acceptodds
Under review as a conference paper at ICLR 2027

TreeSBM: Tree-Valued Schrödinger Bridges for Protein Evolution Forecasting

Abstract

Protein evolution is a stochastic process in which ancestral sequences diversify into lineages that accumulate mutations and generate complex evolutionary trees. Existing protein language models, diffusion models, and flow-based generators forecast individual sequences, while phylogenetic methods reconstruct historical trees from observed descendants. The inverse problem, however, remains largely unexplored: *given a protein sequence, can we forecast the distribution of future evolutionary lineages that may emerge under selection pressures?* To answer this question, we introduce **Tree**-Valued **S**chrödinger **B**ridge **M**atching (**TreeSBM**), a generative framework for protein evolution forecasting. TreeSBM learns a tree-valued Schrödinger bridge between a reference evolutionary process and a distribution of observed evolutionary trees, transporting a root protein sequence into a distribution over future mutational histories. During inference, TreeSBM generates distributions of evolutionary trees forward in time conditioned on a single ancestral sequence, jointly modeling sequence evolution, lineage diversification, and branching topology. Across viral evolution forecasting benchmarks, TreeSBM supports realistic and high fitness evolutionary tree generation, strong future lineage forecasting, and probabilistic exploration of evolutionary spaces. On SARS-CoV-2 generated variants, TreeSBM recovers pandemic-grade threats and other observed high fitness mutations. These results establish tree-valued Schrödinger bridges as a framework for forecasting branching processes in protein evolution. For reproducibility, we host our code, data-processing scripts, and evaluation pipelines at this anonymous repository: [https://anonymous.4open.science/r/TreeSBM-D7EC](https://anonymous.4open.science/r/TreeSBM-D7EC).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein evolution forecasting, tree-valued generative models, Schrödinger bridge matching, viral evolution, phylogenetics

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 4 reviews, above 0% of the others.
Reviewer_mhtdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses prospective forecasting of protein evolution from a single ancestral sequence x0, aiming to sample future evolutionary trees T_future ~ p(T_future | x0) that jointly specify descendant sequences, branching topology, and branch lengths. The authors propose Tree-Valued Schrödinger Bridge Matching (TreeSBM), which combines a reference evolutionary process (frozen ESM-2 mutation probabilities exponentially tilted by a fitness function F(x)) with a 4-layer sparse graph transformer over intermediate trees T_t and four per-node MLP heads predicting sitewise amino-acid substitution log-rate offsets, Poisson branching intensities, branch lengths, and lineage stopping probabilities. Training samples intermediate subtrees between the root-only tree T0 = {x0} and FastTree/Nextstrain-reconstructed phylogenies T1, supervising the heads on node/edge statistics of T1 plus a semigroup regularizer. Experiments evaluate TreeSBM on four NCBI viral datasets (SARS-CoV-2 Spike, Influenza A H3N2 HA, Influenza A H1N1 HA, and HIV-1 Env) under geographic, temporal, and epidemic-aware splits, alongside SARS-CoV-2 variant-of-concern (VOC) recovery and ablations.

Critical strengths

None

Critical weaknesses

1: Although the paper is aimed for predictive proposes, it only focuses on "predicting" very old SARS-Cov-2 VOCs appeared 5 years ago. SARS-CoV-2 is a virus continuously mutating at tremendous speed, and every year around 10 new nextclade major lineage and around 1,000 new pango lineages are designated. However, the paper does not discuss about any recent data at all. For example, in 2026 by 2026-10-20, we have 6 new major lineages 26A-26F that achieved very large spread XFG.1.1, XFG.3.4.1, PQ.16.1.1, RK.1, RF.5 and PJ.2.1. and we have 25A-25I major clades LP.8.1, NB.1.8.1, XFG, XFC, XFJ, MC.10.2.1, NW.1.2, BA.3.2 and PY.1.

The authors do not mention any of the 15 major lineages which are actually dominating the world in 2025 or 2026. Also the discussed "predicted" mutations match very few of the actual happening mutations in recent major variants.

2: Fundamental disconnect between the Tree-Valued Schrödinger Bridge theory and the implemented heuristics.

  • Theory vs. implementation (Section 3.3, Algorithms 1–4): Eq. (6) parameterizes Doob-transformed rates via a tree-space potential difference g_theta(T', t) - g_theta(T, t), but the implementation trains independent node-level heads with direct supervised losses against static terminal-tree targets from T1 (Poisson NLL against T1 child count, which is deterministically 2 for internal nodes and 0 for leaves in bifurcating FastTree trees; MSE against mean T1 child branch length; BCE against leaf status; sitewise mutation losses against x1). SampleBridgeState (Algorithm 2) heuristically prunes branches and independently subsamples mutations M_t(v) subseteq M_1(v), which is not the conditional bridge P^0(T_t | T_0, T_1) of Algorithm 3 and mismatches inference (Algorithm 4, Table B.3). Line 218 also injects "per column Shannon entropy", which is undefined at inference given only a single sequence x0.
Questions for authors

1: What's the performance of treeSBM on recent major variants appeared in 2025 and 2026.

2: Please clarify the disconnections between the theory and implemented heuristics, show clearly why the implemented heuristics are perfect approximations of the proposed theory.

4 comments

We thank the reviewer for requesting an evaluation of recent lineages. We took this request seriously and completed an additional experiment. We also recognize the initial rejection recommendations and believe the paper merits a higher score after consideration of the expanded evaluation and the implementation details we explain here.

W1 and Q1. Recent major lineages

the paper does not discuss about any recent data at all.

