly, the unsupervised attention mechanisms is a useful probing tool for neural networks by tracking the dependency and critical learning stages across layers and recurrent time steps of deep networks.

Self-Assembling 2D Arrays with de Novo Protein Building Blocks.

Modular self-assembly of biomolecules in two dimensions (2D) is straightforward with DNA but has been difficult to realize with proteins, due to the lack of modular specificity similar to Watson-Crick base pairing. Here we describe a general approach to design 2D arrays using de novo designed pseudosymmetric protein building blocks. A homodimeric helical bundle was reconnected into a monomeric building block, and the surface was redesigned in Rosetta to enable self-assembly into a 2D array in the C12 layer symmetry group. Two out of ten designed arrays assembled to micrometer scale under negative stain electron microscopy, and displayed the designed lattice geometry with assembly size up to 100 nm under atomic force microscopy. The design of 2D arrays with pseudosymmetric building blocks is an important step toward the design of programmable protein self-assembly via pseudosymmetric patterning of orthogonal binding interfaces.

Predicting human decision making in psychological tasks with recurrent neural networks.

Unlike traditional time series, the action sequences of human decision making usually involve many cognitive processes such as beliefs, desires, intentions, and theory of mind, i.e., what others are thinking. This makes predicting human decision-making challenging to be treated agnostically to the underlying psychological mechanisms. We propose here to use a recurrent neural network architecture based on long short-term memory networks (LSTM) to predict the time series of the actions taken by human subjects engaged in gaming activity, the first application of such methods in this research domain. In this study, we collate the human data from 8 published literature of the Iterated Prisoner's Dilemma comprising 168,386 individual decisions and post-process them into 8,257 behavioral trajectories of 9 actions each for both players. Similarly, we collate 617 trajectories of 95 actions from 10 different published studies of Iowa Gambling Task experiments with healthy human subjects. We train our prediction networks on the behavioral data and demonstrate a clear advantage over the state-of-the-art methods in predicting human decision-making trajectories in both the single-agent scenario of the Iowa Gambling Task and the multi-agent scenario of the Iterated Prisoner's Dilemma. Moreover, we observe that the weights of the LSTM networks modeling the top performers tend to have a wider distribution compared to poor performers, as well as a larger bias, which suggest possible interpretations for the distribution of strategies adopted by each group.

Influence of ocular dominance columns and patchy callosal connections on binocularity in lateral striate cortex: Long Evans versus albino rats.

In albino rats, it has been reported that lateral striate cortex (V1) is highly binocular, and that input from the ipsilateral eye to this region comes through the callosum. In contrast, in Long Evans rats, this region is nearly exclusively dominated by the contralateral eye even though it is richly innervated by the callosum (Laing, Turecek, Takahata, & Olavarria, 2015). We hypothesized that the inability of callosal connections to relay ipsilateral eye input to lateral V1 in Long Evans rats is a consequence of the existence of ocular dominance columns (ODCs), and of callosal patches in register with ipsilateral ODCs in the binocular region of V1 (Laing et al., 2015). We therefore predicted that in albino rats input from both eyes intermix in the binocular region, without segregating into ODCs, and that callosal connections are not patchy. Confirming our predictions, we found that inputs from both eyes, studied with the transneuronal tracer WGA-HRP, are intermixed in the binocular zone of albinos, without segregating into ODCs. Similarly, we found that callosal connections in albino rats are not patchy but instead are distributed homogeneously throughout the callosal region in V1. We propose that these changes allow the transcallosal passage of ipsilateral eye input to lateral striate cortex, increasing its binocularity. Thus, the binocular region in V1 of albino rats includes lateral striate cortex, being therefore about 25% larger in area than the binocular region in Long Evans rats. Our findings provide insight on the role of callosal connections in generating binocular cells.

Dynamic Forecasting of Zika Epidemics Using Google Trends.

