acceptodds
Under review as a conference paper at ICLR 2027

Probabilistic Mixed-Effects Graph Autoencoding for Individualized Multi-Paradigm Functional Connectome Modeling

Abstract

Repeated observations of the same individual are common in biomedical data, and multi-paradigm fMRI is a prominent example: every scan reflects the current task while sharing a stable individual signature with the participant’s other scans. Most representation-learning methods treat scans as independent samples or leave this shared component implicit. We formulate functional connectome reconstruction under repeated measurements as a conditional neural mixed-effects problem and propose P-MEGA, a probabilistic mixed-effects graph autoencoder that decomposes each functional connectome into a scan-specific population-level prediction and a low-rank participant-specific random effect shared across scans, which we refer to as the participant’s brain fingerprint. A signed multi-window graph encoder and a shared edge decoder form the population-level pathway, while a permutation-invariant set encoder parameterizes an amortized Gaussian posterior over the participant-specific effect from an arbitrary set of support scans, enabling inference for unseen participants with all model parameters shared across the cohort. We evaluate P-MEGA on two large multi-paradigm cohorts widely used to study individual differences and brain development: the Human Connectome Project (HCP; 865 young adults, 9 paradigms) and the Philadelphia Neurodevelopmental Cohort (PNC; 1,272 youths aged 8–21, 3 paradigms). P-MEGA attains the best out-of-fold reconstruction among the evaluated baselines. A support-swap test, which fixes the target scan and changes only whose scans are used to infer the fingerprint, shows that fingerprints inferred from the same participant’s other paradigms explain the target’s residual connectivity better than those from demographically matched or random participants in both cohorts. The inferred fingerprints are reproducible across training seeds and folds and show network-level associations with cognition and age.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Functional connectivity, graph autoencoder, mixed-effects models, multi-paradigm fMRI, variational inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WYegOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

P-MEGA combines a scan-conditioned graph autoencoder with a probabilistic low-rank participant effect inferred from multiple scans. Experiments on HCP and PNC report improved connectome reconstruction, and target-excluded support-swap tests show that same-participant support improves residual reconstruction over matched or random donors. The paper also examines downstream prediction and network-level organization of the learned effects.

Critical strengths
  1. The target-excluded support-swap experiment provides meaningful evidence of participant-specific information beyond demographic matching and generic corrections.
  2. Evaluation on two cohorts with participant-separated cross-validation, multiple training seeds, and retrained component ablations supports the reported reconstruction gains.
Critical weaknesses
  1. The interpretation of the participant effect as a distinct stable trait requires further justification because the scan-conditioned population pathway can also retain participant information.
  2. Training with a fixed support count within each cohort leaves the explicit log K dependence in Equation (7) underdetermined at other support sizes.
  3. The central transfer experiment does not establish the method's comparative advantage. Table 3 demonstrates participant-specific information by changing support identity within P-MEGA, but does not compare against the deterministic participant variant or LMM-VAE under the same target-excluded support protocol.
  4. The reliability of the probabilistic component's gain is unclear. Table 2 reports lower MSE than the deterministic variant in both cohorts, but only provides point estimates; uncertainty estimates for these differences are needed to assess how consistently probabilistic modeling improves reconstruction.
  5. Individual fingerprint reproducibility is not directly demonstrated. Section 4.6 reports stability of network-pair rankings of across-participant variability, which does not establish that the same participant's projected effect remains stable across model fits or support subsets.
  6. The reconstruction advantage is uneven across cohorts. Table 1 shows a 6.2% MSE improvement over the strongest HCP baseline, but only 0.6% MSE and 0.003 correlation improvements over Graph-AE on PNC, limiting the demonstrated benefit in the second cohort.
Questions for authors
  1. How does P-MEGA compare with the deterministic participant variant and LMM-VAE using identical target-excluded support sets?
  2. How consistent are the Table 2 gains over deterministic and direct high-dimensional effects across seeds and folds, with paired uncertainty estimates?
  3. How stable is each participant's projected effect across training seeds and disjoint support subsets, beyond the aggregate variability-map agreement?
  4. Is the small PNC advantage over Graph-AE consistent across paradigms, and what are its paired participant-level confidence intervals?
  5. Are the cognitive gains from adding the participant code reliable under paired comparisons, and do they persist with frozen encoders?
5 comments
AuthorsOpenReview ↗

Response to Reviewer WYeg (1/4)

Thank you very much for the review and for framing each concern as a concrete check. All five questions name a comparison or an uncertainty estimate that can be computed, and we have computed every one of them. In brief: under identical target-excluded support P-MEGA stays ahead of LMM-VAE and the deterministic variant, every paired interval for the ablation gains excludes zero, individual effects are reproducible across seeds and disjoint data, and the PNC advantage over Graph-AE has the same sign in all three paradigms.

W1. Participant effect and the population pathway

The population pathway is scan-conditioned and can carry participant information, as Section 3.4 notes. Three results characterize what the participant effect holds in addition.

It adds information beyond the population pathway. In the support swap the target scan and its population prediction are fixed, so whatever that pathway retains is identical in every condition. Same-participant support still explains the target residual better than matched or random support in both cohorts (Table 3; recomputed in Table R6, in our response to Reviewer sYH3).

It is not tied to the scan it helps reconstruct. Inferred only from the participant's other paradigms, the effect recovers nearly all of the reconstruction gain: excluding the target changes MSE by 0.3% in both cohorts (Table R1).

It is stable for the individual. The same participant's projected effect correlates at 0.714 (HCP) and 0.682 (PNC) across independently trained models, against 0.014 and 0.010 between participants, and it agrees more within than between people when inferred from disjoint paradigm subsets (Table R3).

W2. Fixed support count and the log K input

Correct. Training uses each participant's full scan set, so the support count is constant within a cohort (9 in HCP, 3 in PNC), and the count coefficient in Eq. (7) is identified only at that count, as Appendix H.5 states.