Table 3 already trains the H3N2 model on sequences collected through 2023 and tests it on 2024 sequences. To extend the evaluation to SARS-CoV-2 in 2026, we evaluated a panel of recent lineages. For the 2026 lineages, constructing moderately sized trees from nonidentical sequences was a low-data problem, because sequences showed almost 100 percent identity with minimal divergence. We therefore built one combined tree from the 2026 Omicron lineages that share ancestry (XFG.1.1, PY.1.1.1, PQ.17, and MC.10.1.7). We also evaluated JN.1 as an earlier comparison.

Table M1. TreeSBM forecasts on recent SARS-CoV-2 lineages. Coverage is the fraction of observed sequences within two substitutions of a generated leaf. Min. distance is the minimum distance to an observed leaf. RBD fraction is the proportion of mutations in the receptor-binding domain.

SetReported panel dateCoverage ≤ 2Union site recallAA|hitMin. distanceEVEscape ΔConservationRBD fraction
Combined 20262026-01-010.5050.11009.130−0.7691.0000.222
JN.12024-01-160.0830.5850.00818.000−1.1991.0000.224

For the combined 2026 set, 50.5% of observed sequences lie within two substitutions of a generated leaf, and generated trees leave every conserved site unchanged. JN.1 reaches a union site recall of 0.585, so the model recovers most of the sites that changed. We additionally report nearest-leaf distances of 9.130 and 18.000 to characterize sequence agreement beyond the coverage threshold. The negative EVEscape Δ values indicate that the generated mutations score lower in predicted escape than the observed ones on these saltational lineages. We report AA|hit separately to distinguish recovery of mutation positions from recovery of the corresponding amino-acid identities.

These results address the request for recent data directly. The experiment extends the evaluation beyond the VOC analysis in Section 4.4 to lineages that emerged in 2026. The second comment addresses the relationship between the theory and the implementation.

W2 and Q2. Theory and implementation

Fundamental disconnect between the Tree-Valued Schrödinger Bridge theory and the implemented heuristics.

We appreciate the opportunity to explain how we implement the motivating bridge formulation. We learn sequence-space corrections and train additional heads for branching, branch length, and stopping. In the revision, we will connect each operation to its training objective and explain the approximation involved.

Motivating structure. The bridge is defined by Eq. (5), and its optimal rates take the Doob form of Eq. (4), Rt⋆=Rt0,φt(T′)/φt(T)R^\star_t=R^0_t\\,\varphi_t(\mathcal{T}')/\varphi_t(\mathcal{T}). The theoretical tree-space parameterization multiplies the reference rates by exp⁡(gθ(T′,t)−gθ(T,t))\exp(g_\theta(\mathcal{T}',t)-g_\theta(\mathcal{T},t)) (Eq. 6), and for mutations the offset is a log-correction on the ESM-2 rate at each site (Eq. 7).

Implemented components. We use node-level heads to predict sequence edits and tree-generation quantities. The sequence heads minimize the KL divergence between an endpoint-weighted ESM-2 target and the predicted categorical distribution, which is the bridge objective on sequence space. The topology heads are trained on node statistics of the observed tree. Most reconstructed phylogenies bifurcate, so child-count targets take the values 0 and 2, and the branching head learns which active lineages split along the bridge clock.

Algorithms 2, 3, and 4. Algorithm 2 samples training states and Algorithm 4 generates trees at inference, and we will describe their separate roles in the revision. Algorithm 2 builds an intermediate state by retaining a subset of the endpoint mutations, Mt(v)⊆M1(v)M_t(v)\subseteq M_1(v), on lineages alive at a sampled date. This construction matches generation, where the model extends lineages that are alive after an intermediate date, and it gives each head a training state of the type it meets at inference. We will state the design choices of Algorithm 2 explicitly.

Entropy feature. The per-column Shannon entropy is defined sitewise over the multiple sequence alignment in Eq. (14) and is used during training. The inference procedure starts from a single sequence x0x_0 and does not use it. The ablation without per-site entropy in Table 7 gives a mutation recall of 0.084, compared with 0.080 for the full model.

Empirical support for the approximation. The semigroup regularizer constrains the learned transitions to compose consistently across bridge times. The ablations in Table 7 show that bridge matching raises mutation recall from 0.008 to 0.080 and AA|hit from 0.087 to 0.375. On held-out geographic SARS-CoV-2 roots, TreeSBM reaches a mutation-site recall of 0.219 against 0.048 for the frozen pLM (Table 4), and the new 2026 experiment in our first comment tests the model on recent-lineage sequences.

Beyond the completed recent-lineage evaluation, we are currently running ProteinGym comparisons against the ESM-2 reference and computing generalized Robinson–Foulds distance alongside strict RF. These analyses will extend our assessment of variant-effect prediction and partial clade agreement. We will report the results during the discussion as they become available.

We thank the reviewer again for prompting a clearer explanation and the additional recent-lineage experiment. We have explained the sequence objective, tree supervision, and ancestral-sequence inference procedure, alongside the tenfold mutation-recall gain from bridge matching. We hope these additions are sufficient to merit a score increase and support for acceptance! Thank you!

Reviewer_mhtdOpenReview ↗

Thanks for the rebuttal.

The rebuttal reveals that the author is not familiar with evolution at all, and are likely mostly AI slops.

I'll maintain my scoring of the lowest score.

While we respect the reviewer’s decision to maintain the lowest score, asserting that we are "not familiar with evolution at all" and describing our responses as “AI Slops” (without providing any justification or explanation) is quite disrespectful, and we will notify the AC of this unprofessionalism. We have been nothing but professional in our responses, and nothing we wrote warrants that level of disrespect. As a note, even if we wanted to use AI tools to help craft our response, the sensitive nature of this work (protein and viral evolution) flags and restricts usage anyway.

