Deep learning for predicting subtype classification and survival of lung adenocarcinoma on computed tomography

ObjectivesThe subtype classification of lung adenocarcinoma is important for treatment decision. This study aimed to investigate the deep learning and radiomics networks for predicting histologic subtype classification and survival of lung adenocarcinoma diagnosed through computed tomography (CT) images.MethodsA dataset of 1222 patients with lung adenocarcinoma were retrospectively enrolled from three medical institutions. The anonymised preoperative CT images and pathological labels of atypical adenomatous hyperplasia, adenocarcinoma in situ, minimally invasive adenocarcinoma, invasive adenocarcinoma (IAC) with five predominant components were obtained. These pathological labels were divided into 2-category classification (IAC; non-IAC), 3-category and 8-category. We modeled the classification task of histological subtypes based on modified ResNet-34 deep learning network, radiomics strategies and deep radiomics combined algorithm. Then we established the prognostic models in lung adenocarcinoma patients with survival outcomes. The accuracy (ACC), area under ROC curves (AUCs) and C-index were primarily performed to evaluate the algorithms.ResultsThis study included a training set (n = 802) and two validation cohorts (internal, n = 196; external, n = 224). The ACC of deep radiomics algorithm in internal validation achieved 0.8776, 0.8061 in the 2-category, 3-category classification, respectively. Even in 8 classifications, the AUC ranged from 0.739 to 0.940 in internal set. Further, we constructed a prognosis model that C-index was 0.892(95% CI: 0.846–0.937) in internal validation set.ConclusionsThe automated deep radiomics based triage system has achieved the great performance in the subtype classification and survival predictability in patients with CT-detected lung adenocarcinoma nodules, providing the clinical guide for treatment strategies.     

Trait–fitness associations do not predict within-species phenotypic evolution over 2 million years

Long-term patterns of phenotypic change are the cumulative results of tens of thousands to millions of years of evolution. Yet, empirical and theoretical studies of phenotypic selection are largely based on contemporary populations. The challenges in studying phenotypic evolution, in particular trait–fitness associations in the deep past, are barriers to linking micro- and macroevolution. Here, we capitalize on the unique opportunity offered by a marine colonial organism commonly preserved in the fossil record to investigate trait–fitness associations over 2 Myr. We use the density of female polymorphs in colonies of Antartothoa tongima as a proxy for fecundity, a fitness component, and investigate multivariate signals of trait–fitness associations in six time intervals on the backdrop of Pleistocene climatic shifts. We detect negative trait–fitness associations for feeding polymorph (autozooid) sizes, positive associations for autozooid shape but no particular relationship between fecundity and brood chamber size. In addition, we demonstrate that long-term trait patterns are explained by palaeoclimate (as approximated by ∂¹⁸O), and to a lesser extent by ecological interactions (i.e. overgrowth competition and substrate crowding). Our analyses show that macroevolutionary outcomes of trait evolution are not a simple scaling-up from the trait–fitness associations.     

Haisu: Hierarchically supervised nonlinear dimensionality reduction

We propose a novel strategy for incorporating hierarchical supervised label information into nonlinear dimensionality reduction techniques. Specifically, we extend t-SNE, UMAP, and PHATE to include known or predicted class labels and demonstrate the efficacy of our approach on multiple single-cell RNA sequencing datasets. Our approach, “Haisu,” is applicable across domains and methods of nonlinear dimensionality reduction. In general, the mathematical effect of Haisu can be summarized as a variable perturbation of the high dimensional space in which the original data is observed. We thereby preserve the core characteristics of the visualization method and only change the manifold to respect known or assumed class labels when provided. Our strategy is designed to aid in the discovery and understanding of underlying patterns in a dataset that is heavily influenced by parent-child relationships. We show that using our approach can also help in semi-supervised settings where labels are known for only some datapoints (for instance when only a fraction of the cells are labeled). In summary, Haisu extends existing popular visualization methods to enable a user to incorporate labels known a priori into a visualization, including their hierarchical relationships as defined by a user input graph.     