Table 1 is evaluated at the training count. The target-excluded rows of Table R1 are evaluated one count below it (8 support scans in HCP, 2 in PNC), and MSE changes by 0.3% in both cohorts relative to the complete condition. Table 3 averages over support counts starting from one. The support-swap comparison in Table 3 holds support count and paradigm composition equal across true, matched and random support, so the log K input is identical in the conditions being compared and the contrast depends only on whose scans are encoded. The evidence vector itself is a mean over the support scans and is defined at any count.

Table R8 reports the target-excluded MSE of the current model at each support count. The error decreases monotonically with the support count, from 0.02408 with one support scan to 0.01991 with eight in HCP, with most of the change reached by four (0.02047), and from 0.04106 with one to 0.03849 with two in PNC. The exact posterior of Appendix H.3 is defined at every support count without a learned count coefficient; with the trained model, its target-excluded error falls monotonically from 0.02043 at K = 1 to 0.01936 at K = 8 in HCP and is below the amortized error at every count (Table R2b, in our response to Reviewer NZMY).

Table R8. Target-excluded MSE of P-MEGA by support count K, averaged over the support subsets used in Table 3 (mean over three training seeds). The K = 8 (HCP) and K = 2 (PNC) entries are the target-excluded rows of Table R1, shown with their seed SD.

CohortK = 1K = 2K = 3K = 4K = 5K = 6K = 7K = 8
HCP0.024080.021820.020930.020470.020210.020050.019960.01991 ± 0.00008
PNC0.041060.03849 ± 0.00007——————
AuthorsOpenReview ↗

Response to Reviewer WYeg (2/4)

W3 / Q1. Comparison under identical target-excluded support sets

Table R1 reports this comparison. In every target-excluded row the participant term is computed from the same support set: all of the participant's other paradigms (8 support scans in HCP, 2 in PNC). For LMM-VAE the participant intercept is the average of the random-effect encoder's posterior means over those scans.

Table R1. Five-fold OOF reconstruction (mean ± SD over three training seeds). The paradigm template and General FC are parameter-free and are reported without a seed SD.

Method (what the participant term sees)HCP MSE ↓HCP r ↑PNC MSE ↓PNC r ↑
Paradigm template (no participant term)0.030750.6950.065310.502
Graph-AE (no participant term)0.02501 ± 0.001350.698 ± 0.0020.03858 ± 0.000120.596 ± 0.001
Graph-VAE (no participant term)0.03268 ± 0.001150.668 ± 0.0020.04178 ± 0.000240.561 ± 0.001
LMM-VAE, as in Table 1 (all scans incl. target)0.02116 ± 0.000450.705 ± 0.0060.04139 ± 0.001280.560 ± 0.007
LMM-VAE, target-excluded0.02107 ± 0.000590.706 ± 0.0080.04415 ± 0.001330.522 ± 0.008
General FC (other scans only, target-blind)0.027270.7210.058040.582
P-MEGA, population only0.02972 ± 0.000090.563 ± 0.0070.04847 ± 0.000330.407 ± 0.005
P-MEGA deterministic variant, target-excluded0.02212 ± 0.000170.712 ± 0.0060.04119 ± 0.000110.542 ± 0.003
P-MEGA, target-excluded0.01991 ± 0.000080.727 ± 0.0040.03849 ± 0.000070.597 ± 0.002
P-MEGA, complete (Table 1)0.01985 ± 0.000080.728 ± 0.0010.03836 ± 0.000050.599 ± 0.001

Against LMM-VAE. Under identical target-excluded support, P-MEGA is 5.5% below LMM-VAE in MSE in HCP and 12.8% below it in PNC, with higher edge-wise correlation in both cohorts (0.727 against 0.706; 0.597 against 0.522).

Against the deterministic participant variant. With the deterministic variant's participant code computed from the same support scans, P-MEGA is 10.0% below it in MSE in HCP and 6.6% below it in PNC, with higher correlation in both cohorts (0.727 against 0.712; 0.597 against 0.542). Excluding the target raises the deterministic variant's MSE by 6.4% in HCP and 5.5% in PNC relative to its complete condition (Table R4), against 0.3% for P-MEGA in both cohorts.

W4 / Q2. Paired uncertainty for the Table 2 gains

Table R4 reports the deterministic, direct high-dimensional and no-participant-random-effect variants retrained with three seeds on the same five folds as the main model, together with paired participant-level differences against P-MEGA. These values replace the corresponding rows of Table 2. Table 2 reported a single training seed for each variant and for a separately trained full-model reference; Table R4 reports means over the three main-model seeds and uses the main model of Table 1 as the reference, which is why both the variant and the reference values differ from Table 2.

Table R4. MSE in the complete condition (mean ± SD over three training seeds) and paired difference (variant minus P-MEGA) with 95% participant-bootstrap CI. Positive differences favor P-MEGA.

VariantHCP MSE ± SDHCP Δ [95% CI]PNC MSE ± SDPNC Δ [95% CI]
P-MEGA0.01985 ± 0.00008—0.03836 ± 0.00005—
Deterministic participant effect0.02079 ± 0.00013+0.00094 [0.00083, 0.00101]0.03906 ± 0.00011+0.00070 [0.00059, 0.00083]
Direct high-dimensional effect0.02054 ± 0.00010+0.00070 [0.00062, 0.00075]0.03897 ± 0.00007+0.00061 [0.00052, 0.00069]
No participant random effect0.02603 ± 0.00043+0.00618 [0.00602, 0.00637]0.03996 ± 0.00014+0.00160 [0.00145, 0.00172]

P-MEGA has lower MSE than all three variants in both cohorts, and all six intervals exclude zero. Across seeds, every difference is at least six times the larger of the two seed SDs. Across folds, each difference has the same sign in all five folds. The static-adjacency and dynamic-focused variants were retrained in the same way (MSE 0.02102 ± 0.00027 and 0.03533 ± 0.00016 in HCP; 0.03934 ± 0.00009 and 0.04508 ± 0.00013 in PNC), so every row of the revised Table 2 is a three-seed mean against the main model, and the six models are ranked identically in both cohorts.

