Many different measurement techniques are used to record neural activity in the brains of different organisms, including fMRI, EEG, MEG, lightsheet microscopy and direct recordings with electrodes. Each of these measurement modes have their advantages and disadvantages concerning the resolution of the data in space and time, the directness of measurement of the neural activity and which organisms they can be applied to. For some of these modes and for some organisms, significant amounts of data are now available in large standardized open-source datasets. I will report on our efforts to apply causal discovery algorithms to, among others, fMRI data from the Human Connectome Project, and to lightsheet microscopy data from zebrafish larvae. In particular, I will focus on the challenges we have faced both in terms of the nature of the data and the computational features of the discovery algorithms, as well as the modeling of experimental interventions.
4. resting state fMRI data in the HCP
4
• The subjects are asked to lie with eyes open, with “relaxed” fixation
on a white cross (on a dark background), think of nothing in
particular, and not to fall asleep.
• Four ~15-minute rfMRI runs are acquired in two separate fMRI
sessions
• Acquisition parameters: 1200 volumes, TR = 720 ms, voxel size =
2mm isotropic, 72 slices
11. 7
10-100ms interactions
~100billionneurons
5-6s hemodynamic delay
• indirect measurement of neural activity by
BOLD signal
• aggregation in space from voxel to parcel
• aggregation in time by an order of magnitude
~90,000voxelsBOLDsignal 720ms measurement interval
}<latexit sha1_base64="DsDSimPTxv/Qm15+jZsruPb2YT4=">AAAB6XicdVDLSsNAFL3xWeur6tLNYBFclaQJNu6KblxWsQ9oQ5lMJ+3QySTMTIQS+gduXCji1j9y5984fQgqeuDC4Zx7ufeeMOVMadv+sFZW19Y3Ngtbxe2d3b390sFhSyWZJLRJEp7ITogV5UzQpmaa004qKY5DTtvh+Grmt++pVCwRd3qS0iDGQ8EiRrA20m1v2i+V7YrtedWqjwxxa37NNcTzzy9cFzkVe44yLNHol957g4RkMRWacKxU17FTHeRYakY4nRZ7maIpJmM8pF1DBY6pCvL5pVN0apQBihJpSmg0V79P5DhWahKHpjPGeqR+ezPxL6+b6cgPcibSTFNBFouijCOdoNnbaMAkJZpPDMFEMnMrIiMsMdEmnKIJ4etT9D9pVSuOXXFuvHL9chlHAY7hBM7AgRrU4Roa0AQCETzAEzxbY+vRerFeF60r1nLmCH7AevsEE0uNuA==</latexit><latexit