Review 23: Neuropixels 2.0: A miniaturized high-density probe for stable, long-term brain recordings

Neuropixels 2.0: A miniaturized high-density probe for stable, long-term brain recordings by Nick Steinmetz.

• Paper notes for Neuropixels 2.0
  • Notably, this is a pretty big collaboration. The following labs participated:
    • O’Keefe Lab at UCL
    • Moser Lab at NTNU
    • Lee Lab at Janelia Research
    • Dudman lab at Janelia Research
    • Hausser Lab at UCL
    • Steinmetz Lab at UW
    • Svoboda Lab at Janelia
    • Carandini/ Harris Lab
    • Hantman Lab
    • Haesler Lab
  • Neuropixels new probe :
    • has over 5000 sites, features
    • has 2 probes and a head stage
    • records at 786 sites at once
    • weighs over 1 gram
    • enables recording from over > 10000 recording sites during free behavior in mice
  • Stably recording neurons over days / weeks during long terms processes like learning and memory is challenging, but important for understanding neural coding.
  • Many attempts to record from devices that are flexible and less than $\mu$m in size.
    • Paper states that downside to these approaches are: make insertion difficult and do not scale at large numbers of recording sites per shank.
    • I don’t understand how small flexible devices could make insertion more difficult.
  • More rigid and larger devices (Utah array, wire tetrodes, silicon probes) record high-quality signals for 8 weeks. However, no consistent recordings of individual neurons over the scale of months.
  • Neuropixels: dense coverage along a line, another array: across a plane.
  • algorithm stabilizes device to brain motion post hoc.
  • two recording channels be simultaneously recorded in the same channel with noise penalty in snr.
    • At this point, I realized that there is a lot of past literature about these recording probes that I am missing. This feels like watching avengers before watching all the prequel movies.
  • Results
    • 4 shanks, 1280 sites per shank, 5120 total sites.
    • single wide band 14-bit data stream.
    • Hardware switches can swap recording streams enabling recording from thousands of streams/experiment.
    • can cover a plan of 750 x 720 $\mu$m during recording.
    • Fig. 1
      • Neuropixels 2.0 is much less wide than the previous version.
      • Shows LFP recording and spike waveform recording (overlapping channels).
      • Probe rasters and reproduced raster of dorsal striatum firing pattern across ten trials.
    • 7 / 8 probes recovered in working condition.
    • max time recording from implant: 309 days.
    • Fig 2.
      • the proportion of spikes has a decaying distribution over the # of days after implant. A similar trend is shown in C. and D.
      • hippocampus has a high proportion of firing neurons.
      • comparison of spike recording stats between labs that used the device.
        • I think this plot is hard to interpret because I know there are different conditions for each lab so plot does not necessarily show apples to apples.
    • Paper claims brain motion leads to progressively less spikes during the duration of a recording.
    • But they have an unsupervised learning algorithm to adjust for motion post hoc and re-detect spikes. This method increases recording stability (3E).
    • Fig 3.
      • The results from this stabilization procedure a visually very convincing.
    • Now they expand the timescale alignment to long intervals.
      • successful tracking across days, and weeks (figure 4). 83% tracking accuracy after 9 weeks.
      • wonder what that accuracy is across more than three months? But tracking 83% of neurons seems really good.
    • Average signals from two sites along a recording line and then decompose them.
    • Fig 4
      • bank 1 and bank 2 separated to two sets of neurons and the switch helps record mapping between banks using mismatch.
      • when both banks are recording noise is lower.
      • 3 configurations: both banks on, bank 1 on, bank 2 on.