AuthorsOpenReview ↗

Response to Reviewer WYeg (3/4)

W5 / Q3. Stability of each participant's projected effect

Table R3 reports both analyses: across training seeds and across disjoint support subsets.

Table R3. Individual-level reproducibility of the projected participant effect.

MeasureHCPPNC
Across seeds (same participants, same split, support fixed): within-person r / between-person r, centred BμsB\mu_s0.714 / 0.0140.682 / 0.010
Network-pair ICC(2,1) across seeds, median [quartiles]0.752 [0.715, 0.780]0.769 [0.734, 0.799]
Disjoint paradigm subsets (same model): within-person r / between-person r, centred BμsB\mu_s0.552 / 0.0020.386 / 0.001

Across training seeds. Within each outer fold, the held-out participants and their support (all of the participant's scans) are fixed, and the three training seeds give three fits. For each fit we compute BμsB\mu_s in connectome space and subtract the mean effect of that fold's training participants. Within-person r is the Pearson correlation across edges between the same participant's centred effects under two seeds; between-person r is the same correlation for different participants. Values are averaged over 15 comparisons (5 folds × 3 seed pairs). A participant's effect correlates at 0.714 (HCP) and 0.682 (PNC) between independently trained models, while the between-person value is near zero (0.014 and 0.010).

At the network level. ICC(2,1) (two-way random effects, absolute agreement, single measure; Shrout & Fleiss, 1979) is computed with participants as targets and the three training seeds as raters. For each participant and each network pair, the summary is the mean absolute value of the centred effect over the edges in that block; within-network blocks are included, giving 66 pairs in HCP and 105 in PNC. ICCs are computed within each fold on its held-out participants; we report the median over network pairs, averaged over the five folds. The median is 0.752 in HCP and 0.769 in PNC.

Across disjoint support subsets. With the model fixed, a participant's paradigms are split into two disjoint subsets and the effect is inferred from each (11 randomly sampled splits per model in HCP, 165 comparisons; 3 splits per model in PNC, 45 comparisons). Effects are centred as in the across-seed analysis, by subtracting the mean effect of the fold's training participants inferred from the same paradigm subset. Averaged over all splits, a participant's effect agrees across disjoint data (0.552 in HCP, 0.386 in PNC), while the between-person value is near zero (0.002 and 0.001).

Table R3 will be added to Section 4.6.

W6 / Q4. PNC advantage over Graph-AE by paradigm

Yes. Table R5 gives paired participant-level differences between complete P-MEGA and Graph-AE in PNC for each paradigm.

Table R5. P-MEGA (complete) minus Graph-AE in PNC. Negative ΔMSE and positive Δr favor P-MEGA. Intervals are 95% participant-bootstrap CIs; p is from the participant-level permutation test with 1,000 permutations.

ParadigmΔMSE [95% CI]pΔr [95% CI]
Emotion identification−0.00021 [−0.00033, −0.00009]≤ 0.001+0.0030 [+0.0022, +0.0039]
n-back−0.00017 [−0.00027, −0.00007]≤ 0.001+0.0029 [+0.0020, +0.0038]
Rest−0.00027 [−0.00042, −0.00012]≤ 0.001+0.0028 [+0.0019, +0.0038]
All paradigms−0.00021 [−0.00030, −0.00014]≤ 0.001+0.0029 [+0.0023, +0.0035]

The advantage has the same sign and a similar size in all three paradigms, and every interval excludes zero for both MSE and correlation. Table R5 will be added to Appendix D alongside the per-paradigm results.

In PNC, the advantage over Graph-AE extends beyond reconstruction. P-MEGA's scan representation uu gives the highest sex AUROC among the PNC models (0.724, against 0.679 for Graph-AE), and adding the participant code gives the lowest reading-achievement RMSE (0.960 for [u∥μs][u \parallel \mu_s], against 0.976; Table 4). The participant pathway carries person-specific information in this cohort: same-participant support explains the held-out target residual better than a matched donor by 0.0346 in residual r [0.0325, 0.0371] (Table R6, in our response to Reviewer sYH3), and adding μs\mu_s to uu lowers reading-achievement RMSE under both fine-tuned and frozen encoders (Table R7).

Reviewer_NZMYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

P-MEGA models functional connectomes through a population autoencoding pathway and an additive low-rank participant effect shared across paradigms. A permutation-invariant encoder infers a Gaussian distribution over this effect from support scans, enabling inference for participants unseen during training. The target scan enters the population pathway, while held-out-support evaluation excludes it from the participant pathway. Experiments on HCP and PNC examine reconstruction, support identity, support size, phenotypic relevance, and the network organization of the two components. Analytical appendices characterize the conditional model and its posterior.

Critical strengths
  1. The support-swap control directly tests the central individualization claim. Table 3 holds the target and population prediction fixed while changing whose scans supply the participant effect. Same-participant support improves residual correlation over demographically matched support in both cohorts, with reported intervals excluding zero. This is consequential evidence of transferable participant-specific information beyond a generic residual correction.

  2. The repeated-measures structure is explicit and interpretable in connectome space. The model pools support across paradigms through a shared participant effect and provides an inference procedure for new participants. Retrained ablations in Table 2 and the analytical conditional model help establish the role of this component. This combination makes the contribution more informative than an unrestricted autoencoder with an additional latent vector.

