acceptodds
Withdrawn from ICLR 2027

ARGFlows: Amortized Posterior Sampling of Ancestral Recombination Graphs with GFlowNets

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Ancestral recombination graphs (ARGs) describe the genealogical and recombi- nation history of sampled genomes, underpinning inference of demography, selec- tion, and mutation age. Despite decades of work, posterior inference over ARGs remains challenging. The space combines a variable number of discrete events with continuous event times and is sharply multimodal because structurally dif- ferent histories can explain the same sequence data almost equally well. In this paper we adopt generative flow networks (GFlowNets) to tackle Bayesian infer- ence over ARGs. Because GFlowNets sample combinatorial structures in pro- portion to reward, they are a natural choice for a posterior whose mass is spread across many distinct histories, and where posterior sampling requires sustained exploration of alternative histories. Across three simulated datasets spanning dif- ferent recombination rates, ARGFlow achieves competitive reconstruction against ARGInfer and SINGER while representing uncertainty in coalescence times and local genealogical structure.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
GFlowNets, amortized inference, Bayesian inference, posterior sampling, ancestral recombination graphs, population genetics, computational biology, coalescent theory, probabilistic inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UT6eOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a framework for Bayesian inference over ancestral recombination graphs (ARGs) with generative flow networks (GFlowNets). The model samples ARGs using an acyclic sequential decision process over backward-time coalescence and recombination events. The state retains the chronological history of ancestral events constructed so far, the current backward time and the set of active ancestral lineages. The policy is parameterized by a transformer model and at each step predicts first the event type, then the event parameters (either a pair of lineages to coalesce or a lineage and breakpoint for recombination), and finally the waiting time to the next ancestral event. The GFlowNet reward is the unnormalized Bayesian posterior probability, and the model is trained via Subtrajectory Balance loss. The presented approach achieves competitive reconstruction against ARGInfer and SINGER baselines on simulated datasets.

Critical strengths

The paper presents a methodologically solid and straightforward application of the GFlowNet framework to a new domain of problems, namely sampling ARGs from Bayesian posteriors. I have almost no prior knowledge on the genomic domain and ARGs themselves, but I checked all parts related to the application of GFlowNets, namely the environment construction, the model parameterization and the training objective, and everything looks solid to me. I appreciated the detailed derivation and discussion of all methodological details in Appendix that are needed to implement the approach.

The novelty is clearly there, and the application itself bears scientific importance, as no prior works developed amortized sampling approaches for Bayesian posteriors over ARGs.

I also appreciated the honest reporting of numerical results, faithfully showing where the approach falls behind the baselines. And, despite this, I find the experimental results solid as the paper presents the first amortized learning-based approach, compared to prior MCMC-based methods. The only minor issue I want to point out is the choice to highlight in bold the results of the proposed method in Table 2, rather than conventionally highlighting the best results, which can mislead some readers in my opinion.

Critical weaknesses

The paper is missing some important references. The source paper for one of the baselines, namely ARGInfer, is not cited [1]. The paper that developed methodology for GFlowNets for working with continuous variables is not cited [2]. In addition, there is another prior work on sampling discrete graphs together with continuous parameters with GFlowNets [3]. There is also a work that applies deep reinforcement learning to generate ARGs [4], while GFlowNets themselves are known to be equivalent to specific regularized RL methods [5], so I think it is worth a small discussion.

The method is evaluated only on simulated data, while prior works, e.g. SINGER, are also evaluated on real-world genomic data.

While the paper generally has good writing, I would suggest adding a section to Appendix explaining the background on the mathematical model for ARGs. Since I had no prior knowledge on them, some specifics of the task and the methodology were very hard to grasp on the first read.

References:

[1] Mahmoudi et al. Bayesian inference of ancestral recombination graphs. PLOS Computational Biology 2022.
[2] Lahlou et al. A theory of continuous generative flow networks. ICML 2023.
[3] Deleu et al. Joint Bayesian Inference of Graphical Structure and Parameters with a Single Generative Flow Network. NeurIPS 2023.
[4] Raymond et al. Constructing ancestral recombination graphs through reinforcement learning. Frontiers in Genetics 2025.
[5] Tiapkin et al. Generative Flow Networks as Entropy-Regularized RL. AISTATS 2024.

Questions for authors

I think that this is a borderline yet solid paper, so I recommend acceptance, given that my concerns related to presentation and references would be fixed. I do not need additional clarification on any questions from the authors.