sha1_base64="DsDSimPTxv/Qm15+jZsruPb2YT4=">AAAB6XicdVDLSsNAFL3xWeur6tLNYBFclaQJNu6KblxWsQ9oQ5lMJ+3QySTMTIQS+gduXCji1j9y5984fQgqeuDC4Zx7ufeeMOVMadv+sFZW19Y3Ngtbxe2d3b390sFhSyWZJLRJEp7ITogV5UzQpmaa004qKY5DTtvh+Grmt++pVCwRd3qS0iDGQ8EiRrA20m1v2i+V7YrtedWqjwxxa37NNcTzzy9cFzkVe44yLNHol957g4RkMRWacKxU17FTHeRYakY4nRZ7maIpJmM8pF1DBY6pCvL5pVN0apQBihJpSmg0V79P5DhWahKHpjPGeqR+ezPxL6+b6cgPcibSTFNBFouijCOdoNnbaMAkJZpPDMFEMnMrIiMsMdEmnKIJ4etT9D9pVSuOXXFuvHL9chlHAY7hBM7AgRrU4Roa0AQCETzAEzxbY+vRerFeF60r1nLmCH7AevsEE0uNuA==</latexit><latexit sha1_base64="DsDSimPTxv/Qm15+jZsruPb2YT4=">AAAB6XicdVDLSsNAFL3xWeur6tLNYBFclaQJNu6KblxWsQ9oQ5lMJ+3QySTMTIQS+gduXCji1j9y5984fQgqeuDC4Zx7ufeeMOVMadv+sFZW19Y3Ngtbxe2d3b390sFhSyWZJLRJEp7ITogV5UzQpmaa004qKY5DTtvh+Grmt++pVCwRd3qS0iDGQ8EiRrA20m1v2i+V7YrtedWqjwxxa37NNcTzzy9cFzkVe44yLNHol957g4RkMRWacKxU17FTHeRYakY4nRZ7maIpJmM8pF1DBY6pCvL5pVN0apQBihJpSmg0V79P5DhWahKHpjPGeqR+ezPxL6+b6cgPcibSTFNBFouijCOdoNnbaMAkJZpPDMFEMnMrIiMsMdEmnKIJ4etT9D9pVSuOXXFuvHL9chlHAY7hBM7AgRrU4Roa0AQCETzAEzxbY+vRerFeF60r1nLmCH7AevsEE0uNuA==</latexit><latexit sha1_base64="DsDSimPTxv/Qm15+jZsruPb2YT4=">AAAB6XicdVDLSsNAFL3xWeur6tLNYBFclaQJNu6KblxWsQ9oQ5lMJ+3QySTMTIQS+gduXCji1j9y5984fQgqeuDC4Zx7ufeeMOVMadv+sFZW19Y3Ngtbxe2d3b390sFhSyWZJLRJEp7ITogV5UzQpmaa004qKY5DTtvh+Grmt++pVCwRd3qS0iDGQ8EiRrA20m1v2i+V7YrtedWqjwxxa37NNcTzzy9cFzkVe44yLNHol957g4RkMRWacKxU17FTHeRYakY4nRZ7maIpJmM8pF1DBY6pCvL5pVN0apQBihJpSmg0V79P5DhWahKHpjPGeqR+ezPxL6+b6cgPcibSTFNBFouijCOdoNnbaMAkJZpPDMFEMnMrIiMsMdEmnKIJ4etT9D9pVSuOXXFuvHL9chlHAY7hBM7AgRrU4Roa0AQCETzAEzxbY+vRerFeF60r1nLmCH7AevsEE0uNuA==</latexit>
