acceptodds
Under review as a conference paper at ICLR 2027

MALM: Learning Latent Network Mediators from Paired Connectomes

Abstract

Mediation analysis begins with specifying the mediator, which defines the mechanism being tested. In brain-network mediation, structural connectivity (SC) and functional connectivity (FC) are often treated as candidate network mediators, but they are coupled, noisy, and complementary measurements of brain connectivity, making parallel or serial multimodal formulations potentially restrictive. We propose MALM, an end-to-end Bayesian framework that learns an integrated network mediator from paired SC and FC and uses that learned network directly as the mediator. Its central contribution is to couple multimodal representation learning with exposure-to-mediator and mediator-to-outcome pathway estimation on the same region-pair coordinates. Dual cross-modal attention integrates the observed networks, while a structured mediation layer connects the mediator to closed-form direct, indirect, and total effects with an edge-level decomposition, without imposing an inter-modality ordering. Joint posterior inference quantifies uncertainty over model unknowns and induced mediation quantities, including learned network mediators, pathway maps, global effects, and active mediating brain edges. On synthetic data, MALM improves effect estimation and pathway recovery when connectome views are coupled, interact non-additively, or violate a fixed serial ordering. In HCP and ABCD cohorts, MALM estimates significant connectivity-mediated effects in both cohorts and recovers a substantially more consistent mediating network than existing methods, concentrated in default-mode, visual, and somatomotor systems and supporting a mediation pathway that remains stable across developmental stages.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Causal mediation analysis, Latent network mediator, Multimodal brain connectivity, Cross-modal attention

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_68S9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MALM, a Bayesian framework for mediation analysis using paired structural (SC) and functional (FC) connectivity networks. Rather than modeling SC and FC as separate or sequential mediators, MALM learns a unified network-valued mediator through cross-modal attention and gated fusion. The method uses structured attention and sparsity regularization to identify edge-level mediation pathways. Under specified causal assumptions, the authors derive natural direct, indirect, and total effects, and use stochastic gradient Langevin dynamics (SGLD) for Bayesian uncertainty estimation. The authors evaluate MALM on seven synthetic scenarios, comparing it with four baseline methods and conducting ablation studies. The results show improved indirect-effect estimation and interval coverage, particularly when cross-modal interactions are present. The method is also applied to the HCP and ABCD cohorts, using BDNF Val66Met genotype as the exposure and cognitive performance as the outcome. The authors report negative indirect effects in both cohorts and identify mediating connections involving default-mode, visual, and somatomotor networks. The identified pathways also show greater cross-cohort overlap than those of competing methods.

Critical strengths
  1. The main contribution is learning a unified network-valued mediator from paired SC and FC networks, rather than treating them as independent or sequential mediators. This is useful when the two modalities are closely related and contain complementary information. The use of cross-modal attention, gated fusion, and edge-level pathway modeling provides a reasonable way to combine representation learning with mediation analysis.
  2. The evaluation includes seven simulation settings, four baseline methods, and ablation studies. The experiments cover different relationships between SC and FC, including cross-modal interactions and nonlinear dependencies. The analyses on HCP and ABCD further examine whether the identified pathways are consistent across cohorts. The reported cross-cohort Jaccard index of 0.87, compared with 0.53 and 0.36 for competing methods, is encouraging, although its interpretation depends on the reliability of the posterior inference and pathway-selection procedure.
Critical weaknesses
  1. MALM treats a representation learned from SC and FC as a causal mediator, but this relies on strong assumptions, particularly representation sufficiency (Assumption 1) and the identification conditions in Assumption 4. Representation sufficiency requires that SC–FC configurations mapped to the same mediator have identical potential outcomes under a fixed exposure. This is much stronger than predictive sufficiency, and the proposed training procedure does not guarantee it. The paper also does not examine whether this assumption is reasonable in practice. Moreover, the mediator is learned jointly with the exposure and outcome pathways. While Section 3 defines the causal effects using a fixed population encoder, it is unclear how this relates to the data-adaptive learning procedure or how outcome-guided representation learning affects causal inference. Proposition 3.1 derives the effect decomposition under the stated assumptions, but does not show that the learned mediator satisfies them. This distinction matters because the paper's main claim concerns causal mediation rather than prediction.