Critical weaknesses
  1. The reproducibility results establish stable population organization more directly than reliable individual fingerprints. Section 4.6 reports agreement across seeds and folds in network-pair rankings of across-participant variability. These rankings can be stable without each participant’s projected effect being reproducible across fits. The support-swap results establish useful individual information, but address transfer to target residuals rather than reliability of the estimated fingerprint itself. This distinction matters for the concluding claim that the resulting effects are reproducible across fits. The evidence supports an individualized residual representation more clearly than a stable measurement of each person’s trait.

  2. The justification and uncertainty properties of the amortized posterior need further explanation. Appendix H.3 derives a tractable conditional Gaussian posterior for a fixed global model. The learned diagonal posterior is therefore an approximation to an available reference, but its practical benefit and approximation error are not sufficiently characterized. This is especially relevant when the support count changes: Appendix H.5 correctly notes that the count effect is underdetermined if training uses a fixed count. Reconstruction gains or narrower posterior outputs do not alone establish posterior accuracy or calibration across support sizes. The analytical solution does not make the proposed framework redundant, but it provides a concrete reference for understanding what the learned inference procedure contributes and where it may be unreliable.

Questions for authors
  1. Do the same individuals receive reproducible fingerprints, beyond stable population variance maps? Section 4.6 reports cross-fit agreement in network-pair rankings of across-participant variability. Those statistics would remain unchanged if participant identities were permuted separately within each fit, so they cannot establish individual-level reproducibility. Table 3 supports useful identity information within a fitted model, but addresses a different question. For the same held-out participants in a fixed outer split, how consistent are their projected posterior-mean effects, BμsB\mu_s, across training seeds when the support scans are held fixed? Please identify a matched-participant analysis in connectome space that compares within-person agreement across fits with between-person similarity. This avoids the latent-coordinate rotation issue in H.5 and directly tests the concluding claim that the individual effects are reproducible across fits. If the evidence concerns only population variance rankings, please restrict that claim accordingly.

  2. What does the learned posterior add to the available conditional solution, and what does its uncertainty represent? Appendix H.3 already provides closed-form inference for unseen participants and varying support counts once the global model is fixed. What practical reason motivates learning this update instead? There is also a specific uncertainty issue: for the same support count, Equation 19 gives the same posterior covariance for all participants, and Equation 23 provides the mean-field variance reference. Equation 7, however, allows the scale head to depend on each participant's scan summaries. Does the learned variance vary across participants at a fixed support count? If so, is this approximation error relative to the stated conditional model, or is a different uncertainty model intended? Please relate the learned posterior mean and variance to these analytical references, and explain what supports its behavior at other support counts given the limitation acknowledged in H.5. A decrease in average posterior variance alone does not establish calibration.

5 comments
AuthorsOpenReview ↗

Response to Reviewer NZMY (1/3)

We sincerely thank the reviewer for a thoughtful and encouraging review. We are glad that the support swap reads as a direct test of the central claim and that the repeated-measures structure is seen as interpretable in connectome space. We are particularly grateful for the two questions, which go to the heart of what the paper claims. The first made clear that our reproducibility evidence was population-level and set out exactly the analysis needed at the individual level; we ran it as described (W1). The second led us to run the conditional solution of Appendix H.3 on held-out participants, which clarified what the learned posterior contributes and changed how the revision presents inference (W2). The paper is stronger for both.

W1 / Q1. Individual-level reproducibility of the fingerprint

Yes. We ran the matched-participant analysis you describe (Table R3). The Section 4.6 statistics are population-level, as you note; Table R3 is the individual-level counterpart and will be added to Section 4.6.

Table R3. Individual-level reproducibility of the projected participant effect.

MeasureHCPPNC
Across seeds (same participants, same split, support fixed): within-person r / between-person r, centred BμsB\mu_s0.714 / 0.0140.682 / 0.010
Network-pair ICC(2,1) across seeds, median [quartiles]0.752 [0.715, 0.780]0.769 [0.734, 0.799]
Disjoint paradigm subsets (same model): within-person r / between-person r, centred BμsB\mu_s0.552 / 0.0020.386 / 0.001

Each participant's fingerprint agrees across independently trained models. Within each outer fold, the held-out participants and their support (all of the participant's scans) are fixed, and the three training seeds give three fits. For each fit we compute BμsB\mu_s in connectome space and subtract the mean effect of that fold's training participants. Within-person r is the Pearson correlation across edges between the same participant's centred effects under two seeds; between-person r is the same correlation for different participants. Values are averaged over 15 comparisons (5 folds × 3 seed pairs), with 173 held-out participants per fold in HCP and about 254 in PNC. A participant's effect correlates at 0.714 (HCP) and 0.682 (PNC) between independently trained models, while the between-person value is near zero (0.014 and 0.010). Unlike the variance rankings, these statistics depend on participant identity: permuting identities within a fit reduces the within-person value to the between-person one. The comparison is made in connectome space, so the latent rotations of Appendix H.5 do not enter.

The agreement holds at the network level. ICC(2,1) (two-way random effects, absolute agreement, single measure; Shrout & Fleiss, 1979) is computed with participants as targets and the three training seeds as raters. For each participant and each network pair, the summary is the mean absolute value of the centred effect over the edges in that block; within-network blocks are included, giving 66 pairs in HCP (Shen-268) and 105 in PNC (Power-264). ICCs are computed within each fold on its held-out participants. We report the median over network pairs, averaged over the five folds; quartiles are averaged in the same way. The median is 0.752 in HCP and 0.769 in PNC.

It also holds across disjoint data from the same person. With the model fixed, a participant's paradigms are split into two disjoint subsets and the effect is inferred from each (11 randomly sampled splits per model in HCP, 165 comparisons; 3 splits per model in PNC, 45 comparisons). Effects are centred as in the across-seed analysis, by subtracting the mean effect of the fold's training participants inferred from the same paradigm subset. Averaged over all splits, a participant's effect agrees across disjoint data (0.552 in HCP, 0.386 in PNC), while the between-person value is near zero (0.002 and 0.001).