}<latexit sha1_base64="DsDSimPTxv/Qm15+jZsruPb2YT4=">AAAB6XicdVDLSsNAFL3xWeur6tLNYBFclaQJNu6KblxWsQ9oQ5lMJ+3QySTMTIQS+gduXCji1j9y5984fQgqeuDC4Zx7ufeeMOVMadv+sFZW19Y3Ngtbxe2d3b390sFhSyWZJLRJEp7ITogV5UzQpmaa004qKY5DTtvh+Grmt++pVCwRd3qS0iDGQ8EiRrA20m1v2i+V7YrtedWqjwxxa37NNcTzzy9cFzkVe44yLNHol957g4RkMRWacKxU17FTHeRYakY4nRZ7maIpJmM8pF1DBY6pCvL5pVN0apQBihJpSmg0V79P5DhWahKHpjPGeqR+ezPxL6+b6cgPcibSTFNBFouijCOdoNnbaMAkJZpPDMFEMnMrIiMsMdEmnKIJ4etT9D9pVSuOXXFuvHL9chlHAY7hBM7AgRrU4Roa0AQCETzAEzxbY+vRerFeF60r1nLmCH7AevsEE0uNuA==</latexit><latexit sha1_base64="DsDSimPTxv/Qm15+jZsruPb2YT4=">AAAB6XicdVDLSsNAFL3xWeur6tLNYBFclaQJNu6KblxWsQ9oQ5lMJ+3QySTMTIQS+gduXCji1j9y5984fQgqeuDC4Zx7ufeeMOVMadv+sFZW19Y3Ngtbxe2d3b390sFhSyWZJLRJEp7ITogV5UzQpmaa004qKY5DTtvh+Grmt++pVCwRd3qS0iDGQ8EiRrA20m1v2i+V7YrtedWqjwxxa37NNcTzzy9cFzkVe44yLNHol957g4RkMRWacKxU17FTHeRYakY4nRZ7maIpJmM8pF1DBY6pCvL5pVN0apQBihJpSmg0V79P5DhWahKHpjPGeqR+ezPxL6+b6cgPcibSTFNBFouijCOdoNnbaMAkJZpPDMFEMnMrIiMsMdEmnKIJ4etT9D9pVSuOXXFuvHL9chlHAY7hBM7AgRrU4Roa0AQCETzAEzxbY+vRerFeF60r1nLmCH7AevsEE0uNuA==</latexit><latexit sha1_base64="DsDSimPTxv/Qm15+jZsruPb2YT4=">AAAB6XicdVDLSsNAFL3xWeur6tLNYBFclaQJNu6KblxWsQ9oQ5lMJ+3QySTMTIQS+gduXCji1j9y5984fQgqeuDC4Zx7ufeeMOVMadv+sFZW19Y3Ngtbxe2d3b390sFhSyWZJLRJEp7ITogV5UzQpmaa004qKY5DTtvh+Grmt++pVCwRd3qS0iDGQ8EiRrA20m1v2i+V7YrtedWqjwxxa37NNcTzzy9cFzkVe44yLNHol957g4RkMRWacKxU17FTHeRYakY4nRZ7maIpJmM8pF1DBY6pCvL5pVN0apQBihJpSmg0V79P5DhWahKHpjPGeqR+ezPxL6+b6cgPcibSTFNBFouijCOdoNnbaMAkJZpPDMFEMnMrIiMsMdEmnKIJ4etT9D9pVSuOXXFuvHL9chlHAY7hBM7AgRrU4Roa0AQCETzAEzxbY+vRerFeF60r1nLmCH7AevsEE0uNuA==</latexit><latexit sha1_base64="DsDSimPTxv/Qm15+jZsruPb2YT4=">AAAB6XicdVDLSsNAFL3xWeur6tLNYBFclaQJNu6KblxWsQ9oQ5lMJ+3QySTMTIQS+gduXCji1j9y5984fQgqeuDC4Zx7ufeeMOVMadv+sFZW19Y3Ngtbxe2d3b390sFhSyWZJLRJEp7ITogV5UzQpmaa004qKY5DTtvh+Grmt++pVCwRd3qS0iDGQ8EiRrA20m1v2i+V7YrtedWqjwxxa37NNcTzzy9cFzkVe44yLNHol957g4RkMRWacKxU17FTHeRYakY4nRZ7maIpJmM8pF1DBY6pCvL5pVN0apQBihJpSmg0V79P5DhWahKHpjPGeqR+ezPxL6+b6cgPcibSTFNBFouijCOdoNnbaMAkJZpPDMFEMnMrIiMsMdEmnKIJ4etT9D9pVSuOXXFuvHL9chlHAY7hBM7AgRrU4Roa0AQCETzAEzxbY+vRerFeF60r1nLmCH7AevsEE0uNuA==</latexit>
parcel 1
parcel 405
…
What are we measuring?
