Accelerating genomic workflows using NVIDIA Parabricks.

BACKGROUND: As genome sequencing becomes better integrated into scientific research, government policy, and personalized medicine, the primary challenge for researchers is shifting from generating raw data to analyzing these vast datasets. Although much work has been done to reduce compute times using various configurations of traditional CPU computing infrastructures, Graphics Processing Units (GPUs) offer opportunities to accelerate genomic workflows by orders of magnitude. Here we benchmark one GPU-accelerated software suite called NVIDIA Parabricks on Amazon Web Services (AWS), Google Cloud Platform (GCP), and an NVIDIA DGX cluster. We benchmarked six variant calling pipelines, including two germline callers (HaplotypeCaller and DeepVariant) and four somatic callers (Mutect2, Muse, LoFreq, SomaticSniper). RESULTS: We achieved up to 65 × acceleration with germline variant callers, bringing HaplotypeCaller runtimes down from 36 h to 33 min on AWS, 35 min on GCP, and 24 min on the NVIDIA DGX. Somatic callers exhibited more variation between the number of GPUs and computing platforms. On cloud platforms, GPU-accelerated germline callers resulted in cost savings compared with CPU runs, whereas some somatic callers were more expensive than CPU runs because their GPU acceleration was not sufficient to overcome the increased GPU cost. CONCLUSIONS: Germline variant callers scaled well with the number of GPUs across platforms, whereas somatic variant callers exhibited more variation in the number of GPUs with the fastest runtimes, suggesting that, at least with the version of Parabricks used here, these workflows are less GPU optimized and require benchmarking on the platform of choice before being deployed at production scales. Our study demonstrates that GPUs can be used to greatly accelerate genomic workflows, thus bringing closer to grasp urgent societal advances in the areas of biosurveillance and personalized medicine.

Disruption of NMDAR-dependent burst firing by dopamine neurons provides selective assessment of phasic dopamine-dependent behavior.

Midbrain dopamine (DA) neurons fire in 2 characteristic modes, tonic and phasic, which are thought to modulate distinct aspects of behavior. However, the inability to selectively disrupt these patterns of activity has hampered the precise definition of the function of these modes of signaling. Here, we addressed the role of phasic DA in learning and other DA-dependent behaviors by attenuating DA neuron burst firing and subsequent DA release, without altering tonic neural activity. Disruption of phasic DA was achieved by selective genetic inactivation of NMDA-type, ionotropic glutamate receptors in DA neurons. Disruption of phasic DA neuron activity impaired the acquisition of numerous conditioned behavioral responses, and dramatically attenuated learning about cues that predicted rewarding and aversive events while leaving many other DA-dependent behaviors unaffected.

High-frequency, short-latency disinhibition bursting of midbrain dopaminergic neurons.

During reinforcement and sequence learning, dopaminergic neurons fire bursts of action potentials. Dopaminergic neurons in vivo receive strong background excitatory and inhibitory inputs, suggesting that one mechanism by which bursts may be produced is disinhibition. Unfortunately, these inputs are lost during slice preparation and are not precisely controlled during in vivo experiments. In the present study we show that dopaminergic neurons can be shifted into a balanced state in which constant synaptic N-methyl-d-aspartate (NMDA) and GABA(A) conductances are mimicked either pharmacologically or using dynamic clamp. From this state, a disinhibition burst can be evoked by removing the background inhibitory conductance. We demonstrate three functional characteristics of network-based disinhibition that promote high-frequency, short-latency bursting in dopaminergic neurons. First, we found that increasing the total background NMDA and GABA(A) synaptic conductances increased the intraburst firing frequency and reduced its latency. Second, we found that the disinhibition burst is sensitive to the proportion of background inhibitory input that is removed. In particular, we found that high-frequency, short-latency bursts were enhanced by increasing the degree of disinhibition. Third, the time course over which inhibition is removed had a large effect on the burst, namely, that synchronous removal of weak inhibitory inputs produces bursts of high intraburst frequency and shorter latency. Our results suggest that fast, more precisely timed bursts can be evoked by complete and synchronous disinhibition of dopaminergic neurons in a high-conductance state.

A dynamic role for GABA receptors on the firing pattern of midbrain dopaminergic neurons.

