Genomewide association studies (GWAS) have identified a large number of loci associated with neuropsychiatric traits, however, understanding the molecular mechanisms underlying these loci remains difficult. To help prioritize causal variants and interpret their functions, computational methods have been developed to predict regulatory effects of non-coding variants. An emerging approach to variant annotation is deep learning models that predict regulatory functions from DNA sequences alone. While such models have been trained on large publicly available dataset such as ENCODE, neuropsychiatric trait-related cell types are under-represented in these datasets, thus there is an urgent need of better tools and resources to annotate variant functions in such cellular contexts. To fill this gap, we collected a large collection of neurodevelopment-related cell/tissue types, and trained deep Convolutional Neural Networks (ResNet) using such data. Furthermore, our model, called MetaChrom, borrows information from public epigenomic consortium to improve the accuracy via transfer learning. We show that MetaChrom is substantially better in predicting experimentally determined chromatin accessibility variants than popular variant annotation tools such as CADD and delta-SVM. By combining GWAS data with MetaChrom predictions, we prioritized 31 SNPs for Schizophrenia, suggesting potential risk genes and the biological contexts where they act. In summary, MetaChrom provides functional annotations of any DNA variants in the neuro-development context and the general method of MetaChrom can also be extended to other disease-related cell or tissue types.

Generative Adversarial Networks (GANs) have shown impressive gains in image synthesis. GAN inversion was recently studied to understand and utilize the knowledge it learns, where a real image is inverted back to a latent code and can thus be reconstructed by the generator. Although increasing the number of latent codes can improve inversion quality to a certain extent, we find that important details may still be neglected when performing feature composition over all the intermediate feature channels. To address this issue, we propose a Prior multi-Subspace Feature Composition (PmSFC) approach for high-fidelity inversion. Considering that the intermediate features are highly correlated with each other, we incorporate a self-expressive layer in the generator to discover meaningful subspaces. In this case, the features at a channel can be expressed as a linear combination of those at other channels in the same subspace. We perform feature composition separately in the subspaces. The semantic differences between them benefit the inversion quality, since the inversion process is regularized based on different aspects of semantics. In the experiments, the superior performance of PmSFC demonstrates the effectiveness of prior subspaces in facilitating GAN inversion together with extended applications in visual manipulation.

Data interpolation is typically used to explore and understand the latent representation learnt by a deep network. Naive linear interpolation may induce mismatch between the interpolated data and the underlying manifold of the original data. In this paper, we propose an Adversarial Adaptive Interpolation (AdvAI) approach for facilitating representation learning and image synthesis in autoencoders. To determine an interpolation path that stays on the manifold, we incorporate an interpolation correction module, which learns to offset the deviation from the manifold. Further, we perform matching with a prior distribution to control the characteristics of the representation. The data synthesized from random codes along with interpolation-based regularization are in turn used to constrain the representation learning process. In the experiments, the superior performance of the proposed approach demonstrates the effectiveness of AdvAI and associated regularizers in a variety of downstream tasks.

Autoencoders enjoy a remarkable ability to learn data representations. Research on autoencoders shows that the effectiveness of data interpolation can reflect the performance of representation learning. However, existing interpolation methods in autoencoders do not have enough capability of traversing a possible region between two datapoints on a data manifold, and the distribution of interpolated latent representations is not considered.To address these issues, we aim to fully exert the potential of data interpolation and further improve representation learning in autoencoders. Specifically, we propose the multidimensional interpolation to increase the capability of data interpolation by randomly setting interpolation coefficients for each dimension of latent representations. In addition, we regularize autoencoders in both the latent and the data spaces by imposing a prior on latent representations in the Maximum Mean Discrepancy (MMD) framework and encouraging generated datapoints to be realistic in the Generative Adversarial Network (GAN) framework. Compared to representative models, our proposed model has empirically shown that representation learning exhibits better performance on downstream tasks on multiple benchmarks.

In light of increased co-prescription of multiple drugs, the ability to discern and predict drugdrug interactions (DDI) has become crucial to guarantee the safety of patients undergoing treatment with multiple drugs.However, information on DDI profiles is incomplete and the experimental determination of DDIs is labor-intensive and time-consuming.Although previous studies have explored various feature spaces for in silico screening of interacting drug pairs, their use of conventional cross-validation prevents them from achieving generalizable performance on drug pairs where neither drug is seen during training.Here we demonstrate for the first time targets of adversely interacting drug pairs are significantly more likely to have synergistic genetic interactions than non-interacting drug pairs.Leveraging genetic interaction features and a novel training scheme, we construct a gradient boosting-based classifier that achieves robust DDI prediction even for drugs whose interaction profiles are completely unseen during training.We demonstrate that in addition to classification power -including the prediction of 432 novel DDIs-our genetic interaction approach offers interpretability by providing plausible mechanistic insights into the mode of action of DDIs. Author summaryAdverse drug-drug interactions are adverse side effects caused by taking two or more drugs together.As co-prescription of multiple drugs becomes an increasingly prevalent practice, affecting 42.2% of Americans over 65 years old, adverse drug-drug interactions have become a serious safety concern, accounting for over 74,000 emergency room visits and 195,000 hospitalizations each year in the United States alone.Since experimental determination of adverse drug-drug interactions is labor-intensive and time-consuming, various machine learning-based computational approaches have been developed for predicting drug-drug interactions.Considering the fact that drugs effect through binding and modulating the function of their targets, we have explored whether drug-drug interactions can be predicted from the genetic interaction between the gene targets of two drugs, which characterizes the unexpected fitness effect when two genes are simultaneously knocked out.Furthermore, we have built a fast and robust classifier that achieves accurate

Existing domain adaptation approaches generalize models trained on the labeled source domain data to the unlabeled target domain data by forcing feature distributions of two domains closer. However, these approaches are likely to ignore the semantic information during the feature alignment between source and target domain. In this paper, we propose a new unsupervised domain adaptation framework to learn the cross-domain features and disentangle the semantic information concurrently. Specifically, we firstly combine the task-specific classification and domain adversarial learning to obtain the cross-domain features by mapping the data of both domains with the shared feature extractor. Secondly, we integrate the domain adversarial learning and the within-domain reconstruction to disentangle the semantic information from the domain information. Thirdly, we include a cross-domain transformation to further refine the feature extractor, which in turn improves the performances of the task classifier. We compare our proposed model to previous state-of-the-art methods on domain adaptation digit classification tasks. Experimental results show that our model achieves better performances than the other counterparts, which demonstrates the superiority and effectiveness of our model.

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech–two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, enabling experiments that previously took weeks to now run in days. This allows us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

We propose a convolutional neural network architecture with k-max pooling layer for semantic modeling of music. The aim of a music model is to analyze and represent the semantic content of music for purposes of classification, discovery, or clustering. The k-max pooling layer is used in the network to make it possible to pool the k most active features, capturing the semantic-rich and time-varying information about music. Our network takes an input music as a sequence of audio words, where each audio word is associated with a distributed feature vector that can be fine-tuned by backpropagating errors during the training. The architecture allows us to take advantage of the better trained audio word embeddings and the deep structures to produce more robust music representations. Experiment results with two different music collections show that our neural networks achieved the best accuracy in music genre classification comparing with three state-of-art systems.