12. Fast Greedy Equivalence Search (FGES)
8
Chickering (2002) JMLR
Ramsey (2017) IJDSA
• greedy local search for a DAG
• uses Bayesian Information Criterion (BIC) to determine which edge to
add/delete at any stage in the algorithm
• one free “sparsity” parameter that regulates fit vs. model complexity
• is consistent: given its assumptions it will identify the Markov
equivalence class of the true DAG in the large sample limit
13. Fast Greedy Equivalence Search (FGES)
8
• Assumptions
- acyclic causal structure
- no confounding
- linear Gaussian structural equation model
- iid data
Chickering (2002) JMLR
Ramsey (2017) IJDSA
• greedy local search for a DAG
• uses Bayesian Information Criterion (BIC) to determine which edge to
add/delete at any stage in the algorithm
• one free “sparsity” parameter that regulates fit vs. model complexity
• is consistent: given its assumptions it will identify the Markov
equivalence class of the true DAG in the large sample limit
14. Fast Greedy Equivalence Search (FGES)
8
• Assumptions
- acyclic causal structure
- no confounding
- linear Gaussian structural equation model
- iid data
Chickering (2002) JMLR
Ramsey (2017) IJDSA
• greedy local search for a DAG
• uses Bayesian Information Criterion (BIC) to determine which edge to
add/delete at any stage in the algorithm
• one free “sparsity” parameter that regulates fit vs. model complexity
• is consistent: given its assumptions it will identify the Markov
equivalence class of the true DAG in the large sample limit
• scales to >100,000 variables
- implementation in java in the Tetrad code
package
- our python implementation here: https://
github.com/eberharf/fges-py/
23. Fast Greedy Equivalence Search
• acyclic causal structure
• no unmeasured confounding
• very scalable
Causal Orientation
24. Fast Greedy Equivalence Search
• acyclic causal structure
• no unmeasured confounding
• very scalable
Answer Set Programming for Causality
• cyclic or acyclic causal structure
• unmeasured confounding permitted
• does not scale
Causal Orientation
25. 56
59
81
82108
109
110
Fast Greedy Equivalence Search
• acyclic causal structure
• no unmeasured confounding
• very scalable
Answer Set Programming for Causality
• cyclic or acyclic causal structure
• unmeasured confounding permitted
• does not scale
Causal Orientation
28. true causal structure
Observation to Intervention
x y
z
equivalence class of causal
structures that we learned
x y
z
intervened causal structure
x y
z
29. 15
3.0 s
EPI scan EPI scan
Stimulation
period (30 s)
Electrical stimulation
Stimulus
isolator
Waveform
generation
computer
+ -
B
100 ms
a,b = 0.25 ms
c = 0.75 ms
a b c
d
e
R
(-2.4, -8.2,Subj.292)
L
electrical stimulation fMRI
31. peak activation
depletion
-10 -5 0 5
-40-200204060
run 1, mean of 10 stimulus blocks stacked
timesteps before and after onset of stimulus
activation
-10 -5 0 5
-40-2002040
run 2, mean of 10 stimulus blocks stacked
timesteps before and after onset of stimulus
activation
-10 -5 0 5
-60-40-20020406080
run 3, mean of 10 stimulus blocks stacked
timesteps before and after onset of stimulus
activation
-10 -5 0 5
-60-40-2002040
run 4, mean of 10 stimulus blocks stacked
timesteps before and after onset of stimulus
activation
stimulation stimulation
stimulation stimulation
32. 18
parcel intervention On/Off on first 5 observation
alL_AngularGyrus x
R_InsularCortex x x
R_TemporalPole x x x
R_SuperiorTemporalGyrusPosterior x
R_MiddleTemporalGyrusAnterior x
R_MiddleTemporalGyrusTemporoOccipital x
R_InferiorTemporalGyrusAnterior x
R_InferiorTemporalGyrusTemporoOccipital x
R_FrontalOrbitalCortex x
R_ParahippocampalGyrusAnterior x
R_ParahippocampalGyrusPosterior x
R_TemporalFusiformCortexPosterior x
R_PlanumPolare x
LeftCaudate x
LeftAmygdala x
RightHippocampus x x x
After adjusting for hemodynamic delay and depletion…
33. 18
parcel intervention On/Off on first 5 observation
alL_AngularGyrus x
R_InsularCortex x x