Classical phenotyping and deep learning concur on genetic control of stomatal density and area in sorghum

Stomatal density (SD) and stomatal complex area (SCA) are important traits that regulate gas exchange and abiotic stress response in plants. Despite sorghum (Sorghum bicolor) adaptation to arid conditions, the genetic potential of stomata-related traits remains unexplored due to challenges in available phenotyping methods. Hence, identifying loci that control stomatal traits is fundamental to designing strategies to breed sorghum with optimized stomatal regulation. We implemented both classical and deep learning methods to characterize genetic diversity in 311 grain sorghum accessions for stomatal traits at two different field environments. Nearly 12,000 images collected from abaxial (Ab) and adaxial (Ad) leaf surfaces revealed substantial variation in stomatal traits. Our study demonstrated significant accuracy between manual and deep learning methods in predicting SD and SCA. In sorghum, SD was 32%–39% greater on the Ab versus the Ad surface, while SCA on the Ab surface was 2%–5% smaller than on the Ad surface. Genome-Wide Association Study identified 71 genetic loci (38 were environment-specific) with significant genotype to phenotype associations for stomatal traits. Putative causal genes underlying the phenotypic variation were identified. Accessions with similar SCA but carrying contrasting haplotypes for SD were tested for stomatal conductance and carbon assimilation under field conditions. Our findings provide a foundation for further studies on the genetic and molecular mechanisms controlling stomata patterning and regulation in sorghum. An integrated physiological, deep learning, and genomic approach allowed us to unravel the genetic control of natural variation in stomata traits in sorghum, which can be applied to other plants.

High-throughput phenotyping using deep learning tools integrated with genome-wide association studies revealed genes that control SD and area in grain sorghum.     

Machine Learning for Automatic Prediction of the Quality of Electrophysiological Recordings

The quality of electrophysiological recordings varies a lot due to technical and biological variability and neuroscientists inevitably have to select “good” recordings for further analyses. This procedure is time-consuming and prone to selection biases. Here, we investigate replacing human decisions by a machine learning approach. We define 16 features, such as spike height and width, select the most informative ones using a wrapper method and train a classifier to reproduce the judgement of one of our expert electrophysiologists. Generalisation performance is then assessed on unseen data, classified by the same or by another expert. We observe that the learning machine can be equally, if not more, consistent in its judgements as individual experts amongst each other. Best performance is achieved for a limited number of informative features; the optimal feature set being different from one data set to another. With 80–90% of correct judgements, the performance of the system is very promising within the data sets of each expert but judgments are less reliable when it is used across sets of recordings from different experts. We conclude that the proposed approach is relevant to the selection of electrophysiological recordings, provided parameters are adjusted to different types of experiments and to individual experimenters.     

Genotypic Diversity within a Single Pseudomonas aeruginosa Strain Commonly Shared by Australian Patients with Cystic Fibrosis

