Discrepancy-Guided Domain-Adaptive Data Augmentation.

Data augmentation has been observed playing a crucial role in achieving better generalization in many machine learning tasks, especially in unsupervised domain adaptation (DA). It is particularly effective on visual object recognition tasks as images are high-dimensional with an enormous range of variations that can be simulated. Existing data augmentation techniques, however, are not explicitly designed to address the differences between different domains. Expert knowledge about the data is required, as well as manual efforts in finding the optimal parameters. In this article, we propose a novel domain-adaptive augmentation method by making use of a state-of-the-art style transfer method and domain discrepancy measurement. Specifically, we measure the discrepancy between source and target domains, and use it as a guide to augment the original source samples using style transferred source-to-target samples. The proposed domain-adaptive augmentation method is data and model agnostic that can be easily incorporated with state-of-the-art DA algorithms. We show empirically that, by using this domain-adaptive augmentation, we are able to gradually reduce the discrepancy between the source and target samples, and further boost the adaptation performance using different DA algorithms on three popular domain adaption datasets.

Ranked List Loss for Deep Metric Learning.

The objective of deep metric learning (DML) is to learn embeddings that can capture semantic similarity and dissimilarity information among data points. Existing pairwise or tripletwise loss functions used in DML are known to suffer from slow convergence due to a large proportion of trivial pairs or triplets as the model improves. To improve this, ranking-motivated structured losses are proposed recently to incorporate multiple examples and exploit the structured information among them. They converge faster and achieve state-of-the-art performance. In this work, we unveil two limitations of existing ranking-motivated structured losses and propose a novel ranked list loss to solve both of them. First, given a query, only a fraction of data points is incorporated to build the similarity structure. Consequently, some useful examples are ignored and the structure is less informative. To address this, we propose to build a set-based similarity structure by exploiting all instances in the gallery. The learning setting can be interpreted as few-shot retrieval: given a mini-batch, every example is iteratively used as a query, and the rest ones compose the gallery to search, i.e., the support set in few-shot setting. The rest examples are split into a positive set and a negative set. For every mini-batch, the learning objective of ranked list loss is to make the query closer to the positive set than to the negative set by a margin. Second, previous methods aim to pull positive pairs as close as possible in the embedding space. As a result, the intraclass data distribution tends to be extremely compressed. In contrast, we propose to learn a hypersphere for each class in order to preserve useful similarity structure inside it, which functions as regularisation. Extensive experiments demonstrate the superiority of our proposal by comparing with the state-of-the-art methods on the fine-grained image retrieval task. Our source code is available online: https://github.com/XinshaoAmosWang/Ranked-List-Loss-for-DML.

Cross-View Discriminative Feature Learning for Person Re-Identification.

The viewpoint variability across a network of non-overlapping cameras is a challenging problem affecting person re-identification performance. In this paper, we investigate how to mitigate the cross-view ambiguity by learning highly discriminative deep features under the supervision of a novel loss function. The proposed objective is made up of two terms, the steering meta center term and the enhancing centers dispersion term, that steer the training process to mining effective intra-class and inter-class relationships in the feature domain of the identities. The effect of our loss supervision is to generate a more expanded feature space of compact classes where the overall level of the inter-identities' interference is reduced. Compared with the existing metric learning techniques, this approach has the advantage of achieving a better optimization because it jointly learns the embedding and the metric contextually. Our technique, by dismissing side-sources of performance gain, proves to enhance the CNN invariance to viewpoint without incurring increased training complexity (like in Siamese or triplet networks) and outperforms many related state-of-the-art techniques on Market-1501 and CUHK03.

Controlled in-cell activation of RNA therapeutics using bond-cleaving bio-orthogonal chemistry.

