Temporally delayed linear modelling (TDLM) measures replay in both animals and humans

  1. Yunzhe Liu  Is a corresponding author
  2. Raymond J Dolan
  3. Cameron Higgins
  4. Hector Penagos
  5. Mark W Woolrich
  6. H Freyja Ólafsdóttir
  7. Caswell Barry
  8. Zeb Kurth-Nelson
  9. Timothy E Behrens
  1. State Key Laboratory of Cognitive Neuroscience and Learning, IDG/McGovern Institute for Brain Research, Beijing Normal University, China
  2. Chinese Institute for Brain Research, China
  3. Max Planck University College London Centre for Computational Psychiatry and Ageing Research, United Kingdom
  4. Wellcome Centre for Human Neuroimaging, University College London, United Kingdom
  5. Wellcome Centre for Integrative Neuroimaging, University of Oxford, United Kingdom
  6. Center for Brains, Minds and Machines, Picower Institute for Learning and Memory, Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, United States
  7. Donders Institute for Brain Cognition and Behaviour, Radboud University, Netherlands
  8. Research Department of Cell and Developmental Biology, University College London, United Kingdom
  9. DeepMind, United Kingdom
18 figures and 1 additional file

Figures

Figure 1 with 1 supplement
Task design and illustration of temporal delayed linear modelling (TDLM).

(a) Task design in both simulation and real MEG data. Assuming there is one sequence, A->B->C->D, indicated by the four objects at the top. During the task, participants are shown the objects and …

Figure 1—figure supplement 1
Source localization of stimuli-evoked neural activity in MEG.
Figure 2 with 1 supplement
Obtaining different state spaces.

(a) Assuming we have two abstract codes, each abstract code has two different sensory codes (left panel). The MEG/EEG data corresponding to each stimulus is a conjunctive representation of sensory …

Figure 2—figure supplement 1
Sequences of abstract code.

(a) Illustration of the relationship between sensory code and (abstract) structural code. Structural code cannot be accessed directly but can be indirectly obtained from the conjunctive code …

Effects of temporal, spatial correlations, and classifier regularization on temporal delayed linear modelling (TDLM).

(a) Simple linear regression or cross-correlation approach relies on an asymmetry of forward and backward transitions; therefore, subtraction is necessary (left panel). TDLM instead relies on …

Statistical inference.

(a) P-P plot of one-sample t test (blue) and Wilcoxon signed-rank test (red) against zero. This is performed in simulated MEG data, assuming autocorrelated state time courses, but no real sequences. …

Source localization of replay onset.

(a) Temporal delayed linear modelling indexes the onset of a sequence based on the identified optimal state-to-state time lag (left panel). Sequence onset during resting state from one example …

Temporal delayed linear modelling (TDLM) vs. existing rodent replay methods.

(a) The Radon method tries to find the best fitting line (solid line) of the decoded positions as a function of time. The red bars indicate strong reactivation at a given location. (b) The linear …

Temporal delayed linear modelling (TDLM) applied to real rodent data.

(a) The experimental design of Ólafsdóttir et al., 2016. Rats ran on Z maze for 30 min, followed by 90 min rest. (b) An example rate map for a place cell. The left panel shows its spatial …

Pairwise sequence and spatial modulation effect.

(a) Within each scale, strengths of each pairwise forward sequences in the region of interest (ROI) (significant replay speeds, compare with Figure 7d, green shading) are ordered from the start of …

Appendix 1—figure 1
Extension to temporal delayed linear modelling (TDLM): multi-step sequences.

(a) TDLM can quantify not only pairwise transition, but also longer length sequences. It does so by controlling for evidence of shorter length to avoid false positives. (b) Method applied to human …

Appendix 3—figure 1
Sequences of sequences.

(a) Temporal delayed linear modelling (TDLM) can also be used iteratively to capture the repeating pattern of a sequence event itself. Illustration in the top panel describes the ground truth in the …

Appendix 4—figure 1
Sequence detection in EEG data (from one participant).

(a) Task design. At each trial, the participant starts at state A, and he/she needs to select either ‘BDF’ or ‘CEG’ path, based on the final reward receipt at terminal states F or G. All seven …

Appendix 5—figure 1
Parametric relationship between space and time vs. graph transitions.

(a) A scheme for skipping sequence (left). Both Radon and linear weighted correlation methods aim to capture a parametric relationship between space and time. Temporal delayed linear modelling (TDLM)…

Author response image 1
Illustration of three replay detection methods on the decoded time by position/state spaces.

a. The Radon method tries to find the best fitting line (solid line) of the decoded positions as a function of time. The red bars indicate strong reactivation at given locations. b. The linear …

Author response image 2
"line search” approach vs. TDLM on the simulated spiking data (assuming single ground-truth sequence).

(a) The rate map of the simulated place cells (n=40) over a linearized space with 80 positions. It is smoothed with 2 sample gaussian kernel, to mimic overlapping place fields. (b) We simulated a …

Author response image 3
“line search” approach vs. TDLM on the simulated spiking data (assuming two ground-truth sequences).

(a) The rate map of the simulated place cells (n=40) over a linearized space with 80 positions, smoothed with 2 sample gaussian kernel, to mimic overlapping place fields. (b) We simulated two ground …

Author response image 4
Multi-scale TDLM.

(a) Illustration of a change in state space for the same replay speed. (b) A possible scenario for the application of multi-scale TDLM, where only subsets of state on a path were reactivated. The …

Author response image 5
TDLM applied to real rodent data.

(A) Sequence strength as a function of speed and direction is shown for outbound rate map. Dotted line is the permutation threshold. We have both significant forward (blue) and backward (red) …

Author response image 6
Circular time shift vs. state identity-based permutation on real human whole brain MEG data.

(a) The permutation (blue) is done by circularly shifting the time dimension of each state on the decoded state space of the MEG data during pre-stimuli resting time, where the ground truth is no …

Additional files

Download links