In cystic fibrosis (CF), Pseudomonas aeruginosa undergoes intra-strain genotypic and phenotypic diversification while establishing and maintaining chronic lung infections. As the clinical significance of these changes is uncertain, we investigated intra-strain diversity in commonly shared strains from CF patients to determine if specific gene mutations were associated with increased antibiotic resistance and worse clinical outcomes. Two-hundred-and-one P. aeruginosa isolates (163 represented a dominant Australian shared strain, AUST-02) from two Queensland CF centres over two distinct time-periods (2001–2002 and 2007–2009) underwent mexZ and lasR sequencing. Broth microdilution antibiotic susceptibility testing in a subset of isolates was also performed. We identified a novel AUST-02 subtype (M3L7) in adults attending a single Queensland CF centre. This M3L7 subtype was multi-drug resistant and had significantly higher antibiotic minimum inhibitory concentrations than other AUST-02 subtypes. Prospective molecular surveillance using polymerase chain reaction assays determined the prevalence of the ‘M3L7’ subtype at this centre during 2007–2009 (170 patients) and 2011 (173 patients). Three-year clinical outcomes of patients harbouring different strains and subtypes were compared. MexZ and LasR sequences from AUST-02 isolates were more likely in 2007–2009 than 2001–2002 to exhibit mutations (mexZ: odds ratio (OR) = 3.8; 95% confidence interval (CI): 1.1–13.5 and LasR: OR = 2.5; 95%CI: 1.3–5.0). Surveillance at the adult centre in 2007–2009 identified M3L7 in 28/509 (5.5%) P. aeruginosa isolates from 13/170 (7.6%) patients. A repeat survey in 2011 identified M3L7 in 21/519 (4.0%) P. aeruginosa isolates from 11/173 (6.4%) patients. The M3L7 subtype was associated with greater intravenous antibiotic and hospitalisation requirements, and a higher 3-year risk of death/lung transplantation, than other AUST-02 subtypes (adjusted hazard ratio [HR] = 9.4; 95%CI: 2.2–39.2) and non-AUST-02 strains (adjusted HR = 4.8; 95%CI: 1.4–16.2). This suggests ongoing microevolution of the shared CF strain, AUST-02, was associated with an emerging multi-drug resistant subtype and possibly poorer clinical outcomes.     

Detection of the G17V RHOA Mutation in Angioimmunoblastic T-Cell Lymphoma and Related Lymphomas Using Quantitative Allele-Specific PCR

Angioimmunoblastic T-cell lymphoma (AITL) and peripheral T-cell lymphoma, not otherwise specified (PTCL-NOS) are subtypes of T-cell lymphoma. Due to low tumor cell content and substantial reactive cell infiltration, these lymphomas are sometimes mistaken for other types of lymphomas or even non-neoplastic diseases. In addition, a significant proportion of PTCL-NOS cases reportedly exhibit features of AITL (AITL-like PTCL-NOS). Thus disagreement is common in distinguishing between AITL and PTCL-NOS. Using whole-exome and subsequent targeted sequencing, we recently identified G17V RHOA mutations in 60–70% of AITL and AITL-like PTCL-NOS cases but not in other hematologic cancers, including other T-cell malignancies. Here, we establish a sensitive detection method for the G17V RHOA mutation using a quantitative allele-specific polymerase chain reaction (qAS-PCR) assay. Mutated allele frequencies deduced from this approach were highly correlated with those determined by deep sequencing. This method could serve as a novel diagnostic tool for 60–70% of AITL and AITL-like PTCL-NOS.     

Behavioral fingerprints predict insecticide and anthelmintic mode of action

Novel invertebrate‐killing compounds are required in agriculture and medicine to overcome resistance to existing treatments. Because insecticides and anthelmintics are discovered in phenotypic screens, a crucial step in the discovery process is determining the mode of action of hits. Visible whole‐organism symptoms are combined with molecular and physiological data to determine mode of action. However, manual symptomology is laborious and requires symptoms that are strong enough to see by eye. Here, we use high‐throughput imaging and quantitative phenotyping to measure Caenorhabditis elegans behavioral responses to compounds and train a classifier that predicts mode of action with an accuracy of 88% for a set of ten common modes of action. We also classify compounds within each mode of action to discover substructure that is not captured in broad mode‐of‐action labels. High‐throughput imaging and automated phenotyping could therefore accelerate mode‐of‐action discovery in invertebrate‐targeting compound development and help to refine mode‐of‐action categories.     

Association between Vitamin D Insufficiency and Elevated Serum Uric Acid among Middle-Aged and Elderly Chinese Han Women

Background

Association between vitamin D insufficiency and hyperuricemia has not been reported so far. We aimed to study the association of vitamin D insufficiency with elevated serum uric acid among middle-aged and elderly Chinese Han women.