We developed a dynamic forecasting model for Zika virus (ZIKV), based on real-time online search data from Google Trends (GTs). It was designed to provide Zika virus disease (ZVD) surveillance and detection for Health Departments, and predictive numbers of infection cases, which would allow them sufficient time to implement interventions. In this study, we found a strong correlation between Zika-related GTs and the cumulative numbers of reported cases (confirmed, suspected and total cases; p<0.001). Then, we used the correlation data from Zika-related online search in GTs and ZIKV epidemics between 12 February and 20 October 2016 to construct an autoregressive integrated moving average (ARIMA) model (0, 1, 3) for the dynamic estimation of ZIKV outbreaks. The forecasting results indicated that the predicted data by ARIMA model, which used the online search data as the external regressor to enhance the forecasting model and assist the historical epidemic data in improving the quality of the predictions, are quite similar to the actual data during ZIKV epidemic early November 2016. Integer-valued autoregression provides a useful base predictive model for ZVD cases. This is enhanced by the incorporation of GTs data, confirming the prognostic utility of search query based surveillance. This accessible and flexible dynamic forecast model could be used in the monitoring of ZVD to provide advanced warning of future ZIKV outbreaks.

Model-informed risk assessment for Zika virus outbreaks in the Asia-Pacific regions.

Recently, Zika virus (ZIKV) has been recognized as a significant threat to global public health. The disease was present in large parts of the Americas, the Caribbean, and also the western Pacific area with southern Asia during 2015 and 2016. However, little is known about the factors affecting the transmission of ZIKV. We used Gradient Boosted Regression Tree models to investigate the effects of various potential explanatory variables on the spread of ZIKV, and used current with historical information from a range of sources to assess the risks of future ZIKV outbreaks. Our results indicated that the probability of ZIKV outbreaks increases with vapor pressure, the occurrence of Dengue virus, and population density but decreases as health expenditure, GDP, and numbers of travelers. The predictive results revealed the potential risk countries of ZIKV infection in the Asia-Pacific regions between October 2016 and January 2017. We believe that the high-risk conditions would continue in South Asia and Australia over this period. By integrating information on eco-environmental, social-economical, and ZIKV-related niche factors, this study estimated the probability for locally acquired mosquito-borne ZIKV infections in the Asia-Pacific region and improves the ability to forecast, and possibly even prevent, future outbreaks of ZIKV.

An Integrative Analysis Reveals a Central Role of P53 Activation via MDM2 in Zika Virus Infection Induced Cell Death.

Zika virus (ZIKV) infection is an emerging global threat that is suspected to be associated with fetal microcephaly. However, the molecular mechanisms underlying ZIKV disease pathogenesis in humans remain elusive. Here, we investigated the human protein interaction network associated with ZIKV infection using a systemic virology approach, and reconstructed the transcriptional regulatory network to analyze the mechanisms underlying ZIKV-elicited microcephaly pathogenesis. The bioinformatics findings in this study show that P53 is the hub of the genetic regulatory network for ZIKV-related and microcephaly-associated proteins. Importantly, these results imply that the ZIKV capsid protein interacts with mouse double-minute-2 homolog (MDM2), which is involved in the P53-mediated apoptosis pathway, activating the death of infected neural cells. We also found that synthetic mimics of the ZIKV capsid protein induced cell death in vitro and in vivo. This study provides important insight into the relationship between ZIKV infection and brain diseases.

Systematic Genome-wide Screening and Prediction of microRNAs in EBOV During the 2014 Ebolavirus Outbreak.

Recently, several thousand people have been killed by the Ebolavirus disease (EVD) in West Africa, yet no current antiviral medications and treatments are available. Systematic investigation of ebolavirus whole genomes during the 2014 outbreak may shed light on the underlying mechanisms of EVD development. Here, using the genome-wide screening in ebolavirus genome sequences, we predicted four putative viral microRNA precursors (pre-miRNAs) and seven putative mature microRNAs (miRNAs). Combing bioinformatics analysis and prediction of the potential ebolavirus miRNA target genes, we suggest that two ebolavirus coding possible miRNAs may be silence and down-regulate the target genes NFKBIE and RIPK1, which are the central mediator of the pathways related with host cell defense mechanism. Additionally, the ebolavirus exploits the miRNAs to inhibit the NF-kB and TNF factors to evade the host defense mechanisms that limit replication by killing infected cells, or to conversely trigger apoptosis as a mechanism to increase virus spreading. This is the first study to use the genome-wide scanning to predict microRNAs in the 2014 outbreak EVD and then to apply systematic bioinformatics to analyze their target genes. We revealed a potential mechanism of miRNAs in ebolavirus infection and possible therapeutic targets for Ebola viral infection treatment.