Dopaminergic neurons are subject to a significant background GABAergic input in vivo. The presence of this GABAergic background might be expected to inhibit dopaminergic neuron firing. However, dopaminergic neurons are not all silent but instead fire in single-spiking and burst firing modes. Here we present evidence that phasic changes in the tonic activity of GABAergic afferents are a potential extrinsic mechanism that triggers bursts and pauses in dopaminergic neurons. We find that spontaneous single-spiking is more sensitive to activation of GABA receptors than phasic N-methyl-D-aspartate (NMDA)-mediated burst firing in rat slices (P15-P31). Because tonic activation of GABA(A) receptors has previously been shown to suppress burst firing in vivo, our results suggest that the activity patterns seen in vivo are the result of a balance between excitatory and inhibitory conductances that interact with the intrinsic pacemaking currents observed in slices. Using the dynamic clamp technique, we applied balanced, constant NMDA and GABA(A) receptor conductances into dopaminergic neurons in slices. Bursts could be produced by disinhibition (phasic removal of the GABA(A) receptor conductance), and these bursts had a higher frequency than bursts produced by the same NMDA receptor conductance alone. Phasic increases in the GABA(A) receptor conductance evoked pauses in firing. In contrast to NMDA receptor, application of constant AMPA and GABA(A) receptor conductances caused the cell to go into depolarization block. These results support a bidirectional mechanism by which GABAergic inputs, in balance with NMDA receptor-mediated excitatory inputs, control the firing pattern of dopaminergic neurons.

Disinhibition bursting of dopaminergic neurons.

Substantia nigra pars compacta (SNpc) dopaminergic neurons receive strong tonic inputs from GABAergic neurons in the substantia nigra pars reticulata (SNpr) and globus pallidus (GP), and glutamatergic neurons in the subthalamic nucleus. The presence of these tonic inputs raises the possibility that phasic disinhibition may trigger phasic bursts in dopaminergic neurons. We first applied constant NMDA and GABA(A) conductances onto a two-compartment single cell model of the dopaminergic neuron (Kuznetsov et al., 2006). The model exhibited disinhibition bursting upon stepwise removal of inhibition. A further bifurcation analysis suggests that disinhibition may be more robust than excitation alone in that for most levels of NMDA conductance, the cell remains capable of bursting even after a complete removal of inhibition, whereas too much excitatory input will drive the cell into depolarization block. To investigate the network dynamics of disinhibition, we used a modified version of an integrate-and-fire based model of the basal ganglia (Humphries et al., 2006). Synaptic activity generated in the network was delivered to the two-compartment single cell dopaminergic neuron. Phasic activation of the D1-expressing medium spiny neurons in the striatum (D1STR) produced disinhibition bursts in dopaminergic neurons through the direct pathway (D1STR to SNpr to SNpc). Anatomical studies have shown that D1STR neurons have collaterals that terminate in GP. Adding these collaterals to the model, we found that striatal activation increased the intra-burst firing frequency of the disinhibition burst as the weight of this connection was increased. Our studies suggest that striatal activation is a robust means by which disinhibition bursts can be generated by SNpc dopaminergic neurons, and that recruitment of the indirect pathway via collaterals may enhance disinhibition bursting.

Application of a NMDA receptor conductance in rat midbrain dopaminergic neurons using the dynamic clamp technique.

Neuroscientists study the function of the brain by investigating how neurons in the brain communicate. Many investigators look at changes in the electrical activity of one or more neurons in response to an experimentally-controlled input. The electrical activity of neurons can be recorded in isolated brain slices using patch clamp techniques with glass micropipettes. Traditionally, experimenters can mimic neuronal input by direct injection of current through the pipette, electrical stimulation of the other cells or remaining axonal connections in the slice, or pharmacological manipulation by receptors located on the neuronal membrane of the recorded cell. Direct current injection has the advantages of passing a predetermined current waveform with high temporal precision at the site of the recording (usually the soma). However, it does not change the resistance of the neuronal membrane as no ion channels are physically opened. Current injection usually employs rectangular pulses and thus does not model the kinetics of ion channels. Finally, current injection cannot mimic the chemical changes in the cell that occurs with the opening of ion channels. Receptors can be physically activated by electrical or pharmacological stimulation. The experimenter has good temporal precision of receptor activation with electrical stimulation of the slice. However, there is limited spatial precision of receptor activation and the exact nature of what is activated upon stimulation is unknown. This latter problem can be partially alleviated by specific pharmacological agents. Unfortunately, the time course of activation of pharmacological agents is typically slow and the spatial precision of inputs onto the recorded cell is unknown. The dynamic clamp technique allows an experimenter to change the current passed directly into the cell based on real-time feedback of the membrane potential of the cell (Robinson and Kawai 1993, Sharp et al., 1993a,b; for review, see Prinz et al. 2004). This allows an experimenter to mimic the electrical changes that occur at the site of the recording in response to activation of a receptor. Real-time changes in applied current are determined by a mathematical equation implemented in hardware. We have recently used the dynamic clamp technique to investigate the generation of bursts of action potentials by phasic activation of NMDA receptors in dopaminergic neurons of the substantia nigra pars compacta (Deister et al., 2009; Lobb et al., 2010). In this video, we demonstrate the procedures needed to apply a NMDA receptor conductance into a dopaminergic neuron.