For the sake of improving our research, we would like to know what part of evolution are we not understanding: does the reviewer object to our treatment of mutation and selection, the branching process, the connection between the bridge formulation and evolutionary dynamics? Any clarifications here will help us better the research. Thank you.

Reviewer_sCU1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors tackle the problem of modeling realistic protein evolution. This requires (1) generating a phylogenetic tree with branch lengths, together with (2) protein sequences. The proposed method to solve the problem is mathematically principled: fit a Shroedinger bridge with a parameteric model. This requires the fitted process to transport root sequences to real observed trees, while minimizing the KL divergence of the fitted stochastic process wrt the reference process. The authors implement an approximation to this method and go on to show that sequences generated by the model recapitulate SARS CoV variants.

Critical strengths
  • The problem that this paper is trying to solve is interesting: Realistic simulation of protein evolution from a root sequence. This requires (1) generating a phylogenetic tree with branch lengths, together with (2) protein sequences. This is a highly non-trivial problem with important applications such as viral escape prediction.
  • The proposed method to solve the problem is mathematically principled: fit a Shroedinger bridge with a parameteric model. This requires the fitted process to transport root sequences to real observed trees, while minimizing the KL divergence of the fitted stochastic process wrt the reference process.
Critical weaknesses
  • The actual implementation of the method is totally disconnected from the original motivating mathematical foundations. The generative model is fit by minimizing a sum of ad-hoc losses, NOT by minimizing the KL divergence between the stochastic processes subject to the endpoint constraints. Another way to put this is to say that the work basically takes a baseline pLM for sampling mutations, and slaps on top of it additional networks that are used to generate the phylogeny on the go. These networks are trained with ad-hoc losses that try to steer partial trees towards the final trees observed in the data. The original mathematical formalism is completely abandoned. Basically, we end up with a thin wrapper around a usual pLM.
  • Important aspects of the paper are poorly explained, if at all. For example, (i) how does the original optimization objective from Eq. (3) relate to the losses described in Eq. (8). (ii) How exactly is the reference process P^0 defined? Note that section 3.1 "REFERENCE EVOLUTIONARY PROCESS" only described the pLM underlying the mutation process. (iii) How should time t be interpreted? Is the tree depth of T_t equal to t? In general, little or no intuition is provided for what the bridge process is actually learning. (iv) Algorithm "SampleBridgeState" is mean to sample from P^0(T_t | T_0, T_1); why is Algorithm #2 the right way to do this? This is a non-trivial conditional probability distribution. For what I can tell Algorithm #2 is a completely ad-hoc procedure designed to give some "reasonable" tree between T_0 and T_1. What this tree looks like is not described anywhere, I had to read Algorithm #2 to try to descipher what the tree T_t was. (v) What is the fitness prediction F(x) used for exponential tilting?
Questions for authors

In addition to the questions/points raised in the "Critical Weaknesses" section:

  • "We evaluate this framework on SARS the Cov-2" => typo: "We evaluate this framework on the SARS Cov-2"
  • Why does Eq. (2) have a sum and not an integral. Tree space T' is continuous as it has branch lengths.
  • Fig. 1: Illegibly small font size.
  • How exactly are all the loss terms defined? I.e. L_mut, L_cons, L_rate, etc.
  • In Eq. (9), what is G.
  • In section 4.2, how did you compare the generated with the observed trees? Robinson-Foulds is a metric that requires the trees to be defined over the same leaf set. Did you use TreeSBM to infer the topology from the leaves? How did you do this?
  • Table 2: A RF metric of 0.9814 seems really bad, no? This means that 0.9814 of the splits do not exist in the ground truth tree.
  • Table 2: No standard errors to assess significance. Please add e.g. bootstrapped SEs.
  • Section 4.3: How many leaf sequences did you generate per tree? The more sequences the higher the coverage, so without reporting leaf count there is no way to tell if the coverage is any good.
  • Table 7: this table is very suspicious. Why do all ablations have the same Tree-KL of 0.693? How did you compute Tree-KL? Provide SEs in the table. Why does the first row (TreeSBM without bridge matching) do better than TreeSBM full on Cons. retention? The ablations show no sign that TreeSBM full is needed, and reinforces my remark above that TreeSBM is a thin wrapper around a pLM.
6 comments

We thank the reviewer for recognizing the importance of the problem and for explaining the concerns about the implementation. We took the initial rejection recommendations seriously and worked to clarify the objective and expand the evaluation. We believe the paper merits a higher score based on the mutation-recovery gains, the newly fine-tuned baseline comparison, and the additional uncertainty estimates. We explain these results and the underlying method in the two comments that follow.

W1. Relationship between the bridge and the losses

The generative model is fit by minimizing a sum of ad-hoc losses, NOT by minimizing the KL divergence

The sequence component of the objective is the bridge objective on sequence space. Lmut\mathcal{L}_{\mathrm{mut}} and Lcons\mathcal{L}_{\mathrm{cons}} are KL divergences between an endpoint-weighted target built from the ESM-2 reference and the predicted categorical distribution, averaged over mutated and conserved sites. The remaining terms train the branching, branch-length, and stopping heads that generate topology. Eq. (3) states the bridge problem the model approximates, and Eq. (8) to (10) give the loss we minimize. We will state this relationship explicitly in the revision.

