MELLODDY: Cross-pharma Federated Learning at Unprecedented Scale Unlocks Benefits in QSAR without Compromising Proprietary Information.

Federated multipartner machine learning has been touted as an appealing and efficient method to increase the effective training data volume and thereby the predictivity of models, particularly when the generation of training data is resource-intensive. In the landmark MELLODDY project, indeed, each of ten pharmaceutical companies realized aggregated improvements on its own classification or regression models through federated learning. To this end, they leveraged a novel implementation extending multitask learning across partners, on a platform audited for privacy and security. The experiments involved an unprecedented cross-pharma data set of 2.6+ billion confidential experimental activity data points, documenting 21+ million physical small molecules and 40+ thousand assays in on-target and secondary pharmacodynamics and pharmacokinetics. Appropriate complementary metrics were developed to evaluate the predictive performance in the federated setting. In addition to predictive performance increases in labeled space, the results point toward an extended applicability domain in federated learning. Increases in collective training data volume, including by means of auxiliary data resulting from single concentration high-throughput and imaging assays, continued to boost predictive performance, albeit with a saturating return. Markedly higher improvements were observed for the pharmacokinetics and safety panel assay-based task subsets.

PyDESeq2: a python package for bulk RNA-seq differential expression analysis.

SUMMARY: We present PyDESeq2, a python implementation of the DESeq2 workflow for differential expression analysis on bulk RNA-seq data. This re-implementation yields similar, but not identical, results: it achieves higher model likelihood, allows speed improvements on large datasets, as shown in experiments on TCGA data, and can be more easily interfaced with modern python-based data science tools. AVAILABILITY AND IMPLEMENTATION: PyDESeq2 is released as an open-source software under the MIT license. The source code is available on GitHub at https://github.com/owkin/PyDESeq2 and documented at https://pydeseq2.readthedocs.io. PyDESeq2 is part of the scverse ecosystem.

A kernel-based method to calculate local field potentials from networks of spiking neurons.

BACKGROUND: The local field potential (LFP) is usually calculated from current sources arising from transmembrane currents, in particular in asymmetric cellular morphologies such as pyramidal neurons. NEW METHOD: Here, we adopt a different point of view and relate the spiking of neurons to the LFP through efferent synaptic connections and provide a method to calculate LFPs. RESULTS: We show that the so-called unitary LFPs (uLFP) provide the key to such a calculation. We show experimental measurements and simulations of uLFPs in neocortex and hippocampus, for both excitatory and inhibitory neurons. We fit a "kernel" function to measurements of uLFPs, and we estimate its spatial and temporal spread by using simulations of morphologically detailed reconstructions of hippocampal pyramidal neurons. Assuming that LFPs are the sum of uLFPs generated by every neuron in the network, the LFP generated by excitatory and inhibitory neurons can be calculated by convolving the trains of action potentials with the kernels estimated from uLFPs. This provides a method to calculate the LFP from networks of spiking neurons, even for point neurons for which the LFP is not easily defined. We show examples of LFPs calculated from networks of point neurons and compare to the LFP calculated from synaptic currents. CONCLUSIONS: The kernel-based method provides a practical way to calculate LFPs from networks of point neurons.

Modelling unitary fields and the single-neuron contribution to local field potentials in the hippocampus.

KEY POINTS: We simulate the unitary local field potential (uLFP) generated in the hippocampus CA3, using morphologically detailed models. The model suggests that cancelling effects between apical and basal dendritic synapses explain the low amplitude of excitatory uLFPs. Inhibitory synapses around the soma do not cancel and could explain the high-amplitude inhibitory uLFPs. These results suggest that somatic inhibition constitutes a strong component of LFPs, which may explain a number of experimental observations. ABSTRACT: Synaptic currents represent a major contribution to the local field potential (LFP) in brain tissue, but the respective contribution of excitatory and inhibitory synapses is not known. Here, we provide estimates of this contribution by using computational models of hippocampal pyramidal neurons, constrained by in vitro recordings. We focus on the unitary LFP (uLFP) generated by single neurons in the CA3 region of the hippocampus. We first reproduce experimental results for hippocampal basket cells, and in particular how inhibitory uLFP are distributed within hippocampal layers. Next, we calculate the uLFP generated by pyramidal neurons, using morphologically reconstructed CA3 pyramidal cells. The model shows that the excitatory uLFP is of small amplitude, smaller than inhibitory uLFPs. Indeed, when the two are simulated together, inhibitory uLFPs mask excitatory uLFPs, which might create the illusion that the inhibitory field is generated by pyramidal cells. These results provide an explanation for the observation that excitatory and inhibitory uLFPs are of the same polarity, in vivo and in vitro. These results suggest that somatic inhibitory currents are large contributors to the LFP, which is important information for interpreting this signal. Finally, the results of our model might form the basis of a simple method to compute the LFP, which could be applied to point neurons for each cell type, thus providing a simple biologically grounded method for calculating LFPs from neural networks. In conclusion, computational models constrained by in vitro recordings suggest that: (1) Excitatory uLFPs are of smaller amplitude than inhibitory uLFPs. (2) Inhibitory uLFPs form the major contribution to LFPs. (3) uLFPs can be used as a simple model to generate LFPs from spiking networks.

Contribution of the Axon Initial Segment to Action Potentials Recorded Extracellularly.