In the revision, the abstract, contribution (iii) and the conclusion state that individual effects are reproducible across training seeds and agree more within than between participants across disjoint support subsets, with Table R3 as the evidence; reproducibility across outer folds is stated only for the population-level variability maps of Section 4.6.

AuthorsOpenReview ↗

Response to Reviewer NZMY (2/3)

W2 / Q2. The learned posterior and the conditional solution of Appendix H.3

We ran the conditional solution on held-out participants. With the trained global parameters fixed, Eqs. (19)–(20) give the exact posterior from each participant's own support scans, and we compare it with the amortized posterior of Eq. (7) on the same targets and support sets.

What the learned posterior adds. At test time, the conditional solution is the better estimator for the additive Gaussian effect of Eq. (1). It gives lower held-out error in both cohorts and both support conditions (Table R2a) and at every support count (Table R2b), and a larger participant-specific effect in the support swap: the true-versus-matched difference in residual correlation is +0.0390 in HCP and +0.0808 in PNC, against +0.0158 and +0.0346 with the amortized posterior (Table R6, in our response to Reviewer sYH3). The learned posterior is the variational distribution in the training objective of Eq. (9); it does not rely on the conjugate structure, so the same training procedure applies when the participant effect enters nonlinearly (Section 5). Once the global parameters are fixed, the conditional solution is available, and the revision uses it as the reference and reports both inference modes.

Table R2a. Held-out reconstruction MSE with the amortized posterior (Eq. 7) and with the exact conditional posterior (Eqs. 19–20), from the same trained models (five-fold OOF, mean ± SD over three training seeds).

CohortSupportAmortized posteriorExact posterior
HCPTarget-excluded (K = 8)0.01991 ± 0.000080.01936 ± 0.00005
HCPComplete (K = 9)0.01985 ± 0.000080.01907 ± 0.00005
PNCTarget-excluded (K = 2)0.03849 ± 0.000070.03806 ± 0.00011
PNCComplete (K = 3)0.03836 ± 0.000050.03585 ± 0.00006

Table R2b. HCP, target-excluded, by support count K (mean over three training seeds; the amortized MSE column repeats Table R8, in our response to Reviewer WYeg). The two variance ratios compare the amortized total posterior variance ∑kσsk2\sum_k \sigma_{sk}^2, averaged over participants, with the exact total tr(V∗)\mathrm{tr}(V^{\ast}) of Eq. (19) and the mean-field total of Eq. (23); both references are the same for every participant at a given K. CV is the across-participant coefficient of variation of the learned total variance, which is zero under Eq. (19). Coverage is the per-edge fraction of held-out target values inside the central 50% / 90% predictive interval.

KMSE, amortizedMSE, exactLearned / exact varianceLearned / mean-field varianceCVCoverage 50% / 90%, amortizedCoverage 50% / 90%, exact
10.024080.020434.54714.0780.2530.508 / 0.8810.536 / 0.899
20.021820.019822.3557.3160.2130.525 / 0.8910.543 / 0.903
30.020930.019611.5954.9640.1950.533 / 0.8960.546 / 0.905
40.020470.019511.2093.7630.1970.538 / 0.8990.547 / 0.905
50.020210.019450.9753.0350.1860.541 / 0.9010.548 / 0.906
60.020050.019410.8182.5470.1860.543 / 0.9020.549 / 0.906
70.019960.019380.7052.1960.1860.544 / 0.9020.549 / 0.906
80.019910.019360.6201.9330.1880.545 / 0.9020.549 / 0.906
AuthorsOpenReview ↗

Response to Reviewer NZMY (3/3)

The learned variance does vary across participants at a fixed support count. Its coefficient of variation is 0.19–0.25 at every K, whereas Eq. (19) gives one covariance for everyone. This variation is approximation error rather than an intended uncertainty model: a participant's learned variance does not track that participant's held-out error (HCP, K = 8: Spearman ρ = −0.20, 95% CI [−0.26, −0.13]).

Relation to Eqs. (19) and (23). The learned total variance exceeds the mean-field reference at every K (14.1 times at K = 1, 1.9 times at K = 8) and crosses the exact total between K = 4 and K = 5, from 4.5 times tr(V∗)\mathrm{tr}(V^{\ast}) at K = 1 to 0.62 times at K = 8. It therefore falls faster with support count than the exact posterior, which is the behaviour Appendix H.5 anticipated when the count coefficient is identified only at the training count. The learned mean has nearly the same scale as the exact one (norm ratio ∥Bμs∥/∥Bμs∗∥=0.95\Vert B\mu_s \Vert / \Vert B\mu_s^{\ast} \Vert = 0.95 at K = 8) and is less accurate (Tables R2a–b).

Other support counts and calibration. The exact posterior is defined at every K without a learned count coefficient. With the trained model, its error decreases monotonically from K = 1 to K = 8 and is below the amortized error at every K (Table R2b). Per-edge predictive coverage is within five percentage points of nominal for both posteriors, but these intervals also contain the observation noise σx2\sigma_x^2 and do not isolate the posterior width. We therefore do not interpret the learned posterior widths and make no calibration claim for them.

In the revision: (i) Section 3.5 gives Eqs. (19)–(20) as exact test-time inference for the additive Gaussian effect and describes the amortized posterior as the variational distribution used in training, and contribution (i) is worded accordingly; (ii) Table 1 reports both inference modes, and Section 4.4 adds the true-versus-matched contrast under exact inference; (iii) Tables R2a–b are added to Appendix H; (iv) the statement in Section 4.4 that the mean posterior variance contracts with support is removed, and Appendix H.5 refers to Table R2b instead of deferring calibration.

