KineticNet: Deep learning a transferable kinetic energy functional for orbital-free density functional theory.

Orbital-free density functional theory (OF-DFT) holds promise to compute ground state molecular properties at minimal cost. However, it has been held back by our inability to compute the kinetic energy as a functional of electron density alone. Here, we set out to learn the kinetic energy functional from ground truth provided by the more expensive Kohn-Sham density functional theory. Such learning is confronted with two key challenges: Giving the model sufficient expressivity and spatial context while limiting the memory footprint to afford computations on a GPU and creating a sufficiently broad distribution of training data to enable iterative density optimization even when starting from a poor initial guess. In response, we introduce KineticNet, an equivariant deep neural network architecture based on point convolutions adapted to the prediction of quantities on molecular quadrature grids. Important contributions include convolution filters with sufficient spatial resolution in the vicinity of nuclear cusp, an atom-centric sparse but expressive architecture that relays information across multiple bond lengths, and a new strategy to generate varied training data by finding ground state densities in the face of perturbations by a random external potential. KineticNet achieves, for the first time, chemical accuracy of the learned functionals across input densities and geometries of tiny molecules. For two-electron systems, we additionally demonstrate OF-DFT density optimization with chemical accuracy.

Automated tracing of myelinated axons and detection of the nodes of Ranvier in serial images of peripheral nerves.

The development of realistic neuroanatomical models of peripheral nerves for simulation purposes requires the reconstruction of the morphology of the myelinated fibres in the nerve, including their nodes of Ranvier. Currently, this information has to be extracted by semimanual procedures, which severely limit the scalability of the experiments. In this contribution, we propose a supervised machine learning approach for the detailed reconstruction of the geometry of fibres inside a peripheral nerve based on its high-resolution serial section images. Learning from sparse expert annotations, the algorithm traces myelinated axons, even across the nodes of Ranvier. The latter are detected automatically. The approach is based on classifying the myelinated membranes in a supervised fashion, closing the membrane gaps by solving an assignment problem, and classifying the closed gaps for the nodes of Ranvier detection. The algorithm has been validated on two very different datasets: (i) rat vagus nerve subvolume, SBFSEM microscope, 200 × 200 × 200 nm resolution, (ii) rat sensory branch subvolume, confocal microscope, 384 × 384 × 800 nm resolution. For the first dataset, the algorithm correctly reconstructed 88% of the axons (241 out of 273) and achieved 92% accuracy on the task of Ranvier node detection. For the second dataset, the gap closing algorithm correctly closed 96.2% of the gaps, and 55% of axons were reconstructed correctly through the whole volume. On both datasets, training the algorithm on a small data subset and applying it to the full dataset takes a fraction of the time required by the currently used semiautomated protocols. Our software, raw data and ground truth annotations are available at http://hci.iwr.uni-heidelberg.de/Benchmarks/. The development version of the code can be found at https://github.com/RWalecki/ATMA.

Neuron Segmentation With High-Level Biological Priors.

We present a novel approach to the problem of neuron segmentation in image volumes acquired by an electron microscopy. Existing methods, such as agglomerative or correlation clustering, rely solely on boundary evidence and have problems where such an evidence is lacking (e.g., incomplete staining) or ambiguous (e.g., co-located cell and mitochondria membranes). We investigate if these difficulties can be overcome by means of sparse region appearance cues that differentiate between pre- and postsynaptic neuron segments in mammalian neural tissue. We combine these cues with the traditional boundary evidence in the asymmetric multiway cut (AMWC) model, which simultaneously solves the partitioning and the semantic region labeling problems. We show that AMWC problems over superpixel graphs can be solved to global optimality with a cutting plane approach, and that the introduction of semantic class priors leads to significantly better segmentations.

Three-dimensional quantitative similarity-activity relationships (3D QSiAR) from SEAL similarity matrices.

The program SEAL is suited to describe the electrostatic, steric, hydrophobic, and hydrogen bond donor and acceptor similarity of different molecules in a quantitative manner. Similarity scores AF can be calculated for pairs of molecules, using either a certain molecular property or a sum of weighted properties. Alternatively, their mutual similarity can be derived from distances d or covariances c between SEAL-based property fields that are calculated in a regular grid. For a set of N chemically related molecules, such values form an N x N similarity matrix which can be correlated with biological activities, using either regression analysis and an appropriate variable selection procedure or partial least-squares (PLS) analysis. For the Cramer steroid data set, the test set predictivities (r2pred = 0.53-0.84) of different PLS models, based on a weighted sum of molecular properties, are superior to published results of CoMFA and CoMSIA studies (r2pred = 0.31-0.40), regardless of whether a common alignment or individual, pairwise alignments of all molecules are used in the calculation of the similarity matrices. Training and test set selections have a significant influence on the external predictivities of the models. Although the SEAL similarity score between two molecules is a single number, its value is based on the 3D properties of both molecules. The term 3D quantitative similarity-activity analyses (3D QSiAR) is proposed for approaches which correlate 3D structure-derived similarity matrices with biological activities.