Temporal control of siRNA activation is a major challenge for RNAi-based therapeutics. The majority of the reported siRNA delivery systems rely on environmental factors, such as differences in extracellular and intracellular redox potential, ATP concentration, or pH to activate an siRNA payload. However dynamic endogenous environments are far too complex to rely on for controllable siRNA release and can result in premature siRNA activation prior to reaching the intended biological target. In addition, there are uncertainties about timing, degree and rate of the siRNA activation with spontaneous release approaches. Herein we describe a bio-orthogonal chemistry approach to address this important challenge. With our approach we were able achieve two major goals: complete siRNA inactivation upon immobilization of the payload on the surface of iron oxide nanoparticles and controlled in-cell activation with the addition of a small non-toxic chemical trigger after sufficient cellular uptake of the nanoparticles was confirmed. We have demonstrated our in-cell activation approach using two siRNAs against green fluorescent protein (GFP) and cyclin dependent kinase 8 (CDK8) in GFP expressing MDA-MB-231 cell line. We anticipate that this methodology will potentially advance the clinical translation of RNAi-based therapeutics, as the described bio-orthogonal chemistry can be generalized for any siRNA of choice.

Universal sensor array for highly selective system identification using two-dimensional nanoparticles.

A typical lock-and-key sensing strategy, relying only on the most dominant interactions between the probe and target, could be too limiting. In reality, the information received upon sensing is much richer. Non-specific events due to various intermolecular forces contribute to the overall received information with different degrees, and when analyzed, could provide a much more powerful detection opportunity. Here, we have assembled a highly selective universal sensor array using water-soluble two-dimensional nanoparticles (nGO, MoS2 and WS2) and fluorescent DNA molecules. The array is composed of 12 fluorescently silent non-specific nanoreceptors (2D-nps) and used for the identification of three radically different systems; five proteins, three types of live breast cancer cells and a structure-switching event of a macromolecule. The data matrices for each system were processed using Partial Least Squares (PLS) discriminant analysis. In all of the systems, the sensor array was able to identify each object or event as separate clusters with 95% confidence and without any overlap. Out of 15 unknown entities with unknown protein concentrations tested, 14 of them were predicted successfully with correct concentration. 8 breast cancer cell samples out of 9 unknown entities from three cell types were predicted correctly. During the assembly of each nanoprobe, the intrinsic non-covalent interactions between unmodified 2D nanoparticles and ssDNAs were exploited. The unmodified 2D materials offer remarkable simplicity in the layout and the use of ssDNAs as probes provides limitless possibilities because the natural interaction of a ssDNA and 2D surface can be fine-tuned with the nucleobase composition, oligonucleotide length and type of 2D nanomaterial. Therefore, the approach described here can be advanced and fine-tuned indefinitely for meeting a particular sensing criterion. Though we have only studied three distinct elements, this approach is universal enough to be applied to a wide-range of systems.

Single-trigger dual-responsive nanoparticles for controllable and sequential prodrug activation.

Here we have developed a novel approach where two synergistically acting drugs were completely inactivated upon chemical immobilization on a nanoparticle template and activated in response to a chemical stimulus. The activation rate of each drug payload is controlled using a biologically inert bioorthogonal chemistry approach. By exploiting the subtle differences in the 'click-to-release' bioorthogonal reaction, we engineered a single delivery platform capable of releasing the payloads in a time-staggered manner in response to a single dose of a highly specific, yet reactive, small molecule. Incorporation of both di-axial, 'fast release', and di-equatorial, 'slow release', TCO linkers into our nanodrug assembly inhibited the activity of the drug molecules and enabled us to control the timing and activation of each payload. This single-trigger dual-responsive nanoparticle construct and its release kinetics were characterized using two molecular fluorescent probes and tested in vitro for efficient delivery of molecular payloads. In this manuscript we show that this approach was also successful in the treatment of triple negative BT-20 breast cancer cells. Our nanodrug loaded with the slow-releasing doxorubicin and fast-releasing PAC-1 prodrugs displayed a greater therapeutic response than the nanodrug which released both payloads simultaneously.

Low picomolar, instrument-free visual detection of mercury and silver ions using low-cost programmable nanoprobes.