Methods

We collected data from participants residing in Jinchang district of Suzhou from January to May, 2010. Serum uric acid, 25-hydroxy vitamin D and other traditional biomarkers including fasting plasma glucose and blood lipids were determined in 1726 women aged above 30 years. Association between vitamin D insufficiency and elevated uric acid was analyzed in premenopausal and postmenopausal women, respectively.

Results

Among postmenopausal women, 25-hydroxy vitamin D level of participants with elevated uric acid was lower than that of those with normal uric acid (median [interquartile range]: 35[28–57] vs 40[32–58], µg/L; P = 0.006). Elevated uric acid was more prevalent in participants with vitamin D insufficiency compared to those without vitamin D insufficiency (16.50% vs 8.08%; P<0.001). Association between vitamin D insufficiency and elevated uric acid was not significant among premenopausal women. However, participants with vitamin D insufficiency were more likely to have elevated uric acid compared with those without vitamin D insufficiency among postmenopausal women (OR, 95% CI: 2.38, 1.47–3.87). Moreover, after excluding individuals with diabetes and/or hypertension, the association of vitamin D insufficiency with elevated uric acid was still significant (OR, 95% CI: 2.48, 1.17–5.44).

Conclusions

Vitamin D insufficiency was significantly associated with elevated uric acid among postmenopausal Chinese Han women. This study suggested that a clinical trial should be conducted to confirm the association of vitamin D insufficiency with hyperuricemia.     

Neurally-constrained modeling of human gaze strategies in a change blindness task

Despite possessing the capacity for selective attention, we often fail to notice the obvious. We investigated participants’ (n = 39) failures to detect salient changes in a change blindness experiment. Surprisingly, change detection success varied by over two-fold across participants. These variations could not be readily explained by differences in scan paths or fixated visual features. Yet, two simple gaze metrics–mean duration of fixations and the variance of saccade amplitudes–systematically predicted change detection success. We explored the mechanistic underpinnings of these results with a neurally-constrained model based on the Bayesian framework of sequential probability ratio testing, with a posterior odds-ratio rule for shifting gaze. The model’s gaze strategies and success rates closely mimicked human data. Moreover, the model outperformed a state-of-the-art deep neural network (DeepGaze II) with predicting human gaze patterns in this change blindness task. Our mechanistic model reveals putative rational observer search strategies for change detection during change blindness, with critical real-world implications.     

An extensive candidate gene approach to speciation: diversity,divergence and linkage disequilibrium in candidate pigmentation genes across the European crow hybrid zone

Colouration patterns have an important role in adaptation and speciation. The European crow system, in which all-black carrion crows and grey-coated hooded crows meet in a narrow hybrid zone, is a prominent example. The marked phenotypic difference is maintained by assortative mating in the absence of neutral genetic divergence, suggesting the presence of few pigmentation genes of major effect. We made use of the rich phenotypic and genetic resources in mammals and identified a comprehensive panel of 95 candidate pigmentation genes for birds. Based on functional annotation, we chose a subset of the most promising 37 candidates, for which we developed a marker system that demonstrably works across the avian phylogeny. In total, we sequenced 107 amplicons (∼3 loci per gene, totalling 60 kb) in population samples of crows (n=23 for each taxon). Tajima''s D, Fu''s F_S, DHEW and HKA (Hudson–Kreitman–Aguade) statistics revealed several amplicons that deviated from neutrality; however, none of these showed significantly elevated differentiation between the two taxa. Hence, colour divergence in this system may be mediated by uncharacterized pigmentation genes or regulatory regions outside genes. Alternatively, the observed high population recombination rate (4N_er∼0.03), with overall linkage disequilibrium dropping rapidly within the order of few 100 bp, may compromise the power to detect causal loci with nearby markers. Our results add to the debate as to the utility of candidate gene approaches in relation to genomic features and the genetic architecture of the phenotypic trait in question.     