R_TemporalPole x x x
R_SuperiorTemporalGyrusPosterior x
R_MiddleTemporalGyrusAnterior x
R_MiddleTemporalGyrusTemporoOccipital x
R_InferiorTemporalGyrusAnterior x
R_InferiorTemporalGyrusTemporoOccipital x
R_FrontalOrbitalCortex x
R_ParahippocampalGyrusAnterior x
R_ParahippocampalGyrusPosterior x
R_TemporalFusiformCortexPosterior x
R_PlanumPolare x
LeftCaudate x
LeftAmygdala x
RightHippocampus x x x
—— one subject —— | population
After adjusting for hemodynamic delay and depletion…
34. Population Prior to Improve Learning for Single Subject
BIC(H) = − 2log(ℒ) + λ log(n)Bayesian Information Criterion
35. Population Prior to Improve Learning for Single Subject
BIC(H) = − 2log(ℒ) + λ log(n)Bayesian Information Criterion
Prior probability on the presence
of an edge
ϕij
36. Population Prior to Improve Learning for Single Subject
BIC(H) = − 2log(ℒ) + λ log(n)Bayesian Information Criterion
Prior probability on the presence
of an edge
ϕij
Effect of the prior probability
when adding an edge i—j
BIC(Hij) − BIC(H) − α log(
ϕij
1 − ϕij
)
37. Population Prior to Improve Learning for Single Subject
BIC(H) = − 2log(ℒ) + λ log(n)Bayesian Information Criterion
Prior probability on the presence
of an edge
ϕij
Effect of the prior probability
when adding an edge i—j
BIC(Hij) − BIC(H) − α log(
ϕij
1 − ϕij
)
sparsity
prior weight
39. Evaluation without ground truth
20
• single-subject connectivity = high-confidence edges inferred on
MyConnectome data (84 scanning sessions of one individual)
- distinct from HCP data; different scanner
40. Evaluation without ground truth
20
• single-subject connectivity = high-confidence edges inferred on
MyConnectome data (84 scanning sessions of one individual)
- distinct from HCP data; different scanner
• human-generic edges = high-confidence edges learned from
Human Connectome population data
41. Evaluation without ground truth
20
• single-subject connectivity = high-confidence edges inferred on
MyConnectome data (84 scanning sessions of one individual)
- distinct from HCP data; different scanner
• human-generic edges = high-confidence edges learned from
Human Connectome population data
• subject-specific edges = high-confidence edges from single
subject that are not human-generic edges
42. Evaluation without ground truth
20
• single-subject connectivity = high-confidence edges inferred on
MyConnectome data (84 scanning sessions of one individual)
- distinct from HCP data; different scanner
• human-generic edges = high-confidence edges learned from
Human Connectome population data
• subject-specific edges = high-confidence edges from single
subject that are not human-generic edges
‣ single-subject connectivity = 60% human-generic, 40%
subject-specific
43. Evaluation without ground truth
20
• single-subject connectivity = high-confidence edges inferred on
MyConnectome data (84 scanning sessions of one individual)
- distinct from HCP data; different scanner
• human-generic edges = high-confidence edges learned from
Human Connectome population data
• subject-specific edges = high-confidence edges from single
subject that are not human-generic edges
‣ single-subject connectivity = 60% human-generic, 40%
subject-specific
• test data: two sessions (~1,000 timepoints) of single-subject
MyConnectome data
44. Evaluation without ground truth
20
• single-subject connectivity = high-confidence edges inferred on
MyConnectome data (84 scanning sessions of one individual)
- distinct from HCP data; different scanner
• human-generic edges = high-confidence edges learned from
Human Connectome population data
• subject-specific edges = high-confidence edges from single
subject that are not human-generic edges
‣ single-subject connectivity = 60% human-generic, 40%
subject-specific
• test data: two sessions (~1,000 timepoints) of single-subject
MyConnectome data
➡ How much does the prior help to learn the single-subject
connectivity from the test data?