  2. Equation (14) combines the conditional density of the learned mediator with reconstruction likelihoods for SC and FC, although the mediator is deterministically defined as Mi=Φθ(Si,Fi)M_i=\Phi_\theta(S_i,F_i). Could the authors clarify whether this objective corresponds to a properly normalized likelihood under an explicit generative model, or whether it should instead be interpreted as a composite or generalized Bayesian objective? Furthermore, Appendix C reports a fixed learning rate and an independently tuned noise-scale parameter for SGLD. Please provide the exact parameter update rule and explain how the injected noise relates to the step size and the intended posterior distribution.

Questions for authors
  1. Assumption 1 requires the learned mediator to be sufficient for the potential outcome under interventions on SC and FC. How can this assumption be justified or assessed for the proposed learned representation? In particular, how does the model ensure that different SC–FC configurations mapped to the same mediator do not have different potential outcomes? Could the authors provide additional theoretical conditions or other evidence showing that the learned representation preserves the required causal structure?
  2. Equation (14) combines the mediator density with SC/FC reconstruction likelihoods, although the mediator is deterministically defined as Mi=Φθ(Si,Fi)M_i=\Phi_\theta(S_i,F_i). Could the authors clarify whether this defines a valid joint likelihood under an explicit generative model or should be viewed as a composite or generalized Bayesian objective? Appendix C describes SGLD with a fixed learning rate and a separately tuned noise scale. Could the authors provide the exact update rule, explain how the noise scale relates to the target posterior, and clarify how burn-in, sample collection, and convergence were assessed?
  3. MALM reports significant indirect effects in both HCP and ABCD, unlike the compared methods. Could the authors provide further evidence that this difference is not mainly due to underestimated posterior uncertainty, greater model capacity, or regularization? Appendix E indicates that SC and FC are reconstructed from modality-specific projections of the shared fused node embeddings, rather than directly from the final mediator matrix Mi=HiBHi⊤M_i=H_iBH_i^\top. Can both observed connectomes be reconstructed using only the final network-valued mediator? If not, how does the auxiliary reconstruction objective ensure that the final mediator retains the information required for causal interpretation?
  4. MALM combines several established techniques, including cross-modal attention, gated fusion, sparsity regularization, and Bayesian inference. While the overall framework is interesting, its methodological novelty is not entirely clear. Could the authors clarify the main contributions beyond integrating existing techniques? Specifically, how does the learned mediator differ from standard multimodal representations? Does MALM introduce a new mediation estimation or identification method, or mainly apply existing mediation theory to a learned representation? What is the advantage of joint learning over a two-stage approach that learns the representation before estimating mediation effects?
Reviewer_ugj3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MALM, which jointly learns a network mediator from SC and FC to analyze the pathways through which exposure influences outcomes via brain networks. Its main contribution is integrating multimodal representation learning with mediation pathway estimation within a single framework, without needing to preset the order of the two modalities. The method preserves region-pair coordinates, and under the stated identifying assumptions and conditional-mean specifications, it can estimate natural direct effects, natural indirect effects, and total effects, and decompose indirect effects down to network edges. Joint Bayesian inference further quantifies the uncertainty in network representations, pathway coefficients, and effect estimates. Comparative experimental results indicate that MALM can improve the estimation accuracy of indirect effects in most scenarios and is particularly suitable for situations involving cross-modal interactions or nonlinear coupling. Ablation analysis supports the role of joint network mediator learning and structured pathway modeling. Real data analysis shows that both datasets exhibit significant brain network mediation effects, and the identified mediation pathways are more consistent than those found by comparison methods. These results support the estimation performance and cross-cohort stability of MALM.