The comparison with a frozen and a fine-tuned ESM-2 shows what the learned bridge rates add. On geographic SARS-CoV-2 (Table 4), mutation-site recall is 0.219 for TreeSBM and 0.048 for the frozen pLM, and AA|hit is 0.375 versus 0.017. Fine-tuning ESM-2 on the same 9,561 sequences gives a site recall of 0.158 and an EVEscape Δ of −0.259, compared with 0.219 and −0.048 for TreeSBM. With TreeSBM, we additionally generate topology, branch lengths, and lineage survival.

W2. Points (i) to (v)

(i) Objective. The full optimization objective is LTreeSBM=Lbridge+λsemiLsemi\mathcal{L}_{\mathrm{TreeSBM}}=\mathcal{L}_{\mathrm{bridge}}+\lambda_{\mathrm{semi}}\mathcal{L}_{\mathrm{semi}}. Lbridge\mathcal{L}_{\mathrm{bridge}} combines the sequence KL terms of Eq. (8) with the topology terms, which are a Poisson negative log-likelihood and softplus intensity for branching, a mean-squared error for branch length, a binary cross-entropy for stopping, and an ESM fitness regularizer.

(ii) Reference process. P0P^0 acts on sequence space. Its single-substitution rates come from ESM-2 token logits at each site, tilted by the sequence-level fitness F(x)F(x), the ESM-2 score of the whole sequence, as QF0(x,x′)∝Q0(x,x′)exp⁡(βF(x′))Q^0_F(x,x')\propto Q^0(x,x')\exp(\beta F(x')). The branching and stopping heads predict topology after the sequence edits.

(iii) Time. t∈[0,1]t\in[0,1] is a normalized bridge clock, with the starting sequences at t=0t=0 and the endpoints at t=1t=1. We use the Nextstrain tree clock to date nodes in training and generation, and use tt to specify normalized progression between sequence endpoints.

(iv) Bridge process and Algorithm 2. The bridge learns mutation trajectories from root to leaf by moving probability mass toward observed endpoints while keeping it close to ESM-2 residue predictions. Algorithm 2 samples training states by extending lineages that are alive after a random intermediate date, so training states match the states the model meets at generation. We will explain each design choice in Algorithm 2 and separates it from the inference procedure in Algorithm 4.

(v) Fitness function. F(x)F(x) is the ESM-2 score of the whole sequence, defined in (ii).

Q1 to Q5. Presentation

We will expand the variable and loss definitions, enlarge the font in Figure 1, and correct the typo. For Eq. (2), the tree space is a union over topologies of continuous branch-length domains, and we will write the master equation as a sum over topologies and an integral over branch lengths.

Q6 and Q7. Leaf matching and strict Robinson-Foulds distance

A RF metric of 0.9814 seems really bad, no? This means that 0.9814 of the splits do not exist in the ground truth tree.

We match each generated leaf to an observed leaf by Hamming distance and compute RF on the matched leaf set. Topology comes from the branching and stopping heads at inference, and most reconstructed trees bifurcate. Strict RF counts a split as correct only when the clade matches exactly, so leaf-placement errors can affect multiple splits. In the table below, all six methods fall between 0.9786 and 0.9881, so strict RF sits near its maximum for every method, including the neutral birth-death and pLM-prior baselines. We are adding generalized RF, which gives partial credit to near-matching clades, next to strict RF in Table 2 and in the ablation tables.

Q8. Standard errors for Table 2

MethodRFQuartetBranch W1Terminal edit
Neutral birth–death0.9801 ± 0.002510.7023 ± 0.01040.000822 ± 0.0001170.004852 ± 0.000718
Empirical birth–death0.9817 ± 0.002750.7021 ± 0.01080.000822 ± 0.0001170.005546 ± 0.000734
pLM prior0.9786 ± 0.002760.7007 ± 0.01080.000822 ± 0.0001170.005191 ± 0.000765
ARTreeFormer0.9843 ± 0.003010.7022 ± 0.01050.000984 ± 0.0001300.005476 ± 0.000727
PhyloVAE0.9881 ± 0.002840.7051 ± 0.01040.000925 ± 0.0001240.005363 ± 0.000751
TreeSBM0.9838 ± 0.002650.7073 ± 0.01040.001000 ± 0.0001360.003705 ± 0.000600

With TreeSBM, we obtained the lowest mean terminal-sequence edit distance among all six methods, including a 28.6% reduction relative to the pLM prior. We report the standard errors for every method so the reviewer can assess the magnitude and uncertainty of each comparison.

Q9. Leaf sequences per tree

TableVirusLeaves per treeTrees KLeaves in the pool
Table 3, Cov.@10H3N2 HA1610160
Table 3, Cov.@100H3N2 HA161001,600
Table 4, Cov.@100SARS-CoV-2 spike, H1N1 HA, H3N2 HA, HIV-1 Env161001,600

All methods in Table 3 appear at the same K. We will add the leaf counts to the tables and captions.

Q10. Table 7

Why do all ablations have the same Tree-KL of 0.693? [...] Why does the first row do better on Cons. retention?

We obtained the same reported Tree-KL value of 0.693 across the component and reference-process ablations. We will state its implementation and normalization explicitly alongside the exact-clade comparison. To assess partially matching clades as well, we are adding generalized RF and will report the completed analysis separately.

The sequence-level columns do separate the variants. Without bridge matching, mutation recall is 0.008 compared with 0.080 for the full model, and AA|hit is 0.087 compared with 0.375. Without bridge matching, conservation retention was 0.876, compared with 0.839 for the full model. With bridge matching, we recovered ten times as many observed mutation sites and achieved substantially higher conditional amino-acid accuracy. We therefore interpret conservation and mutation recovery together.