The EPA's recommended maximum allowable level of inorganic mercury in drinking water is 2 ppb (10 nM). To our knowledge, the most sensitive colorimetric mercury sensor reported to date has a limit of detection (LOD) of 800 pM. Here, we report an instrument-free and highly practical colorimetric methodology, which enables detection of as low as 2 ppt (10 pM) of mercury and/or silver ions with the naked eye using a gold nanoprobe. Synthesis of the nanoprobe costs less than $1.42, which is enough to perform 200 tests in a microplate; less than a penny for each test. We have demonstrated the detection of inorganic mercury from water, soil and urine samples. The assay takes about four hours and the color change is observed within minutes after the addition of the last required element of the assay. The nanoprobe is highly programmable which allows for the detection of mercury and/or silver ions separately or simultaneously by changing only a single parameter of the assay. This highly sensitive approach for the visual detection relies on the combination of the signal amplification features of the hybridization chain reaction with the plasmonic properties of the gold nanoparticles. Considering that heavy metal ion contamination of natural resources is a major challenge and routine environmental monitoring is needed, yet time-consuming, this colorimetric approach may be instrumental for on-site heavy metal ion detection. Since the color transition can be measured in a variety of formats including using the naked eye, a simple UV-Vis spectrophotometer, or recording using mobile phone apps for future directions, our cost-efficient assay and method have the potential to be translated into the field.

Rapid Visual Screening and Programmable Subtype Classification of Ebola Virus Biomarkers.

The massive outbreaks of the highly transmissible and lethal Ebola virus disease were caused by infection with one of the Ebolavirus species. It is vital to develop cost-effective, highly sensitive and selective multitarget biosensing platforms that allow for both the detection and phenotyping. Here, a highly programmable, cost-efficient and multianalyte sensing approach is reported that enables visual detection and differentiation of conserved oligonucleotide regions of all Ebolavirus subtypes known to infect human primates. This approach enables the detection of as little as 400 amols (24 × 106 molecules) of target sequences with the naked eye. Furthermore, the detection assay can be used to classify four virus biomarkers using a single nanoprobe template. This can be achieved by using different combinations of short single stranded initiator molecules, referred to as programming units, which also enable the simultaneous and rapid identification of the four biomarkers in 16 different combinations. The results of 16 × 5 array studies illustrate that the system is extremely selective with no false-positive or false-negative. Finally, the target strands in liquid biopsy mimics prepared from urine specimens are also able to be identified and classified.

Unlocked Nucleic Acids for miRNA detection using two dimensional nano-graphene oxide.

In this study we have used Unlocked Nucleic Acids (UNAs) to discriminate a breast cancer oncomiR from two other miRNAs in the same RNA family using two-dimensional graphene oxide nanoassemblies. Fluorescently labeled single stranded probe strands and graphene oxide nanoassemblies have been used to detect miR-10b and discriminate it from miR-10a, which differs by only a single nucleotide (12th base from the 5' end), and miR-10c, which differs by only two nucleotides (12th and 16th bases from the 5' end). We have determined the discrimination efficacy and detection capacity of a DNA probe with two inserted UNA monomers (UNA2), and compared it to the DNA probe with two purposefully inserted mutations (DNAM2) and full complementary sequence (DNAfull). We have observed that UNA2 is 50 times more powerful than DNAfull in discriminating miR-10b from miR-10c while generating an equally high fluorescence signal. This fluorescence signal was then further enhanced with the use of the highly specific endonuclease dsDNase for an enzymatic amplification step. The results demonstrate that the underutilized UNAs have enormous potential for miRNA detection and offer remarkable discrimination efficacy over single and double mismatches.

Complex Thermodynamic Behavior of Single-Stranded Nucleic Acid Adsorption to Graphene Surfaces.