Preliminary results on the prediction of countershock success with fibrillation power.

In a study of artifact-free ventricular fibrillation episodes in 54 patients, 28 of whom experienced return of spontaneous circulation (ROSC), the power of different indicators to predict the ROSC was compared. Taking the average of sensitivity, specificity and positive and negative predictive value, the dominant frequency reaches 76%, the mean amplitude 72% and fibrillation power 71%. There is little correlation between the three indicators.

Fibrillation power, an alternative method of ECG spectral analysis for prediction of countershock success in a porcine model of ventricular fibrillation.

BACKGROUND: Noninvasive prediction of defibrillation success after cardiac arrest and cardiopulmonary resuscitation (CPR) may help in determining the optimal time for a countershock, and thus increase the chance for survival. METHODS: In a porcine model (n=25) of prolonged cardiac arrest, advanced cardiac life support was provided by administration of two or three doses of either vasopressin or epinephrine after 3 or 8 min of basic life support. After 4 min of ventricular fibrillation and 18 min of life support, defibrillation was attempted. The denoised power spectral density of 10 s intervals of the ventricular fibrillation electrocardiogram (ECG) was estimated from averaged and smoothed Fourier transforms. We have eliminated the spectral contribution of artifacts from manual chest compressions and provide a definition for the contribution of ventricular fibrillation to the power spectral density. This contribution is quantified and termed "fibrillation power". RESULTS: We tested fibrillation power and two established methods in their discrimination of survivors (n=16) vs. non-survivors (n=9) in the last minute before the countershock. A fibrillation power > or =79 dB predicted successful defibrillation with sensitivity, specificity, positive predictive value and negative predictive value of 98%, 98%, 99% and 97% while a mean fibrillation frequency > or =7.7 Hz was predictive with 85%, 83%, 90% and 77% and a mean amplitude > or =0.49 mV was predictive with 95%, 90%, 94% and 91%. CONCLUSIONS: We suggest that fibrillation power is an alternative source of information on the status of a fibrillating heart and that it may match the established mean frequency and amplitude analysis of ECG in predicting successful countershock during CPR.

Carving: scalable interactive segmentation of neural volume electron microscopy images.

Interactive segmentation algorithms should respond within seconds and require minimal user guidance. This is a challenge on 3D neural electron microscopy images. We propose a supervoxel-based energy function with a novel background prior that achieves these goals. This is verified by extensive experiments with a robot mimicking human interactions. A graphical user interface offering access to an open source implementation of these algorithms is made available.

Optimal classification of long echo time in vivo magnetic resonance spectra in the detection of recurrent brain tumors.

We describe the optimal high-level postprocessing of single-voxel (1)H magnetic resonance spectra and assess the benefits and limitations of automated methods as diagnostic aids in the detection of recurrent brain tumor. In a previous clinical study, 90 long-echo-time single-voxel spectra were obtained from 52 patients and classified during follow-up (30/28/32 normal/non-progressive tumor/tumor). Based on these data, a large number of evaluation strategies, including both standard resonance line quantification and algorithms from pattern recognition and machine learning, were compared in a quantitative evaluation. Results from linear and non-linear feature extraction, including ICA, PCA and wavelet transformations, and also the data from resonance line quantification were combined systematically with different classifiers such as LDA, chemometric methods (PLS, PCR), support vector machines and ensemble methods. Classification accuracy was assessed using a leave-one-out cross-validation scheme and the area under the curve (AUC) of the receiver operator characteristic (ROC). A regularized linear regression on spectra with binned channels reached 91% classification accuracy compared with 83% from quantification. Interpreting the loadings of these regressions, we find that lipid and lactate signals are too unreliable to be used in a simple machine rule. Choline and NAA are the main source of relevant information. Overall, we find that fully automated pattern recognition algorithms perform as well as, or slightly better than, a manually controlled and optimized resonance line quantification.

Generation of pseudonative protein structures for threading.

The threading approach to protein structure prediction suffers from the limited number of substantially different folds available as templates. A method is presented for the generation of artificial protein structures, amenable to threading, by modification of native ones. The artificial structures so generated are compared to the native ones and it is shown that, within the accuracy of the pseudoenergy function or force field used, these two types of structures appear equally useful for threading. Since a multitude of pseudonative artificial structures can be generated per native structure, the pool of pseudonative template structures for threading can be enormously enlarged by the inclusion of the pseudonative artificial structures.