We are currently computing generalized Robinson–Foulds distance alongside strict RF, including the topology and ablation comparisons, to quantify partial clade agreement. In parallel, we are running the ProteinGym comparison against the ESM-2 reference to extend the assessment to variant-effect prediction. We will report these additional results as the analyses are completed.

We thank the reviewer again for the detailed questions. We have added standard errors, specified the leaf budgets, and clarified the contribution of bridge matching through the mutation-recovery ablation and the new fine-tuned baseline. Given these additions and the lowest mean terminal edit distance in the updated comparison, we respectfully ask the reviewer to increase the score and reassess confidence in the submission.

Reviewer_sCU1OpenReview ↗

Response to Reviewers

Dear Authors,

Thanks for your response.

It appears to me that you have not updated your manuscript to reflect any of the reviewer feedback. Do you intend to retain the original submission as-is during the review process? Or should we expect an updated manuscript before the author-reviewer discussion ends on November 18th?

Best regards

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

This paper uses a "reference evolutionary process" which amounts to an exponential tilt on top of ESM2's mutation measure by an (as far as I can tell not defined or specified) "fitness predictor," coupled to a poisson process for "mutation times."

The authors do not specify what tilting procedure they use to to tilt the ESM2 diffusion process, nor do they specify what fitness function they tilt it by. As they subsequently seek to keep their evolutionary process "close" in KL to this reference measure it's critical they specify what it actually is.

They then use a second exponential tilt, this time on the trees themselves, claiming to learn a log transition rate predictor g_{\theta} for a given tree, e.g. log R(T, T') = Delta g + const. They implement this site-wise rather than via proposals, so it amounts to a sitewise correction to their initial tilt procedure (which remains unspecified)

In defining the loss used to train this g_{\theta} they introduce several terms L_{const}, L_{mut}, L_{bridge} which are simply never defined at any point. Not defining your loss function at all is such an egregious oversight that I can only assume this paper was not written by a human researcher.

In addition, their benchmarks (Tables 2-4) don't seem to point to their model as being meaningfully better than PLM at predicting viral evolution

Critical strengths

Table 2 does suggest that their tilting procedure is somewhat effective at recovering tree topology; which is interesting.

Critical weaknesses
  • They aren't really doing anything new here - they're adjusting an exponential tilt on ESM2 by a secondary site-wise tilt that comes from the tree level transition probabilities.
  • Those tree level transition probabilities are trained via a loss that they do not properly define.
  • I'm not confident that these benchmarks are fair, given that their model was tuned on the SARS- Cov 2 spike protein and the others were not (each of those models is capable of giving a log likelihood of a given tree, which one could use to fine tune them on the same dataset that was used to train their model).
Questions for authors
  • What is the loss function you used to train your tree level transition rate predictor?
  • how do you parameterize the transition rate? Obviously a |trees| x |trees| matrix is infeasible, so are you doing at a per site level? If so how is this different from fine tuning your initial model with respect to the log likelihood of the trees in the training set?
2 comments

We thank the reviewer for examining the contribution and baseline comparisons so closely. We took the initial rejection recommendations seriously and carried out the requested fine-tuning comparison. We believe the paper merits a higher score because we now demonstrate improved mutation-site recovery against ESM-2 trained on the same sequence collection. We first explain the contribution and objective, then present the new results.

W1. Contribution

They aren't really doing anything new here

Our contribution is to generate a distribution of branching evolutionary histories from one ancestral sequence. To clarify the scope of this task, we distinguish the outputs and conditioning information in the comparisons below.

Method classOutputConditioning
Protein language, diffusion, and flow modelsIndependent sequencesNone or sequence context
Phylogenetic inference (FastTree, BEAST 2)One historical treeObserved leaf sequences
Topology generators (ARTreeFormer, PhyloVAE)Distribution over topologiesFixed taxon set
TreeSBMDistribution over future trees with sequences, topology, branch lengths, and lineage survivalOne ancestral sequence

Each sampled tree contains the full ancestry and descendants of every node, so a forecast contains coherent lineages with shared history. Independent sequence samples cannot represent which lineages persist and branch, and vaccine selection and antiviral design depend on that information. The learned correction is a site-level log-offset trained against observed trees, and the comparisons in the next comment show what it adds over the frozen pLM and over a fine-tuned ESM-2.

W2 and Q1. Loss function and fitness predictor

What is the loss function you used to train your tree level transition rate predictor?

The objective is LTreeSBM=Lbridge+λsemiLsemi\mathcal{L}_{\mathrm{TreeSBM}}=\mathcal{L}_{\mathrm{bridge}}+\lambda_{\mathrm{semi}}\mathcal{L}_{\mathrm{semi}}. Lbridge\mathcal{L}_{\mathrm{bridge}} contains a sequence term, λmutLmut+λconsLcons\lambda_{\mathrm{mut}}\mathcal{L}_{\mathrm{mut}}+\lambda_{\mathrm{cons}}\mathcal{L}_{\mathrm{cons}}, and a topology term. Lmut\mathcal{L}_{\mathrm{mut}} and Lcons\mathcal{L}_{\mathrm{cons}} are KL divergences between the endpoint-weighted ESM-2 target and the predicted categorical distribution, averaged over mutated and conserved sites. The topology term sums a Poisson negative log-likelihood for branching intensity, a mean-squared error for branch length, a binary cross-entropy for stopping, and an ESM fitness regularizer. Lsemi\mathcal{L}_{\mathrm{semi}} penalizes the discrepancy between a direct transition from bridge time ss to tt and the composition of transitions through an intermediate time rr. We will define each term in Section 3.3 of the revision, and our response to Reviewer pRaq states them in full.