In just over a decade since its discovery, research on graphene has exploded due to a number of potential applications in electronics, materials, and medicine. In its water-soluble form of graphene oxide, the material has shown promise as a biosensor due to its preferential absorption of single-stranded polynucleotides and fluorescence quenching properties. The rational design of these biosensors, however, requires an improved understanding of the binding thermodynamics and ultimately a predictive model of sequence-specific binding. Toward these goals, here we directly measured the binding of nucleosides and oligonucleotides to graphene oxide nanoparticles using isothermal titration calorimetry and used the results to develop molecular models of graphene-nucleic acid interactions. We found individual nucleosides binding KD values lie in the submillimolar range with binding order of rG < rA < rC < dT < rU, while 5mer and 15mer oligonucleotides had markedly higher binding affinities in the range of micromolar and submicromolar KD values, respectively. The molecular models developed here are calibrated to quantitatively reproduce the above-mentioned experimental results. For oligonucleotides, our model predicts complex binding features such as double-stacked bases and a decrease in the fraction of graphene stacked bases with increasing oligonucleotide length until plateauing beyond ∼10-15 nucleotides. These experimental and computational results set the platform for informed design of graphene-based biosensors, further increasing their potential and application.

Reprogrammable multiplexed detection of circulating oncomiRs using hybridization chain reaction.

In this study, we have coupled the DNA polymerization capability of hybridization chain reaction (HCR) with the plasmonic properties of gold nanoparticles to develop a reprogrammable and multiplexed detection of three circulating oncomiRs (miR-10b, miR-21 and miR-141) dysregulated in various disease states of breast cancer. We have demonstrated that by simply changing the initiator (label-free short single stranded DNA) content of the HCR, while keeping everything else unchanged, the same nanoparticle assembly can be reprogrammed for the detection of the target oncomiRs individually or simultaneously in all possible combinations. We have shown that as little as 20 femtomoles of each oncomiR can be detected visually without using any analytical instrument. Furthermore, we demonstrated that the target oncomiR can be detected in an RNA pool isolated from a liquid biopsy mimic of breast cancer.

Multiplexed Activity of perAuxidase: DNA-Capped AuNPs Act as Adjustable Peroxidase.

In this study, we have investigated the intrinsic peroxidase-like activity of citrate-capped AuNPs (perAuxidase) and demonstrated that the nanozyme function can be multiplexed and tuned by integrating oligonucleotides on a nanoparticle surface. Systematic studies revealed that by controlling the reaction parameters, the mutiplexing effect can be delayed or advanced and further used for aptasensor applications.

Discriminating a Single Nucleotide Difference for Enhanced miRNA Detection Using Tunable Graphene and Oligonucleotide Nanodevices.

In this study we have reported our efforts to address some of the challenges in the detection of miRNAs using water-soluble graphene oxide and DNA nanoassemblies. Purposefully inserting mismatches at specific positions in our DNA (probe) strands shows increasing specificity against our target miRNA, miR-10b, over miR-10a which varies by only a single nucleotide. This increased specificity came at a loss of signal intensity within the system, but we demonstrated that this could be addressed with the use of DNase I, an endonuclease capable of cleaving the DNA strands of the RNA/DNA heteroduplex and recycling the RNA target to hybridize to another probe strand. As we previously demonstrated, this enzymatic signal also comes with an inherent activity of the enzyme on the surface-adsorbed probe strands. To remove this activity of DNase I and the steady nonspecific increase in the fluorescence signal without compromising the recovered signal, we attached a thermoresponsive PEGMA polymer (poly(ethylene glycol) methyl ether methacrylate) to nGO. This smart polymer is able to shield the probes adsorbed on the nGO surface from the DNase I activity and is capable of tuning the detection capacity of the nGO nanoassembly with a thermoswitch at 39 °C. By utilizing probes with multiple mismatches, DNase I cleavage of the DNA probe strands, and the attachment of PEGMA polymers to graphene oxide to block undesired DNase I activity, we were able to detect miR-10b from liquid biopsy mimics and breast cancer cell lines. Overall we have reported our efforts to improve the specificity, increase the sensitivity, and eliminate the undesired enzymatic activity of DNase I on surface-adsorbed probes for miR-10b detection using water-soluble graphene nanodevices. Even though we have demonstrated only the discrimination of miR-10b from miR-10a, our approach can be extended to other short RNA molecules which differ by a single nucleotide.

Monitoring the multitask mechanism of DNase I activity using graphene nanoassemblies.

