acceptodds
Withdrawn from ICLR 2027

CoFlo: Federated Sampling of Molecular Conformers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative models for molecular conformer generation benefit from large, diverse training datasets. However, much of the available molecular data remains siloed across institutions with different chemical specializations, where privacy constraints and intellectual property concerns hinder centralized training. To address this challenge, we introduce CoFlo, a framework for collaborative conformer generation across distributed clients without sharing their training data or generative-model parameters. Clients independently train flow matching models on their local data, while a lightweight router is trained through federated learning. At inference time, the router combines the fixed client velocity fields to guide a shared sampling trajectory, using mixture weights motivated by a posterior-weighted decomposition of the joint velocity field. We evaluate CoFlo on GEOM-QM9 and GEOM-Drugs under IID and heterogeneous scaffold-, chemotype-, topology-based, and cross-dataset partitions. CoFlo outperforms average client performance across all four metrics in every evaluated setting and surpasses even the best individual client on all four metrics in most dataset–split combinations, including the realistic mixed-dataset setting. These results demonstrate the potential of inference-time collaboration of specialized experts for conformer generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular conformer generation, federated learning, flow matching, mixture of experts, distributed inference

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_svhfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper uses a federated sampling method for molecular conformer generation. Specifically, given experts with one flow matching based conformer generation, this work trains a router (given a molecular graph, predict which expert "handles" that example), and uses the router at inference time to find the right weightings for the velocity coming from the different experts. This router is trained in a federated way, with updates happening locally for each expert and then combined through some global update.

Critical strengths

The method shows good performance when compared to naive velocity averaging or the single best expert. The soft router (velocities weighted by router probabilities) performs similarly to the top-1 router (velocity of the max probability expert is used). Overall performance appears to be good, and the router is learned correctly from the molecular graph.

Critical weaknesses

The methodological novelty is limited. In some aspects this work closely follows the DDM paper (Decentralized Diffusion Models). The same velocity averaging mechanism through a learned router was proposed there. The differences, from my understanding, are: (a) Application to molecular conformer generation; (b) the router is trained differently, with this work training it on a federated manner, with the updates from each expert being synchronized globally; and (c) the router takes as input a molecular graph (DDM takes noisy images as input to the learned router).

The evaluation and scenarios tested seem to be particularly well suited for the graph based router, since the splits considered come, as I understand, from the molecular graph. The router only taking in the graph may be disfavorable in other settings (e.g. different clients using the same graphs but different conformations). In such cases including the noisy conditioning, as in DDM, could matter. Why was this input removed? [This limitation is acknowledged in the paper.]

Questions for authors

Could this method also be used for other domains in which general labels to condition generation are available? (In this work, molecular graphs represent a strong condition, in other domains, such as protein folding, it could be the protein sequence; or outside of AI for bio, in LLMs the prompt and topic of the prompt, etc).

Reviewer_7LbtOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies molecular conformer generation using independently trained models held by different clients. CoFlo trains a graph-only router through federated learning and uses its weights to combine the fixed client velocity fields during sampling. Experiments on GEOM-QM9, GEOM-Drugs, and their mixed federation report improvements over individual clients across several data partitions, with Top-1 routing often performing similarly to the full mixture.

Critical strengths
  1. The overall framework is clearly presented. The evaluation covers several chemically heterogeneous partitions, and the uniform-averaging and Top-1 baselines help assess the role of the router.
Critical weaknesses
  1. The contribution has limited novelty. DDM [1] already derives the expert-field decomposition and combines independently trained generative models using a learned classifier. CoFlo's main addition is to condition the router only on the molecular graph and train it using FedAdam.

  2. The practical advantage of Soft-Router over Top-1 is unclear. Their COV-R differs by less than one percentage point on every mixed-dataset split (lines: 365-368), while Top-1 does sampling with a single expert, which is faster and does not justify the marginal performance boost.

[1] David McAllister, Matthew Tancik, Jiaming Song, and Angjoo Kanazawa. Decentralized Diffusion Models. CVPR, 2025.