Critical strengths
  1. MALM learns a network mediator from paired structural connectivity and functional connectivity, and jointly estimates the exposure-to-mediator and mediator-to-outcome pathways on the same region-pair coordinates. This allows multimodal representations to be directly involved in mediation effect estimation without presetting the serial order of the two modalities. The overall indirect effect can also be decomposed into contributions from individual network edges, providing a clear analytical framework for understanding how exposure influences outcomes through the integrated network. This design addresses practical difficulties in multimodal mediation analysis and has methodological value.
  2. Simulation experiments covered cross-modal interactions, serial relationships, and nonlinear coupling, allowing the method's performance under different relationships between modalities to be tested. Ablation results support the role of joint network mediator learning and structured pathway modeling. Pathway recovery results in two real-world cohorts were also more consistent than comparison methods. These experiments align with the research motivation of the paper and collectively support the method's estimation performance and application potential.
Critical weaknesses
  1. The paper's causal derivation assumes that the network mediator is a sufficient representation of the paired networks for the outcome under intervention. In practice, the method needs to learn this mediator from data, but it has not been adequately explained how the learning objective supports the above assumption. Reconstructing brain networks and predicting cognitive performance can show that the representation contains information, but this is insufficient to establish representation sufficiency. Therefore, existing theory explains well how to compute effects given an appropriate mediator, but the argument for whether the proposed method can learn such a mediator remains insufficient. This directly affects the causal interpretation of the learned networks and their indirect effects.
  2. A central contribution of MALM is its ability to identify interpretable brain connections that mediate the exposure–outcome relationship. However, the synthetic experiments primarily evaluate global NIE estimation through RMSE and interval coverage, while the ablation study in Table 1 uses the same two metrics. The paper does not provide quantitative evaluations of edge-level pathway recovery against known ground truth, such as edge-selection precision, recall, F1, AUPRC, or estimation error of individual mediation effects. This is a substantial omission because accurate estimation of the total indirect effect does not necessarily imply correct identification of the contributing network connections. Since the synthetic scenarios contain predefined exposure and outcome pathway parameters, these experiments could directly evaluate whether MALM recovers the actual mediating edges. Without such comparisons, one of the paper's main methodological contributions remains experimentally untested.
  3. The real-data analysis highlights a substantially higher cross-cohort Jaccard index for MALM (0.87) than for PMM (0.53) and DM (0.36). However, both cohorts use the same 68-region parcellation, and MALM constructs its pathway matrices using region-informed attention and sparsity regularization. These shared structural constraints may favor similar edge-selection patterns across cohorts, independently of whether the recovered pathways reflect reproducible mediation mechanisms. Moreover, Figure 4 visualizes the top 10% of edges, whereas the text describes Jaccard overlap between significant mediating-edge sets, leaving the relationship between these selection criteria unclear. The experiments do not examine how cross-cohort overlap changes under different selection thresholds, randomized exposure labels, or removal of the region-informed pathway structure. Importantly, the existing ablation study evaluates only NIE estimation and interval coverage, rather than pathway reproducibility. Consequently, the reported stability advantage may partly reflect model-induced selection preferences, and the main cross-cohort contribution lacks essential experimental controls.
  4. The learnable recovery policy is a central component of LBTAR, but its training objective in Equation (21) encourages the model to select channels with lower current prediction errors. This does not necessarily lead to the best recovery order. A channel that is difficult to reconstruct may still provide important information for recovering other missing channels, while selecting easily predicted channels first may contribute little to subsequent reconstruction. Moreover, the actual channel selection uses a hard Top-K operation, whereas the policy is trained through a soft, one-step error-based objective. The method therefore does not directly optimize the effect of channel selection on the final reconstruction quality. The experiments also lack comparisons with policies based on expected future error reduction or reconstruction improvement, which would help establish whether the proposed strategy learns an effective recovery order rather than simply prioritizing easier channels.
  5. Equations (22)–(23) introduce teacher forcing, self-rollout, and scheduled sampling for autoregressive channel recovery, but the experimental section does not clearly specify which strategy is used for the main results or how the scheduled sampling probability changes during training. This matters because teacher forcing writes ground-truth EEG signals into the recovery state, whereas inference must rely entirely on previously reconstructed channels. The resulting difference in available context can substantially affect error propagation across recovery stages. Although uncertainty feedback is specifically introduced to address this problem, the paper provides no direct comparison among the three rollout strategies or analysis of reconstruction error at successive stages. These missing experiments make it difficult to determine whether the reported improvements arise from the proposed uncertainty-aware recovery mechanism or from the training and rollout configuration.
Questions for authors
  1. Section 3 assumes that the network mediator is a sufficient representation of the paired networks for the outcome under intervention. Please explain how the proposed learning objective supports this assumption. Apart from reconstruction and prediction performance, is there any theoretical analysis or experimental result indicating that the learned representations can support the causal effect interpretation in the paper?
  2. How do the mediator fitting, outcome fitting, and reconstruction terms in Equation 14 jointly constitute the posterior distribution? Please clarify whether this objective comes from a complete probabilistic model or if it is defined based on a weighted loss. If it is the latter, how do the weights of each term affect the interpretation of the credible intervals?
  3. How should the noise scale and step size of SGLD be coordinated to obtain samples from the target posterior? Are there any sampling diagnostics or sensitivity analyses showing that the interval conclusions of the estimated mediation effects do not significantly change due to the choice of noise scale?
1 comment