Here we have demonstrated that graphene serves as a remarkable platform for monitoring the multitask activity of an enzyme with fluorescence spectroscopy. Our studies showed that four different simultaneous enzymatic tasks of DNase I can be observed and measured in a high throughput fashion using graphene oxide and oligonucleotide nanoassemblies. We have used phosphorothioate modified oligonucleotides to pinpoint the individual and highly specific functions of DNase I with single stranded DNA, RNA, and DNA/DNA and DNA/RNA duplexes. DNase I resulted in fluorescence recovery in the nanoassemblies and enhanced the intensity tremendously in the presence of sequence specific DNA or RNA molecules with different degrees of amplification. Our study enabled us to discover the sources of this remarkable signal enhancement, which has been used for biomedical applications of graphene for sensitive detection of specific oncogenes. The significant difference in the signal amplification observed for the detection of DNA and RNA molecules is a result of the positive and/or reductive signal generating events with the enzyme. In the presence of DNA there are four possible ways that the fluorescence reading is influenced, with two of them resulting in a gain in signal while the other two result in a loss. Since the observed signal is a summation of all the events together, the absence of the two fluorescence reduction events with RNA gives a greater degree of fluorescence signal enhancement when compared to target DNA molecules. Overall, our study demonstrates that graphene has powerful features for determining the enzymatic functions of a protein and reveals some of the unknowns observed in the graphene and oligonucleotide assemblies with DNase I.

Simultaneous detection of circulating oncomiRs from body fluids for prostate cancer staging using nanographene oxide.

Circulating oncomiRs are highly stable diagnostic, prognostic, and therapeutic tumor biomarkers, which can reflect the status of the disease and response to cancer therapy. miR-141 is an oncomiR, which is overexpressed in advanced prostate cancer patients, whereas its expression is at the normal levels in the early stages of the disease. On the other hand, miR-21 is significantly elevated in the early stage, but not in the advanced prostate cancer. Here, we have demonstrated simultaneous detection of exogenous miR-21 and miR-141 from human body fluids including blood, urine and saliva using nanographene oxide. Our system enables us to specifically and reliably detect each oncomiR at different fluorescence emission channels from a large population of RNAs extracted from body fluids. We were also able to determine the content and the ratio of the miR-21 and miR-141 in 10 different miRNA cocktails composed of various, but unknown, concentrations of both oncomiRs. A strong agreement (around 90%) between the experimental results and the actual miRNA compositions was observed. Moreover, we have demonstrated that overexpressed miR-21 or miR-141 increases the fluorescence only at their signature wavelengths of 520 and 670 nm, respectively. The approach in this study combines two emerging fields of nanographene in biomedicine and the role of circulating miRNAs in cancer. Our strategy has the potential to address the current challenges in diagnosis, prognosis and staging of prostate cancer with a non- or minimally invasive approach.

The role of microRNA in resistance to breast cancer therapy.

MicroRNAs (miRNAs) are small noncoding RNA molecules with big implications in cancer. The abnormal expression of specific miRNAs has been linked to development of many cancer types. Dysregulated miRNAs play a significant role in proliferation, invasion, differentiation, apoptosis, and resistance of various cancer cells, and considered as oncogenes or tumor-suppressor genes. Findings have shown abnormal expression of specific miRNAs in breast tumors is a strong indication about the resistance to conventional cancer therapy methods. Acquired cancer resistance is a complex, multifactorial occurrence that requires various mechanisms and processes, however, recent studies have suggested that resistance may be linked to treatment-induced dysregulation of miRNAs. This dysregulation of miRNAs can affect the protein expression in cells, the ability for anti-cancer drugs to reach their targets within cells, and the apoptotic pathways. Controlling the expression of these miRNAs alters the resistant phenotype of breast cancer to a nonresistant one. This review focuses on the role of dysregulated miRNAs in breast cancer that are linked to resistance against chemo-, radiation, hormone, and targeted therapies. Finally, the role of miRNAs in breast cancer metastasis is briefly discussed.