Reviewer_sYH3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper treats repeated fMRI scans of one participant as a mixed-effects problem, splitting each connectome into a scan-level population prediction and a low-rank participant effect shared across that participant's scans. A permutation-invariant set encoder gives an amortized posterior over that effect from any subset of support scans, so it can be inferred for participants never seen in training. The claim is that same-participant support provides individual-specific information beyond a correction shared across participants, and the validation is an identity-controlled support swap: the effect is inferred from the target's own other scans, from a demographically matched participant, or from a random one, with support count and paradigm fixed. A further analysis reports it reproducible across seeds and folds and related to phenotype.

Critical strengths

The support swap is the right experiment for the claim. Holding support count and paradigm fixed while changing donor identity isolates what is specific to the participant, and including both a demographically matched and a random donor separates the individual contribution from a correction that transfers across people. The evaluation around it respects the repeated-measures structure: all scans from a participant stay in one outer fold, the intervals are participant-level bootstraps, and the comparisons are FDR-corrected in both cohorts.

Critical weaknesses

No table compares a target-excluded P-MEGA against any baseline. Table 1 carries the reconstruction claim and reports the complete condition, which the paper says uses the posterior mean from all available scans of the participant, adding that this is an autoencoding analysis so the target scan may contribute to it. Two of the baselines carry scan-level codes and so have no participant-level term, and for LMM-VAE, which the paper calls its closest mixed-effects comparator, what the random effect receives at inference is not stated. Table 3, which does exclude the target, contains only other P-MEGA conditions. The population-only condition is also lower on edge-wise Pearson than the parameter-free paradigm template, 0.563 against 0.695 on HCP, though it is ahead on MSE; since it sets the participant effect to zero in the trained model rather than being a separately trained population model, it does not settle what such a model would reach. Its node features already include a row of the target scan's own correlation matrix. What the comparison with the scan-level baselines does not isolate is the participant offset itself, which in the complete condition has seen the target scan.

The estimator that offset approximates is derived in the paper and never run. Appendix H.3 gives the exact conditional posterior mean as a direction-wise shrinkage of the participant's mean population residual, and Appendix H says the closed forms serve as references for interpretation while the implementation uses the amortized posterior. Given the fitted model's loadings and noise scale, that closed form needs no inference network and no per-participant optimization, and it is computable for held-out participants from their own support scans. Running it would give both a baseline the method has to beat and a direct reading of what amortization costs. The related work also points at closer comparators than the ones evaluated, pooling a participant's rest and task scans into one individual estimate, and the set-conditioned family of neural statisticians and neural processes, and neither is run.

The share of the support-swap gain that is individual rather than donor-transferable also differs about fourfold between cohorts without comment, roughly 6% of the same-participant reduction in HCP and 23% in PNC. The increment is real and the bootstrap intervals separate in both, but a fourfold difference in that share between a developmental cohort and a young-adult one bears on how far the HCP numbers generalize.

A comparison of the target-excluded model against any one of the baselines, together with the closed-form estimator run on held-out participants, would settle most of this for me.

Questions for authors

Q1. Is there a comparison anywhere of the target-excluded model against a baseline? Table 1 uses the complete condition and Table 3 has no baselines, so I could not find the margin that the reconstruction claim needs.

Q2. How does the amortized posterior compare with the closed-form shrinkage estimator of Appendix H.3, evaluated on held-out participants from their own support scans?

Q3. Can related participants span HCP folds? The ethics statement names family-aware analysis as a prerequisite for individual-level use, and heritable connectome structure is what an effect described as participant-specific would borrow from.

7 comments
AuthorsOpenReview ↗

Response to Reviewer sYH3 (1/4)

Thank you for a rigorous review, and for judging the support swap to be the right experiment for the claim. The review names the two analyses that would settle most of its concerns: the target-excluded model against the baselines, and the closed-form estimator of Appendix H.3 run on held-out participants. We ran both. Target-excluded P-MEGA has lower error than every baseline in both cohorts (W1), and exact inference lowers that error further and enlarges the participant-specific difference in the support swap (W2a). The remaining points are addressed in turn.

W1 / Q1. Target-excluded P-MEGA against the baselines

The submission did not contain this comparison: Table 1 used the complete condition and Table 3 compared only P-MEGA conditions. We have now run it (Table R1). In the target-excluded row the participant effect is inferred from all of the participant's other paradigms (8 support scans in HCP, 2 in PNC), and the target scan enters only the population pathway, as it does for the encoder of every learned baseline.

Table R1. Five-fold OOF reconstruction (mean ± SD over three training seeds). The paradigm template and General FC are parameter-free and are reported without a seed SD.

Method (what the participant term sees)HCP MSE ↓HCP r ↑PNC MSE ↓PNC r ↑
Paradigm template (no participant term)0.030750.6950.065310.502
Graph-AE (no participant term)0.02501 ± 0.001350.698 ± 0.0020.03858 ± 0.000120.596 ± 0.001
Graph-VAE (no participant term)0.03268 ± 0.001150.668 ± 0.0020.04178 ± 0.000240.561 ± 0.001
LMM-VAE, as in Table 1 (all scans incl. target)0.02116 ± 0.000450.705 ± 0.0060.04139 ± 0.001280.560 ± 0.007
LMM-VAE, target-excluded0.02107 ± 0.000590.706 ± 0.0080.04415 ± 0.001330.522 ± 0.008
General FC (other scans only, target-blind)0.027270.7210.058040.582
P-MEGA, population only0.02972 ± 0.000090.563 ± 0.0070.04847 ± 0.000330.407 ± 0.005
P-MEGA deterministic variant, target-excluded0.02212 ± 0.000170.712 ± 0.0060.04119 ± 0.000110.542 ± 0.003
P-MEGA, target-excluded0.01991 ± 0.000080.727 ± 0.0040.03849 ± 0.000070.597 ± 0.002
P-MEGA, complete (Table 1)0.01985 ± 0.000080.728 ± 0.0010.03836 ± 0.000050.599 ± 0.001

The margin does not come from the target scan entering the participant effect. Excluding the target from the support set changes P-MEGA's MSE by 0.3% in both cohorts (within one seed SD in HCP) and r by at most 0.002. Against the comparators that carry a participant-level term, target-excluded P-MEGA is 5.5% below target-excluded LMM-VAE in HCP and 12.8% below it in PNC, and 27.0% and 33.7% below General FC, with higher edge-wise correlation in every case. Target-excluded P-MEGA also has lower MSE than Graph-AE, the strongest baseline without a participant term: 20.4% lower in HCP and 0.2% lower in PNC.