Estimating leaf area index using unmanned aerial vehicle data: shallow vs. deep machine learning algorithms

Measuring leaf area index (LAI) is essential for evaluating crop growth and estimating yield, thereby facilitating high-throughput phenotyping of maize (Zea mays). LAI estimation models use multi-source data from unmanned aerial vehicles (UAVs), but using multimodal data to estimate maize LAI, and the effect of tassels and soil background, remain understudied. Our research aims to (1) determine how multimodal data contribute to LAI and propose a framework for estimating LAI based on remote-sensing data, (2) evaluate the robustness and adaptability of an LAI estimation model that uses multimodal data fusion and deep neural networks (DNNs) in single- and whole growth stages, and (3) explore how soil background and maize tasseling affect LAI estimation. To construct multimodal datasets, our UAV collected red–green–blue, multispectral, and thermal infrared images. We then developed partial least square regression (PLSR), support vector regression, and random forest regression models to estimate LAI. We also developed a deep learning model with three hidden layers. This multimodal data structure accurately estimated maize LAI. The DNN model provided the best estimate (coefficient of determination [R²] = 0.89, relative root mean square error [rRMSE] = 12.92%) for a single growth period, and the PLSR model provided the best estimate (R² = 0.70, rRMSE = 12.78%) for a whole growth period. Tassels reduced the accuracy of LAI estimation, but the soil background provided additional image feature information, improving accuracy. These results indicate that multimodal data fusion using low-cost UAVs and DNNs can accurately and reliably estimate LAI for crops, which is valuable for high-throughput phenotyping and high-spatial precision farmland management.

Multimodal data fusion (red–green–blue, multispectral, and thermal infrared) using low-cost unmanned aerial vehicles in a deep neural network and machine learning framework estimates maize leaf area index     

Deep learning allows genome-scale prediction of Michaelis constants from structural features

The Michaelis constant K_M describes the affinity of an enzyme for a specific substrate and is a central parameter in studies of enzyme kinetics and cellular physiology. As measurements of K_M are often difficult and time-consuming, experimental estimates exist for only a minority of enzyme–substrate combinations even in model organisms. Here, we build and train an organism-independent model that successfully predicts K_M values for natural enzyme–substrate combinations using machine and deep learning methods. Predictions are based on a task-specific molecular fingerprint of the substrate, generated using a graph neural network, and on a deep numerical representation of the enzyme’s amino acid sequence. We provide genome-scale K_M predictions for 47 model organisms, which can be used to approximately relate metabolite concentrations to cellular physiology and to aid in the parameterization of kinetic models of cellular metabolism.

To understand the action of an enzyme, we need to know its affinity for its substrates, quantified by Michaelis constants, but these are difficult to measure experimentally. This study shows that a deep learning model that can predict them from structural features of the enzyme and substrate, providing KM predictions for all enzymes across 47 model organisms.     

Network Signatures of Survival in Glioblastoma Multiforme

To determine a molecular basis for prognostic differences in glioblastoma multiforme (GBM), we employed a combinatorial network analysis framework to exhaustively search for molecular patterns in protein-protein interaction (PPI) networks. We identified a dysregulated molecular signature distinguishing short-term (survival<225 days) from long-term (survival>635 days) survivors of GBM using whole genome expression data from The Cancer Genome Atlas (TCGA). A 50-gene subnetwork signature achieved 80% prediction accuracy when tested against an independent gene expression dataset. Functional annotations for the subnetwork signature included “protein kinase cascade,” “IκB kinase/NFκB cascade,” and “regulation of programmed cell death” – all of which were not significant in signatures of existing subtypes. Finally, we used label-free proteomics to examine how our subnetwork signature predicted protein level expression differences in an independent GBM cohort of 16 patients. We found that the genes discovered using network biology had a higher probability of dysregulated protein expression than either genes exhibiting individual differential expression or genes derived from known GBM subtypes. In particular, the long-term survivor subtype was characterized by increased protein expression of DNM1 and MAPK1 and decreased expression of HSPA9, PSMD3, and CANX. Overall, we demonstrate that the combinatorial analysis of gene expression data constrained by PPIs outlines an approach for the discovery of robust and translatable molecular signatures in GBM.     