Questions for authors
  1. How does the velocity parameterization in Eq. (5) follow from the noisy interpolant in Eq. (4) when σ>0\sigma>0? For fixed τ<1\tau<1, Eq. (4) gives

    = X_1-X_0-\frac{\sigma Z}{1-\tau}.$$ Under squared-error clean-coordinate regression, the ideal predictor is $f^*(x,\tau,G)=\mathbb E[X_1\mid X_\tau=x,G]$. Consequently, $$\frac{f^*(x,\tau,G)-x}{1-\tau} = \mathbb E[X_1-X_0\mid X_\tau=x,G] -\frac{\sigma}{1-\tau}\mathbb E[Z\mid X_\tau=x,G],$$ where the first term on the right-hand side is the marginal velocity. The final conditional expectation need not vanish, even though $\mathbb E[Z]=0$. Is the additional noise intended only as training augmentation, or is Eq. (5) intended to recover this marginal velocity?
  2. Table 3 includes centralized baselines, but the authors acknowledge that their configurations differ. Could the authors specify these differences and compare the methods using the same architecture and comparable total training compute, including a centralized model trained on the mixed union? This would clarify the accuracy cost of decentralization.

  3. Why do the Drugs Topology results differ so substantially between Figure 3(a) and Table 3?

Reviewer_K621OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents two methods for federated learning of conformer geometries. Each organization uses its private data to train a flow matching model that predicts the molecular conformers it knows. In the simple baseline, these flow models are averaged, presumably making a better predictor overall than any of the separate models. In the method advocated by the authors, CoFlo, an additional model is trained that takes the 2-d molecular graph as input and uses that to create a weighted-average of the separate flow models, presumably weighted more towards the model trained on more similar molecules. As expected, CoFlo gives more accurate conformers than a flat unweighted average flow field. The key problem behind this study is that pharma and other companies do not want to share their labeled data, and even more they do not want to share their unlabeled data (the list of molecules they are studying), since this would give clues/suggestions to their competitors. As the authors note, it is very questionable how well CoFlo protects this proprietary information.

Critical strengths

The CoFlo method appears to be more accurate than the simpler federated approach of just averaging all the flow fields.

Averaging (or weighted averaging) of flow fields is a very reasonable way to do the averaging (it is often not sensible to average the actual geometries predicted by each sub-model, since the arithmetic average of geometry1 and geometry2 is often a non-physical molecular geometry.)

Critical weaknesses
  1. Most importantly, it is not apparent how CoFlo could ensure the privacy of each organization's set of molecules. The model p(k|G) which gives the weights in the weighted average tells which organization's model is best for predicting molecule G. If the model p(k|G) is close to 1, it is saying that that molecule G or a close analog is a molecule being studied by organization k. This is exactly the information that a pharma company does not want to share with competitors. This flaw in the scheme is acknowledged by the authors. Once they realized this massive problem, which makes CoFlo completely unsuitable for the desired application, why did they still proceed to submit this to ICLR, and in a paper that seems to be suggesting CoFlo is a good idea?
  2. It is unclear to this reviewer which organization trains p(k|G) and how that could be done in this scheme, where each organization has its own private data and labels. But even if you could devise a guaranteed-private way to train p(k|G), one would need to keep it secret, see issue #1
Questions for authors

I think CoFlo has to go back to the drawing board. Maybe there is some clever way to hide p(k|G) inside the final model, so it would be hard/impossible for any of the organizations to deconvolve it? Note that each organization only knows the geometries of the molecules it is studying, so they cannot tell if CoFlo is giving accurate predictions for labels of other molecules. So it is not easy for them to learn much about their competitors' data from the predicted geometries. (Though they could infer that none of their competitors is studying molecule G if the predicted geometry of G was obviously unphysical.) So leakage via the final model predictions is probably not a huge concern. But p(k|G) is exactly the information that the companies want to keep secret from each other.

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

Related papers

Discussion (0)

Sign in to comment.