45. Recall for human-generic and subject-specific Connectivity
human-generic edges,
recall for different
sparsities and prior weights
46. Recall for human-generic and subject-specific Connectivity
human-generic edges,
recall for different
sparsities and prior weights
subject-specific edges,
recall for different
sparsities and prior weights
47. video from Fetcho lab: http://pages.nbb.cornell.edu/neurobio/Fetcho/what-we-do-overview/; image from https://grants.nih.gov/sites/default/files/150312_zebrafish_slides.pdf
Zebrafish Larvae
48. video from Fetcho lab: http://pages.nbb.cornell.edu/neurobio/Fetcho/what-we-do-overview/; image from https://grants.nih.gov/sites/default/files/150312_zebrafish_slides.pdf
Zebrafish Larvae
49. GIF from video Ahrens M.B. et al. Nat. Methods 10, 413–420 (2013)
https://www.sfari.org/2018/09/11/sfari-workshop-discusses-zebrafish-as-experimental-systems-to-study-autism/#ref
Lightsheet Microscopy
50. GIF from video Ahrens M.B. et al. Nat. Methods 10, 413–420 (2013)
https://www.sfari.org/2018/09/11/sfari-workshop-discusses-zebrafish-as-experimental-systems-to-study-autism/#ref
Lightsheet Microscopy
57. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
58. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
• How do we handle time delays?
59. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
• How do we handle time delays?
- (sampling rate)
60. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
• How do we handle time delays?
- (sampling rate)
- slower vs. faster causal effects
61. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
• How do we handle time delays?
- (sampling rate)
- slower vs. faster causal effects
• How to model the interventions?
62. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
• How do we handle time delays?
- (sampling rate)
- slower vs. faster causal effects
• How to model the interventions?
- (soft vs. hard)
63. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
• How do we handle time delays?
- (sampling rate)
- slower vs. faster causal effects
• How to model the interventions?
- (soft vs. hard)
- time course of the intervention
64. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
• How do we handle time delays?
- (sampling rate)
- slower vs. faster causal effects
• How to model the interventions?
- (soft vs. hard)
- time course of the intervention
- electric stimulation vs. task fMRI vs. magnetic stimulation
65. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
• How do we handle time delays?
- (sampling rate)
- slower vs. faster causal effects
• How to model the interventions?
- (soft vs. hard)
- time course of the intervention
- electric stimulation vs. task fMRI vs. magnetic stimulation
• What is the relation between the observational network and
the stimulated network?
66. Open Questions
28
• What are appropriate coarse-grained parcelations of a brain?
- How can we discover them?
• How do we handle time delays?
- (sampling rate)
- slower vs. faster causal effects
• How to model the interventions?
- (soft vs. hard)
- time course of the intervention
- electric stimulation vs. task fMRI vs. magnetic stimulation
• What is the relation between the observational network and
the stimulated network?
• [scalability of methods with weaker assumptions: fast non-
parametric independence tests; non-Gaussian / non-linear
methods etc.]
67. Caltech:
• Ralph Adolphs
• Julien Dubois
• Lynn Paul
• Mike Tyszka
• Krzysztof
Chalupka
• Pietro Perona
• Yusuke Tomina
• Daniel Wagenaar
Undergraduates:
• RJ Antonello
• Lin Lin Lee
• Samuel Liebana
• John Moss
• Ethan Pronovost
• Amy Xiong
• Mark Xu
• Dan Xu
Funding:
• Chen Institute Director Award
• Caltech Carver Mead seed
grant
• AWS credit award
• NSF 1564330
• NSF 1845958
• NIH U01 NS103780-01
Janelia:
• Misha Ahrens
• Yu Mu
• Mikail Rubinov
• Greg Fleischman
UCLA / LA
Biomed:
• Paul Mathews
• Neil Harris
• Katrina Choe
• Joshua
Schoenfield
Collaborators
Thank you!