By now, I still can not see the rebuttal. I keep my Rating.

Reviewer_k5TLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that instead of picking SC or FC as the mediator, or forcing a parallel or serial order between them, you can learn one fused brain network from both and use that as the mediator. MALM does this with cross-attention between the two connectomes, puts a linear mediation model on top, and fits everything jointly in a Bayesian way, which gives closed-form direct and indirect effects plus edge-level attributions. In simulations it beats existing single-modal, parallel, serial and deep mediation methods. On HCP and ABCD, it is the only method that finds a significant indirect effect in both cohorts, and its mediating edges agree well across the two.

Critical strengths
  1. Avoiding a forced SC→FC or FC→SC ordering by learning a joint mediator is a reasonable answer to a real limitation of multimodal mediation.
  2. The closed-form NIE and its edge-wise decomposition keep the global effect and the connection-level attribution in one coordinate system, which is convenient for interpretation.
  3. The identification section is stated carefully, including the representation-sufficiency condition the learned mediator needs.
  4. The synthetic study covers additive, interacting, serial and nonlinear SC–FC settings, and the ablations remove each main component in turn.
Critical weaknesses
  1. Φθ is trained jointly with D and W, so the mediator is chosen partly to make exposure and outcome pathways align. A significant NIE is therefore partly guaranteed by construction, and the identification argument, which assumes a fixed Φθ0, does not cover a map estimated from the same z and y. A permutation test (shuffle z or y and refit) is needed to show the false-positive rate is controlled.
  2. Eq. (14) puts a Gaussian density on Φθ(S, F), a deterministic function of the data, without any change-of-variables term, so it is not a proper likelihood. The encoder can trivially raise it by pulling M_i toward µ(x) + zD. The SGLD "credible intervals" are also not calibrated by design: noise_scale, which directly controls their width, was chosen from {0.05, 0.1, 0.2} by validation RMSE. The coverage results in Fig. 2 and Table 1 are therefore hard to interpret.
  3. The cross-cohort Jaccard of 0.87 is largely structural. D and W are built from q·k attention over region features ρ that are identical in HCP and ABCD, so similar maps across cohorts are expected regardless of the data. Fig. 4 also uses "top 10%" edges while the text says "significant" edges.
  4. The real-data analysis ignores dependence and confounding structure. HCP contains many twins and siblings, and ABCD has 21 sites and siblings, but neither family nor site is modeled or adjusted for, so intervals are likely too narrow. The units of the cognition score are not stated: if it is standardized, a total effect of about −0.4 from a single common SNP is implausibly large; if it is raw, the effect is tiny.
  5. Some synthetic results look inconsistent with the claims. In the additive LA setting, where the observed connectomes are the correctly specified mediator, the "observed mediator" variant should do best, yet it is worse than MALM. MALM's own coverage there is only 0.87, below concat fusion (0.90). The true NIE in the CMI scenarios, which contain an S⊙F term, is never defined.
Questions for authors
  1. With exposure labels permuted or outcome permuted and the full model refit, how often is the NIE "significant" in HCP and ABCD?
  2. How do NIE estimates and intervals change if Φθ is learned without the outcome likelihood and then frozen for mediation?
  3. How sensitive are interval width and coverage to noise_scale? Was it tuned in the simulations too?
  4. What is the Jaccard index if ρ is replaced by random features, or if D and W are learned as free matrices?
  5. How were HCP family structure and ABCD site handled? What are the units of the cognition outcome?
  6. In the CMI scenarios, how is the ground-truth NIE defined?
1 comment

I didn't see any rebuttal from the authors, so the concerns in my review are still open. I'll keep my original score, and I hope the comments are useful for a future revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.