Search

Motivation Molecular carcinogenicity is a preventable cause of cancer, but systematically identifying carcinogenic compounds, which involves performing experiments on animal models, is expensive, time consuming and low throughput. As a result, carcinogenicity information is limited and building data-driven models with good prediction accuracy remains a major challenge. Results In this work, we propose CONCERTO, a deep learning model that uses a graph transformer in conjunction with a molecular fingerprint representation for carcinogenicity prediction from molecular structure. Special efforts have been made to overcome the data size constraint, such as multi-round pre-training on related but lower quality mutagenicity data, and transfer learning from a large self-supervised model. Extensive experiments demonstrate that our model performs well and can generalize to external validation sets. CONCERTO could be useful for guiding future carcinogenicity experiments and provide insight into the molecular basis of carcinogenicity. Availability and implementation The code and data underlying this article are available on github at https://github.com/bowang-lab/CONCERTO

Molecular carcinogenicity is a preventable cause of cancer, however, most experimental testing of molecular compounds is an expensive and time consuming process, making high throughput experimental approaches infeasible. In recent years, there has been substantial progress in machine learning techniques for molecular property prediction. In this work, we propose a model for carcinogenicity prediction, CONCERTO, which uses a graph transformer in conjunction with a molecular fingerprint representation, trained on multi-round muta-genicity and carcinogenicity objectives. To train and validate CONCERTO, we augment the training dataset with more informative labels and utilize a larger external validation dataset. Extensive experiments demonstrate that our model yields results superior to alternate approaches for molecular carcinogenicity prediction.

Remote monitoring of electrical cable conditions is an essential characteristic of the next-generation smart grid, which features the ability to consistently surveil and control the grid infrastructure. In this paper, we propose a technique that harnesses power line modems (PLMs) as sensors for monitoring cable health. We envisage that all or most of these PLMs have already been deployed for data communication purposes and focus on the distribution grid or neighborhood area networks in the smart grid. For such a setting, we propose a machine learning (ML) based framework for automatic cable diagnostics by continuously monitoring the cable status to identify, assess, and locate possible degradations. As part of our technique, we also synthesize state-of-the-art reflectometry methods within the PLMs to extract beneficial features for effective performance of our proposed ML solution. Simulation results demonstrate the effectiveness of our solution under different aging conditions and varying load configurations. Finally, we reflect on our proposed diagnostics method by evaluating its robustness and comparing it with existing alternatives., Submitted to an IEEE journal. A version of this paper won the "Best Paper Award" at IEEE ISPLC 2018

The design and performance evaluation of underwater acoustic (UA) communication systems in shallow water and harbour environments is a continuous challenge due to the numerous degrading factors present in the UA channel, one of which is the presence of noise generated due to nearby shipping activity. However, few research studies have examined the properties of ship-radiated noise in terms of its time-domain statistical characteristics and its negative effects on UA communication systems. We propose the use of unsupervised learning techniques to train generative models that capture the time-domain stochastic behaviours of ship-radiated noise using a publicly available database of long-term acoustic shipping noise recordings. These models can then be used for further analysis of ship-radiated noise and performance evaluation of UA orthogonal frequency-division multiplexing systems in the presence of such interference. For further validation, we include experimentally acquired ship-radiated noise recordings acquired off the coast of Caesarea, Israel. The results indicate a two component Gaussian mixture model serves as a better approximation for high frequency ship-radiated noise while generative adversarial networks produce improved realizations of shipping noise in lower frequencies.

Interference due to ship-radiated noise in the underwater acoustic (UA) channel generates additive distortions that degrade wireless UA communications signals. Compressed sensing (CS) techniques are an approach used to estimate and suppress the impulsive components of ship-radiated noise for orthogonal frequency-division multiplexing (OFDM) systems by exploiting the null sub-carriers not used for data transmission. However, these CS-based estimation methods are constrained to estimating sparse signals and typically require slow iterative solvers. To combat these drawbacks, we propose a deep learning (DL) approach to structured signal recovery for estimating and mitigating the interfering effects of ship-radiated noise for OFDM systems. Our results indicate that the DL models, trained via publicly available long term acoustic data of shipping noise signals, produce measurable mitigation gains to the benchmark CS algorithms. In addition, we show the DL models outperform the benchmark CS estimation methods on new never before “seen” experimentally acquired ship-radiated noise data.

Achieving high data rate robust communication in shallow and harbour underwater acoustic (UA) environments can be a demanding challenge in the presence of shipping noise. Noise generated from nearby passing ships can lead to impulsive agitations which impair UA communication systems. Utilizing the assumption that impulse noise exhibits sparsity, we realize a compressed sensing (CS) based framework for noise estimation exploiting the pilot sub-carriers of UA orthogonal frequency-division modulation systems. Under the CS framework, we propose the use of a empirical Bayesian approach which first characterizes the statistical properties of shipping noise prior to conceiving an estimate. In addition, we invoke the K-SVD algorithm for dictionary learning. K-SVD iteratively forms a sparse representation for the class of shipping noise signals, which is later used for noise estimation. Numerical results show that the empirical Bayesian based signal recovery algorithm yields the best performance for interference estimation.

With an aging power distribution infrastructure, it becomes increasingly important for the next generation smart grid to self-monitor its transmission lines. In this paper, we propose a road-map towards achieving a self-reliant grid surveillance using power line communications (PLC). To this end, we exploit the principle that cable faults or degradations manifest themselves as changes in the PLC channel conditions. In particular, by monitoring the channel transfer functions that are already computed in legacy PLC receivers, we enable power line modems with intelligent grid sensing abilities to identify and assess cable anomalies using machine learning techniques. Through simulations, we show that our proposed monitoring and diagnostics solutions successfully empower power line modems to independently detect and predict the extent of water-tree degradations commonly seen in cross-linked polyethylene insulated power cables.