LMM-VAE at inference. Table 1 did not state this. Our LMM-VAE uses one random intercept per participant. At inference it is the average of the random-effect encoder's posterior means over all of the test participant's scans, target included. The published rule (Simchoni & Rosset, 2024, Section 3.1) estimates each level's random effect from the random-effect encoder's outputs on that level's training observations, which do not exist for held-out participants, so we average the posterior means over the participant's own scans instead. The Table 1 row is therefore a target-included result. Table R1 adds the same rule with the target removed from the average. Target-excluded P-MEGA is below both versions in both cohorts, including the target-included one (5.9% in HCP, 7.0% in PNC).

AuthorsOpenReview ↗

Response to Reviewer sYH3 (2/4)

Population-only against the template. We agree that this condition is not a population model. It is the jointly trained model with the offset set to zero. In the revision we use it only as the fixed reference in the support swap, and Section 4.2 no longer presents the reduction relative to it as a reconstruction gain. Joint training assigns the residual shared across a participant's scans to the offset pathway (Appendix H.2), and Table 3 shows that most of this residual is common across people. Zeroing the offset therefore removes structure that a standalone model would keep, consistent with its correlation falling below the template. The separately trained counterpart is the "No participant random effect" variant, retrained with the three main-model seeds (Table R4, in our response to Reviewer WYeg): MSE 0.02603 ± 0.00043 in HCP and 0.03996 ± 0.00014 in PNC, 15.3% and 38.8% below the template. Its edge-wise r is 0.691 ± 0.004 in HCP and 0.574 ± 0.002 in PNC, against 0.563 and 0.407 for the population-only condition. In HCP this lies within the range of the other models without a participant term (0.668–0.698; template 0.695), and in PNC it exceeds the template (0.502). Target-excluded P-MEGA is 23.5% below it in MSE in HCP and 3.7% below it in PNC, with higher edge-wise r (0.727 against 0.691; 0.597 against 0.574).

Target scan in the node features. Correct. Every learned model in the table encodes the target scan through its scan-level pathway; this is autoencoding in all of those rows. General FC is the one target-blind reference. It is parameter-free: the prediction for a scan is the arithmetic mean of the participant's other Fisher-z FC vectors, with no template correction and nothing estimated from training participants.

Relation to Table 3. Table 3 averages over all support sizes, starting from a single other paradigm, so its true-support row corresponds to a higher MSE than the all-other-paradigms row above.

In the revision, Table 1 will report the target-excluded condition alongside the complete one.

AuthorsOpenReview ↗

Response to Reviewer sYH3 (3/4)

W2a / Q2. Closed-form estimator of Appendix H.3

We ran it. As you note, the closed form needs no inference network and no per-participant optimization. With the trained P-MEGA parameters fixed (DθD_\theta, B=Wb⊤B = W_b^\top and σx2\sigma_x^2), Eqs. (19)–(20) give the exact posterior of each held-out participant's effect from that participant's own support scans; the target scan enters only through Dθ(usp)D_\theta(u_{sp}), as in the target-excluded rows of Table R1. Because it uses the same trained model, the closed form is exact inference in P-MEGA rather than a separate model: it shows what the model achieves when the participant effect is inferred exactly, and how much of that the amortized encoder recovers.

Table R2a. Held-out reconstruction MSE with the amortized posterior (Eq. 7) and with the exact conditional posterior (Eqs. 19–20), from the same trained models (five-fold OOF, mean ± SD over three training seeds).

CohortSupportAmortized posteriorExact posterior
HCPTarget-excluded (K = 8)0.01991 ± 0.000080.01936 ± 0.00005
HCPComplete (K = 9)0.01985 ± 0.000080.01907 ± 0.00005
PNCTarget-excluded (K = 2)0.03849 ± 0.000070.03806 ± 0.00011
PNCComplete (K = 3)0.03836 ± 0.000050.03585 ± 0.00006

Amortization has a measurable cost. With the target excluded, exact inference lowers MSE by 2.8% in HCP and 1.1% in PNC, and with complete support by 3.9% and 6.5%. In HCP, exact inference also gives the lower error at every support count from one to eight (Table R2b, in our response to Reviewer NZMY). This is the amortization gap known from variational autoencoders (Cremer et al., 2018); because the participant effect in Eq. (1) is additive and Gaussian, exact inference removes it at test time.

The margin over the baselines widens. With exact inference and the target excluded, P-MEGA is 8.1% below target-excluded LMM-VAE in HCP and 13.8% below it in PNC, and below Graph-AE by 22.6% in HCP and 1.3% in PNC (in PNC, paired participant-level difference −0.00052, 95% CI [−0.00063, −0.00041]).

The participant-specific result is stronger under exact inference. We repeated the support swap of Table R6 with the exact posterior, keeping the targets, support sets and matched donors. The true-versus-matched difference in residual correlation is +0.0390 in HCP and +0.0808 in PNC, against +0.0158 and +0.0346 with the amortized posterior, and both intervals exclude zero. The person-specific component is therefore not produced by the inference network, and exact inference recovers more of it.