The fitness function F(x)F(x) is the ESM-2 score of the whole sequence, and the reference is QF0(x,x′)∝Q0(x,x′)exp⁡(βF(x′))Q^0_F(x,x')\propto Q^0(x,x')\exp(\beta F(x')) over single-substitution neighbors.

Q2. Parameterization of the transition rate

We parameterize the sequence bridge locally, learning a log-offset for each node, site, and residue, log⁡Rθ(i,a)=log⁡R0(i,a)+cθ(i,a)\log R_\theta(i,a)=\log R^0(i,a)+c_\theta(i,a) with cθ∈RL×20c_\theta\in\mathbb{R}^{L\times 20}. Separate heads predict branching, branch length, and stopping from the same node embeddings.

We condition the learned offset on the current tree state and bridge time, allowing sequence predictions to depend on the developing lineage. To test the benefit beyond adapting the underlying language model, we additionally fine-tuned ESM-2 on the same training sequences. We report that experiment in the next comment.

W3. Fairness of the benchmarks

I'm not confident that these benchmarks are fair, given that their model was tuned on the SARS-Cov 2 spike protein and the others were not

To test this directly, we fine-tuned ESM-2 on the 9,561 distinct SARS-CoV-2 sequences used to train the TreeSBM checkpoint. We corrupted 15% of standard amino-acid positions and minimized cross-entropy at the corrupted positions. This control sees the same sequences as TreeSBM, and we evaluated it on the same splits as Table 4.

Table R1. SARS-CoV-2 spike forecasting on held-out roots. Higher is better except for the absolute value of EVEscape Δ.

SplitModelCoverage@100Site recallAA|hitEVEscape ΔCons. retention
EpidemicFine-tuned ESM-20.8000.2501.000−0.2481.000
EpidemicFrozen pLM0.800 ± 0.2000.3290.088−0.1090.710
EpidemicTreeSBM0.800 ± 0.2000.3570.110−0.0040.753
GeographicFine-tuned ESM-20.7750.1581.000−0.2591.000
GeographicFrozen pLM0.7750.0480.017−0.4240.985
GeographicTreeSBM0.7880.2190.375−0.0480.840

With TreeSBM, we achieved 42.8% higher mutation-site recall on the epidemic split and 38.6% higher recall on the geographic split than with fine-tuned ESM-2. Recall was 0.357 versus 0.250 and 0.219 versus 0.158, respectively. Coverage matched at 0.800 on the epidemic split and increased from 0.775 to 0.788 on the geographic split. The EVEscape Δ of the fine-tuned model was −0.248 and −0.259, compared with −0.004 and −0.048 for TreeSBM, so the mutations it generated departed further from the escape profile of observed mutations. The fine-tuned model reached an AA|hit of 1.000 over a smaller set of recovered sites and left every conserved site unchanged (retention 1.000). We report these complementary measures alongside the broader mutation-site recovery achieved with TreeSBM.

The frozen pLM comparison also separates the methods. On geographic SARS-CoV-2, TreeSBM recovers 4.6 times as many mutated sites (0.219 versus 0.048) and reaches an AA|hit 22 times higher (0.375 versus 0.017). On H3N2 temporal and H1N1 geographic splits, site recall is 0.253 versus 0.059 and 0.241 versus 0.071 (Table 4). TreeSBM reaches these values while generating topology, branch lengths, and lineage survival, alongside these sequence-level predictions.

The reviewer also noted that tree models can score a tree under a log-likelihood. The sequence-level control above gives the closest matched comparison for the mutation component. The tree baselines in Tables 2 and 3 are evaluated on the topology and lineage metrics defined in Section 4.

To extend these completed comparisons, we are currently evaluating TreeSBM against the ESM-2 reference on ProteinGym. We are also computing generalized Robinson–Foulds distance alongside strict RF to assess partial clade agreement. We will add the results during the discussion as these analyses are completed.

We thank the reviewer for requesting this comparison. We have now adapted ESM-2 to the same training sequences and demonstrated higher mutation-site recall with TreeSBM on both splits. Together with the explicit objective and generation of branching histories, these results strengthen the case for the contribution and respectfully ask the reviewer to increase the score and reassess confidence in the submission to support its acceptance, rather than its rejection.

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

This paper proposes TreeSBM, which tries to forecast protein evolution: given a single ancestral sequence, it samples a distribution over future evolutionary trees, covering descendant sequences, branching topology, and branch lengths together. The reference process is ESM-2's mutation measure, exponentially tilted by a fitness predictor. A 4-layer sparse graph transformer runs over intermediate trees and drives four per-node heads: substitution log-rate offsets, Poisson branching intensity, branch length, and stopping probability. These heads are supervised on statistics of terminal trees reconstructed with FastTree and Nextstrain, plus a semigroup regularizer. The experiments cover four viral datasets (SARS-CoV-2 Spike, H3N2 HA, H1N1 HA, HIV-1 Env) under geographic, temporal, and epidemic-aware splits, with SARS-CoV-2 variant-of-concern recovery as the headline application, along with component ablations.

Critical strengths

Forecasting a distribution over complete branching histories, i.e. sequences, topology, and branch lengths jointly, is a well-motivated problem that has not received much attention, and the Schrödinger bridge formulation is a principled place to start.