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

The authors show the application of GFlowNets to learning the posterior distribution over ancestral recombination graphs. Ancestral recombinations are multi-modal in nature with many possible graph structures that could be inferred from the data, making posterior inference to find the distribution over graphs more appropriate than finding one candidate. The formulate the reward using the hudson coalescent-with-recombination prior and the sequence likelihood, and show empirically that their method achieves enhanced performance across multiple metrics compared to existing methods on three simulated datasets.

Critical strengths

GFlowNets have not been applied to this domain before, the authors apply careful consideration of the constraints and rewards of learning ARGs so that GFlowNets can be used to learn them. The application of GFlowNets to this domain seems to come at an advantage over other methods for these datasets compared to other methods as seen in the empirical results section.

Critical weaknesses

The main weakness of this paper is the limited empirical results. Three simulated datasets to assess the method is limiting. If only simulated datasets can be used to assess the method, there should be a larger number of simulated datasets. If there are real datasets that can be used in this domain to assess the method, those results would greatly strengthen the empirical section. The second point relating to the weak empirical results section is that there are no error bars in the results (table 2 for example). Given the weak empirical section as a result of the lack of error bars and limited number of datasets on which the method is tested, I dont think that the fact that the method outperforms baselines on one seed in table 2 is significant evidence to support the robustness of this method. In order for the paper to pass the acceptance threshold, the authors need to provide either a good theoretical reason to motivate the use of GFlowNets for this application, or wide-ranging empirical results to suggest the benefits of using their method over others. As a final point, since the main contribution of this paper is the setting up of the rewards and constraints of the GFlowNet to learn ARGs, I think that the paper could benefit from having the actual equation for the reward in the main text, rather than in the appendix, replacing the standard equations for the rewards for the GFlowNet in equations 9 and 10 and making it more specific to the application to ARG learning. Similarly, the subtrajectory balance equations I believe, are the standard equations introduced in the GFlowNet literature, and this space could be taken up with more details on how GFlowNets are applied to ARG learning.

Questions for authors
  • In ARG learning, are there ‘real’ datasets for which the method can be tested on?
  • What are the main conceptual differences in the approaches for learning the posterior distribution over ARGs, what makes GFlowNets a particularly good approach compared to existing approaches?
Reviewer_Gu5cOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript proposes using GFlowNets to sample ancestral recombination graphs. Authors propose a generation sequence that induces a tree-like state graph and train their model using a simplified version of the well-known subTB loss. They propose a transformer-based architecture to implement the forward policy and a factorized forward transition accounting for the option to coalesce or recombine and the discrete-continuous nature of the target object. They also present small scale (10 haplotypes) comparisons against ARGInfer and SINGER.

Critical strengths
  • The main strength of this work is exploring ancestral recombination graphs as a novel application niche for GFlowNets, which makes sense given prior works as GFlowNets for Bayesian phylogenetic inference focusing on trees.
Critical weaknesses
  • The manuscript has limited technical novelty. The manuscript seems like a fairly straightforward application of GFlowNets, and the loss is the well-known SubTB. The state encoder, which is sold as a contribution, seems more like an engineering detail. The problem-specific design elements, like the forward construction, also seem modest.

  • The authors do not offer a good rationale for why GFlowNets are the right tool here. Every state has a unique parent, so the state graph is a tree and the model is essentially an autoregressive sampler with a tractable density. I would appreciate further elaboration on why using GFlowNets and not some standard VI objective. Also, I would expect at least a wall-clock comparison against existing methods for ARG inference.

  • The manuscript could be better polished. There are missing definitions and relevant elements are left out of the main text. For instance, the marginal over terminals induced by P_F is never properly defined, and TB is mentioned but never written. There is a clear abuse of notation around Eqs. 7–8: P_F is used for the joint and for every conditional, mixing probabilities and densities. From the main text it is not even clear whether there is a Bernoulli at each step to choose between C and R, or what the parametric form of the other distributions are. F(s_t) is referred to in Sec. 4.3 before being defined.

  • It seems the evaluation metrics do not directly assess sampling correctness. As I understand, the metrics in the experiments section mostly measure error wrt ground truth. Even agreement with ARGInfer and SINGER are only proxies which rely on these methods having converged. Since the marginal over terminals for ARGFlow is tractable, metrics like flow consistency in subgraphs (FCS) seem like a more reliable option.

Questions for authors
  • Can you provide wall-clock time comparison against baselines?

  • Judging by the results in Appendix E, it seems like ARGFlow has a relevant mismatch wrt the reward. Could you elaborate further on the reason for this?

  • As far as I can tell, nothing in the construction prevents infinite trajectories unless a maximum length is imposed. If this is the case, what happens to training trajectories that hit maximum length? How does this bias the sampler?

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

Related papers

Discussion (0)

Sign in to comment.