In the revision, Section 3.5 gives Eqs. (19)–(20) as exact test-time inference for the additive Gaussian effect and describes the amortized posterior as the variational distribution used in training (Eq. 9). Table 1 reports both inference modes, and Section 4.4 adds the true-versus-matched contrast under exact inference.

W2b. Closer comparators from the related work

Pooling a participant's rest and task scans. Table R1 now includes a General FC-style reference (Elliott et al., 2019). Elliott et al. compute FC on the concatenated time series after removing task structure; our version averages the participant's other Fisher-z FC vectors, so it is target-blind and parameter-free. The two rows from Table R1:

MethodHCP MSE ↓HCP r ↑PNC MSE ↓PNC r ↑
General FC (other scans only, target-blind)0.027270.7210.058040.582
P-MEGA, target-excluded0.01991 ± 0.000080.727 ± 0.0040.03849 ± 0.000070.597 ± 0.002

With the participant effect inferred from the same other scans, P-MEGA is 27.0% below General FC in MSE in HCP and 33.7% below it in PNC, with higher edge-wise correlation in both cohorts.

Set-conditioned latent-variable models. P-MEGA's inference network belongs to this family: a permutation-invariant set encoder amortizes a Gaussian posterior over a participant-level latent from an arbitrary support set (Eq. 7), as in the Neural Statistician and neural processes. The difference lies in how the latent enters the decoder. Those models condition the decoder on the latent; P-MEGA adds a low-rank effect in connectome space. The additive form is what yields the mixed-effects covariance of Appendix H.4 and the closed-form posterior of Appendix H.3, which gives the exact inference reported in W2a above, and it lets the same inferred effect be swapped between donors with the population prediction held fixed, which is the design of the support-swap test. We did not train a decoder-conditioned variant for this response, because in such a model the latent's effect on the connectome depends on the target scan and could not be held fixed across donors in this way.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of new analyses and revisions

We thank the three reviewers for careful and constructive reviews. All three single out the identity-controlled support swap as evidence for the central claim. The main requests were a target-excluded comparison with the baselines, the closed-form posterior of Appendix H.3, individual-level rather than population-level reproducibility, and paired uncertainty for several gains. The new analyses below use the same five folds and three training seeds as the main results.

Key results

  1. The reconstruction margin does not depend on the target scan (Table R1). Inferring the participant effect only from the participant's other paradigms changes P-MEGA's MSE by 0.3% in both cohorts. Target-excluded P-MEGA is below every baseline in both cohorts, including LMM-VAE under the same support (5.5% lower in HCP, 12.8% in PNC) and a target-blind General FC reference (27.0% and 33.7%).

  2. Exact inference strengthens the results (Tables R2a–b). With the trained parameters fixed, Eqs. (19)–(20) give exact inference in P-MEGA from each held-out participant's own scans. With the target excluded, it lowers held-out error relative to the amortized posterior by 2.8% in HCP and 1.1% in PNC, and it is lower at every support count. It widens the margins over the baselines (8.1% and 13.8% below LMM-VAE, 22.6% and 1.3% below Graph-AE) and raises the true-versus-matched difference in the support swap from +0.0158 to +0.0390 in HCP and from +0.0346 to +0.0808 in PNC. The learned posterior variance varies across participants at a fixed support count without tracking their error, and its scaling with support count departs from Eq. (19), so we do not interpret posterior widths.

  3. Individual effects are reproducible (Table R3). For the same held-out participants, a participant's projected effect correlates at 0.714 (HCP) and 0.682 (PNC) between independently trained models, against 0.014 and 0.010 between participants; network-level ICC(2,1) medians are 0.752 and 0.769. Inferred from disjoint paradigm subsets, it agrees more within than between participants (0.552 and 0.386, against 0.002 and 0.001).

  4. The support swap holds per seed in both cohorts (Table R6). Recomputed per training seed as in Table 1, the ordering of conditions is unchanged, and the 95% interval of the true-versus-matched difference excludes zero in both cohorts. The person-specific share is 5.4% in HCP and 20.0% in PNC; the larger PNC increment and the smaller PNC same-participant reduction both contribute, and the share is now reported per cohort.

  5. The remaining gains hold under paired tests (Tables R4, R5, R7, R8). With every ablation retrained on three seeds, P-MEGA has lower MSE than the deterministic, direct high-dimensional and no-random-effect variants, all paired intervals excluding zero, and the six models rank identically in both cohorts. The PNC advantage over Graph-AE has the same sign in all three paradigms, every interval excluding zero. Adding the participant code lowers cognitive-prediction error under paired comparison, also with frozen encoders. Target-excluded error decreases monotonically with support count.

Revisions. Table 1 reports the target-excluded condition and both inference modes. Section 3.5 adds exact inference (Eqs. 19–20) and describes the amortized posterior as the training-time variational distribution. Section 4.2 no longer presents the population-only reduction as a reconstruction gain. Section 4.4 uses the recomputed support swap, adds the true-versus-matched contrast under exact inference, and removes the statement on posterior-variance contraction. Every row of Table 2 is a three-seed mean, with paired intervals for the participant-effect variants. Tables R2a–b, R3 and R5 are added to the paper. The abstract, contribution (iii) and the conclusion state individual-level reproducibility with Table R3 as evidence, and fold-level reproducibility only for the population-level variability maps. HCP family structure is stated in the methods and limitations.

Where to find the new tables

TableContentIn the response to
R1Target-excluded comparison with all baselinessYH3, WYeg
R2a–bExact vs amortized posterior; posterior variance by support countsYH3 (R2a), NZMY
R3Individual-level reproducibilityNZMY, WYeg
R4Ablations on three seeds with paired intervalsWYeg
R5PNC advantage over Graph-AE by paradigmWYeg
R6Support swap recomputed per training seedsYH3
R7Cognitive prediction, paired and with frozen encodersWYeg
R8Target-excluded error by support countWYeg

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.