Critical weaknesses
  1. I cannot reconstruct the training objective from the paper. Eq. (8) writes L_rate = λ_mut L_mut + λ_cons L_cons, but neither L_mut nor L_cons is ever defined. The other losses, the Poisson NLL on branching, the MSE on branch length, the BCE on stopping, and the ESM fitness regularizer, are also not formulated and defined. G and d(·,·) in Eq. (9) are also undefined. Then Eq. (10) gives L_TreeSBM = L_bridge + λ_semi L_semi, where L_bridge appears for the first time and is never connected back to Eq. (8). So I cannot tell what was actually optimized.

  2. I don't see why a Schrödinger bridge is the right choice here. Section 3.3 defines the bridge as the minimum-KL process to an ESM-2 reference under endpoint constraints, but it does not say what this buys for evolution forecasting, or why it fits this problem better than the alternatives. Why would a directly trained CTMC, flow matching, or even the autoregressive tree-edit model already used as a baseline not work just as well? Moreover, the experiments don't settle it either. In Table 7, removing bridge matching leaves both tree-level metrics completely unchanged (Tree-KL 0.693, Coverage@100 0.775), and in Table 3 the autoregressive baseline recovers clades 2–3× better. I would like the authors to say which property of the bridge makes it the right tool here, and discuss this explicitly.

  3. The paper should report ProteinGym results. Eq. (7) outputs a per-position substitution score, which is exactly the form used for zero-shot variant-effect prediction, and the reference process is ESM-2 itself. Scoring the viral DMS assays in ProteinGym and comparing against the ESM-2 reference alone would directly show whether the learned tilt captures anything beyond its own prior.

Questions for authors

Please see the above weaknesses.

2 comments

We thank the reviewer for reading the paper closely and explaining which parts needed clarification. We took the initial rejection recommendations seriously and have worked to address them through additional experiments and a clearer explanation of the method. We believe the paper merits a higher score, and we appreciate the opportunity to explain why. We developed TreeSBM to generate a distribution over future evolutionary trees, including descendant sequences, topology, branch lengths, and lineage survival, from a single ancestral sequence. Protein language models and diffusion models generate independent sequences, and phylogenetic methods reconstruct past trees from observed leaves. We explain the formulation and completed analyses in the responses that follow.

W1. Training objective

I cannot reconstruct the training objective from the paper.

To address this concern, we explain each component of the objective explicitly. We use the Schrödinger bridge in Eq. (5) to motivate the learned process and optimize the following combination of sequence, topology, and consistency terms

\mathcal{L}_{\mathrm{bridge}}=\lambda_{\mathrm{mut}}\mathcal{L}_{\mathrm{mut}}+\lambda_{\mathrm{cons}}\mathcal{L}_{\mathrm{cons}}+\mathcal{L}_{\mathrm{topo}}.$$ **Sequence bridge matching ($\mathcal{L}_{\mathrm{mut}}$, $\mathcal{L}_{\mathrm{cons}}$).** For an active node and site $i$, the mutation head predicts a categorical distribution over residues, $p_\theta(\cdot\mid\mathcal{T}_t,t)\propto R^0(i,\cdot)\exp(c_\theta(i,\cdot))$, as in Eq. (7). The target is the ESM-2 reference distribution reweighted toward the residue observed in $\mathcal{T}_1$. $\mathcal{L}_{\mathrm{mut}}$ is the KL divergence between target and prediction averaged over mutated sites, where the bridge-state residue differs from the endpoint, and $\mathcal{L}_{\mathrm{cons}}$ is the same quantity averaged over conserved sites. **Topology ($\mathcal{L}_{\mathrm{topo}}$).** This term sums four components. A Poisson negative log-likelihood fits the softplus branching intensity $\lambda_\theta(v)$ to the child count of node $v$ in $\mathcal{T}_1$. A mean-squared error fits the predicted branch length to the mean child branch length in $\mathcal{T}_1$. A binary cross-entropy fits the stopping probability to the indicator that $v$ has no children in $\mathcal{T}_1$. An ESM fitness regularizer equals the negative mean log reference rate, $\log R^0$, of the current bridge residues. **Semigroup regularizer ($\mathcal{L}_{\mathrm{semi}}$).** $G_{s,t}$ denotes the learned transition from bridge time $s$ to $t$, and $d$ measures the discrepancy between $G_{s,t}$ and the composition $G_{r,t}\circ G_{s,r}$ for $s<r<t$. The fitness function $F(x)$ is the ESM-2 score of the full sequence $x$, and it tilts the reference as $Q^0_F(x,x')\propto Q^0(x,x')\exp(\beta F(x'))$ over single-substitution neighbors. Time $t\in[0,1]$ is a normalized bridge clock, with the root-only tree at $t=0$ and the observed tree at $t=1$. Algorithm 2 samples intermediate bridge states during training, and Algorithm 4 generates trees at inference. The next comment addresses the choice of a Schrödinger bridge.

W2. Justification for the Schrödinger bridge

I would like the authors to say which property of the bridge makes it the right tool here, and discuss this explicitly.

We chose bridge matching to learn how ancestral sequences develop toward observed descendants while retaining a pretrained sequence reference. Three aspects motivated this choice. First, we condition training on both endpoints. Each training pair joins a root-only tree to an observed tree, and bridge states Tt\mathcal{T}_t between them allow supervision at every intermediate time without simulating the process. Second, we learn a log-offset cθc_\theta on the ESM-2 reference log-rate (Eq. 7). Each dataset contains only 147 to 404 trees (Table 1), and we use the reference to retain pretrained sequence information when fitting these smaller collections. A transition model fit from scratch would need to estimate site-level rates for every node from these trees alone. Third, we penalize disagreement between direct and composed transitions to encourage consistency over the longer generation horizons.