Deep learning prediction of attention-deficit hyperactivity disorder in African Americans by copy number variation

Current understanding of the underlying molecular network and mechanism for attention-deficit hyperactivity disorder (ADHD) is lacking and incomplete. Previous studies suggest that genomic structural variations play an important role in the pathogenesis of ADHD. For effective modeling, deep learning approaches have become a method of choice, with ability to predict the impact of genetic variations involving complicated mechanisms. In this study, we examined copy number variation in whole genome sequencing from 116 African Americans ADHD children and 408 African American controls. We divided the human genome into 150 regions, and the variation intensity in each region was applied as feature vectors for deep learning modeling to classify ADHD patients. The accuracy of deep learning for predicting ADHD diagnosis is consistently around 78% in a two-fold shuffle test, compared with ∼50% by traditional k-mean clustering methods. Additional whole genome sequencing data from 351 European Americans children, including 89 ADHD cases and 262 controls, were applied as independent validation using feature vectors obtained from the African American ethnicity analysis. The accuracy of ADHD labeling was lower in this setting (∼70–75%) but still above the results from traditional methods. The regions with highest weight overlapped with the previously reported ADHD-associated copy number variation regions, including genes such as GRM1 and GRM8, key drivers of metabotropic glutamate receptor signaling. A notable discovery is that structural variations in non-coding genomic (intronic/intergenic) regions show prediction weights that can be as high as prediction weight from variations in coding regions, results that were unexpected.     

New Tests to Measure Individual Differences in Matching and Labelling Facial Expressions of Emotion,and Their Association with Ability to Recognise Vocal Emotions and Facial Identity

Although good tests are available for diagnosing clinical impairments in face expression processing, there is a lack of strong tests for assessing “individual differences” – that is, differences in ability between individuals within the typical, nonclinical, range. Here, we develop two new tests, one for expression perception (an odd-man-out matching task in which participants select which one of three faces displays a different expression) and one additionally requiring explicit identification of the emotion (a labelling task in which participants select one of six verbal labels). We demonstrate validity (careful check of individual items, large inversion effects, independence from nonverbal IQ, convergent validity with a previous labelling task), reliability (Cronbach’s alphas of.77 and.76 respectively), and wide individual differences across the typical population. We then demonstrate the usefulness of the tests by addressing theoretical questions regarding the structure of face processing, specifically the extent to which the following processes are common or distinct: (a) perceptual matching and explicit labelling of expression (modest correlation between matching and labelling supported partial independence); (b) judgement of expressions from faces and voices (results argued labelling tasks tap into a multi-modal system, while matching tasks tap distinct perceptual processes); and (c) expression and identity processing (results argued for a common first step of perceptual processing for expression and identity).     

Discovering molecular features of intrinsically disordered regions by using evolution for contrastive learning

A major challenge to the characterization of intrinsically disordered regions (IDRs), which are widespread in the proteome, but relatively poorly understood, is the identification of molecular features that mediate functions of these regions, such as short motifs, amino acid repeats and physicochemical properties. Here, we introduce a proteome-scale feature discovery approach for IDRs. Our approach, which we call “reverse homology”, exploits the principle that important functional features are conserved over evolution. We use this as a contrastive learning signal for deep learning: given a set of homologous IDRs, the neural network has to correctly choose a held-out homolog from another set of IDRs sampled randomly from the proteome. We pair reverse homology with a simple architecture and standard interpretation techniques, and show that the network learns conserved features of IDRs that can be interpreted as motifs, repeats, or bulk features like charge or amino acid propensities. We also show that our model can be used to produce visualizations of what residues and regions are most important to IDR function, generating hypotheses for uncharacterized IDRs. Our results suggest that feature discovery using unsupervised neural networks is a promising avenue to gain systematic insight into poorly understood protein sequences.