Action potentials (APs) are electric phenomena that are recorded both intracellularly and extracellularly. APs are usually initiated in the short segment of the axon called the axon initial segment (AIS). It was recently proposed that at the onset of an AP the soma and the AIS form a dipole. We study the extracellular signature [the extracellular AP (EAP)] generated by such a dipole. First, we demonstrate the formation of the dipole and its extracellular signature in detailed morphological models of a reconstructed pyramidal neuron. Then, we study the EAP waveform and its spatial dependence in models with axonal AP initiation and contrast it with the EAP obtained in models with somatic AP initiation. We show that in the models with axonal AP initiation the dipole forms between somatodendritic compartments and the AIS, and not between soma and dendrites as in the classical models. The soma-dendrites dipole is present only in models with somatic AP initiation. Our study has consequences for interpreting extracellular recordings of single-neuron activity and determining electrophysiological neuron types, but also for better understanding the origins of the high-frequency macroscopic extracellular potentials recorded in the brain.

The basis of sharp spike onset in standard biophysical models.

In most vertebrate neurons, spikes initiate in the axonal initial segment (AIS). When recorded in the soma, they have a surprisingly sharp onset, as if sodium (Na) channels opened abruptly. The main view stipulates that spikes initiate in a conventional manner at the distal end of the AIS, then progressively sharpen as they backpropagate to the soma. We examined the biophysical models used to substantiate this view, and we found that spikes do not initiate through a local axonal current loop that propagates along the axon, but through a global current loop encompassing the AIS and soma, which forms an electrical dipole. Therefore, the phenomenon is not adequately modeled as the backpropagation of an electrical wave along the axon, since the wavelength would be as large as the entire system. Instead, in these models, we found that spike initiation rather follows the critical resistive coupling model proposed recently, where the Na current entering the AIS is matched by the axial resistive current flowing to the soma. Besides demonstrating it by examining the balance of currents at spike initiation, we show that the observed increase in spike sharpness along the axon is artifactual and disappears when an appropriate measure of rapidness is used; instead, somatic onset rapidness can be predicted from spike shape at initiation site. Finally, we reproduce the phenomenon in a two-compartment model, showing that it does not rely on propagation. In these models, the sharp onset of somatic spikes is therefore not an artifact of observing spikes at the incorrect location, but rather the signature that spikes are initiated through a global soma-AIS current loop forming an electrical dipole.

Single CA3 pyramidal cells trigger sharp waves in vitro by exciting interneurones.

KEY POINTS: The CA3 hippocampal region generates sharp waves (SPW), a population activity associated with neuronal representations. The synaptic mechanisms responsible for the generation of these events still require clarification. Using slices maintained in an interface chamber, we found that the firing of single CA3 pyramidal cells triggers SPW like events at short latencies, similar to those for the induction of firing in interneurons. Multi-electrode records from the CA3 stratum pyramidale showed that pyramidal cells triggered events consisting of putative interneuron spikes followed by field IPSPs. SPW fields consisted of a repetition of these events at intervals of 4-8 ms. Although many properties of induced and spontaneous SPWs were similar, the triggered events tended to be initiated close to the stimulated cell. These data show that the initiation of SPWs in vitro is mediated via pyramidal cell synapses that excite interneurons. They do not indicate why interneuron firing is repeated during a SPW. ABSTRACT: Sharp waves (SPWs) are a hippocampal population activity that has been linked to neuronal representations. We show that SPWs in the CA3 region of rat hippocampal slices can be triggered by the firing of single pyramidal cells. Single action potentials in almost one-third of pyramidal cells initiated SPWs at latencies of 2-5 ms with probabilities of 0.07-0.76. Initiating pyramidal cells evoked field IPSPs (fIPSPs) at similar latencies when SPWs were not initiated. Similar spatial profiles for fIPSPs and middle components of SPWs suggested that SPW fields reflect repeated fIPSPs. Multiple extracellular records showed that the initiated SPWs tended to start near the stimulated pyramidal cell, whereas spontaneous SPWs could emerge at multiple sites. Single pyramidal cells could initiate two to six field IPSPs with distinct amplitude distributions, typically preceeded by a short-duration extracellular action potential. Comparison of these initiated fields with spontaneously occurring inhibitory field motifs allowed us to identify firing in different interneurones during the spread of SPWs. Propagation away from an initiating pyramidal cell was typically associated with the recruitment of interneurones and field IPSPs that were not activated by the stimulated pyramidal cell. SPW fields initiated by single cells were less variable than spontaneous events, suggesting that more stereotyped neuronal ensembles were activated, although neither the spatial profiles of fields, nor the identities of interneurone firing were identical for initiated events. The effects of single pyramidal cell on network events are thus mediated by different sequences of interneurone firing.

Recurrent synapses and circuits in the CA3 region of the hippocampus: an associative network.

In the CA3 region of the hippocampus, pyramidal cells excite other pyramidal cells and interneurons. The axons of CA3 pyramidal cells spread throughout most of the region to form an associative network. These connections were first drawn by Cajal and Lorente de No. Their physiological properties were explored to understand epileptiform discharges generated in the region. Synapses between pairs of pyramidal cells involve one or few release sites and are weaker than connections made by mossy fibers on CA3 pyramidal cells. Synapses with interneurons are rather effective, as needed to control unchecked excitation. We examine contributions of recurrent synapses to epileptiform synchrony, to the genesis of sharp waves in the CA3 region and to population oscillations at theta and gamma frequencies. Recurrent connections in CA3, as other associative cortices, have a lower connectivity spread over a larger area than in primary sensory cortices. This sparse, but wide-ranging connectivity serves the functions of an associative network, including acquisition of neuronal representations as activity in groups of CA3 cells and completion involving the recall from partial cues of these ensemble firing patterns.