Evidence from the ablations. In Table 7, the variant without bridge matching differs from the full model on every sequence-level metric.

Variant (held-out SARS-CoV-2 roots)Mutation recall ↑AA|hit ↑Cons. retention ↑Tree-KL ↓Coverage@100 ↑
Without bridge matching0.0080.0870.8760.6930.775
TreeSBM full0.0800.3750.8390.6930.775

Bridge matching raises mutation recall tenfold and AA|hit from 0.087 to 0.375. We also report the higher conservation retention without bridge matching, alongside the greater recovery of observed mutations with the full model. Tree-KL and Coverage@100 take identical values in every row of Tables 6 and 7, and Coverage@100 equals 0.775 for the frozen pLM and the autoregressive baseline on the same split (Table 4). We therefore compare mutation recovery and amino-acid identity to distinguish the variants in this setting. The reported Tree-KL values are identical at the displayed precision, alongside the low exact-clade recovery. We are adding generalized Robinson-Foulds distance, which gives partial credit to near-matching clades, to Tables 2, 6, and 7, so that topology is measured with a sensitive metric.

Autoregressive tree-edit baseline. The autoregressive baseline recovers more complete H3N2 clades (clade recall 0.300 versus 0.100, Table 3). TreeSBM exceeds it on the quantity that determines which variants a forecast contains. Mutation recall on H3N2 is 0.156 for TreeSBM and 0.079 for the autoregressive model, and mutation-site recall on geographic SARS-CoV-2 is 0.219 versus 0.006 (Table 4), a 36-fold difference. The frozen language model reaches a site recall of 0.048 on that split and an AA|hit of 0.017, compared with 0.375 for TreeSBM.

Ongoing experiments

To address the reviewer’s request directly, we are currently running the ProteinGym evaluation against the ESM-2 reference. This comparison will assess variant-effect prediction and complement the completed lineage-forecasting experiments. We are also computing generalized Robinson–Foulds distance alongside strict RF to assess partial agreement between generated and observed clades. We will report these results during the discussion as the analyses are completed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the reviewers for reading TreeSBM closely, and providing valuable feedback -- we took all four initial rejection recommendations seriously. We agreed that readers should be able to tell exactly what we trained, why we used a bridge, and how much we gained over a protein language model trained on the same data. During rebuttal, we revised the formulation, trained the requested baseline, and evaluated recent lineages. We ask the Area Chair and reviewers to consider the November 17 manuscript and these new results in the final assessment.

Importance of forecasting a distribution of trees

Starting from one ancestral protein sequence, we want to forecast several possible histories of how its descendants might diversify. A list of independently generated sequences leaves open which descendants share a branch. A historical reconstruction starts with the descendants already in hand. With TreeSBM, we start with the ancestor alone and jointly generate descendant sequences, their branching relationships, and branch lengths. This lets us evaluate two distinct questions. How closely do the generated descendants resemble those we later observe? How closely do their branching relationships agree with the reconstructed tree?

New experiments and revisions for each reviewer

  • Reviewer pRaq asked what objective we trained and why we needed bridge matching. We rewrote the formulation to distinguish the bridge problem that motivates TreeSBM from the loss we actually minimize. We also compared the full model with a version trained without bridge matching. On held-out roots, mutation-site recall rose from 0.008 to 0.080. Among correctly identified sites, amino-acid accuracy rose from 0.087 to 0.375. Tree-KL and coverage were unchanged at the reported precision. The clearest contribution of bridge matching in this experiment was better recovery of sequence changes.

  • Reviewer qRdW asked whether the gains would hold against a protein language model trained on the same data. We fine-tuned ESM-2 and evaluated both models on the same held-out roots. TreeSBM recovered more observed mutation sites on both splits, with recall of 0.357 versus 0.250 on the epidemic split and 0.219 versus 0.158 on the geographic split. Fine-tuned ESM-2 achieved higher amino-acid accuracy at the sites it recovered. Reporting both measures shows exactly where each method performed well.

  • Reviewer sCU1 asked how we compared generated and observed trees and how much uncertainty there was in Table 2. We explained how we matched leaves, reported the number of leaves sampled, and added standard errors for all six methods. TreeSBM achieved the lowest mean terminal-sequence edit distance, 0.003705 ± 0.000600, compared with 0.005191 ± 0.000765 for the protein language model prior, a 28.6% reduction in mean distance. Strict Robinson–Foulds distances were close to their maximum for all six methods. In the revised manuscript, we also evaluate partial agreement between clades so readers can assess topology separately from sequence agreement.

  • Reviewer mhtd asked us to evaluate recent lineages and explain how the implementation relates to the theory. We added an evaluation on a combined panel of four lineages from 2026. 50.5% of observed sequences fell within two substitutions of a generated leaf. This measures how closely we approached observed descendants. We separately reported exact amino-acid recovery at correctly identified mutation sites, which was zero on this panel. We also revised the method section to identify which parts of the model follow from the bridge formulation and how we train the branching and stopping predictions.

Case for acceptance

We believe TreeSBM deserves acceptance for introducing and testing a way to forecast branching protein histories from a single ancestor. The strongest evidence comes from three comparisons. We recovered more observed mutation sites on both splits after training ESM-2 on the same data. We showed which sequence-level gains depended on bridge matching. Across six methods, we generated terminal sequences with the lowest mean edit distance to observed descendants. Though we know it may be difficult given the initial scores, we respectfully ask the Area Chair and reviewers to support acceptance of TreeSBM after additional consideration.

Thank you,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.