Question 1

We thank the reviewer for highlighting this important point. Conventional PLMs are trained with a masked-language-model (MLM) loss on single sequences. That objective optimises per-token reconstruction, not the relationships among sequences, so the embeddings do not preserve evolutionary distance. A simple Hamming count is crude but is monotonically related to divergence, so outperforms these models on tree reconstruction.

Our experiments make this explicit. Swapping PHYLA’s quartet-based tree loss for MLM decreases the ability for PHYLA to do evolutionary reasoning; fine-tuning a head on the 650M-parameter ESM-2 backbone likewise fails to beat Hamming, and removing PHYLA’s multi-sequence attention blocks drops performance dramatically. Scale cannot rescue a model whose objective and architecture is misaligned with the task.

This result matters for the community. It shows that objective and architecture alignment, not parameter count, is the bottleneck, and it motivates our Evolutionary-Reasoning Benchmark and PHYLA’s design. Publishing these findings will prevent researchers from assuming that large MLM-trained PLMs “implicitly learn evolution” and will guide future work toward loss functions and architectures that capture the phylogenetic signal. PHYLA represents the first step in a new class of evolutionary-focused PLMs with its own unique model architectures, loss functions, and benchmarks.

Question 2

We thank the reviewer for raising this point. Our original focus was to test whether general‑purpose PLMs capture phylogenetic structure, but we agree that including a classical pipeline provides an essential upper bound. We have therefore run the standard MAFFT + FastTree workflow on the same TreeBase and TreeFam splits. It achieves the best tree‑reconstruction scores (normRF ≈ 0.65 on TreeBase and ≈ 0.32 on TreeFam), while PHYLA attains 0.73 and 0.58 respectively—closing roughly half of the gap that separates PLMs from classical methods. Importantly, the classical run required ~2 hours on TreeBase and ~66 hours on TreeFam across multiple CPU nodes (30 nodes with 75 GB memory each with  52% usage each), whereas PHYLA produced both trees in under an hour on a single H100 GPU (83% memory usage). These results, now reported in the appendix, clarify the landscape: classical align‑then‑tree pipelines remain the accuracy benchmark, but PHYLA offers a substantial speed advantage and demonstrates that models can approach SOTA accuracy without an explicit alignment step.

Weakness 1

We thank the reviewer for this thoughtful and constructive comment. We agree that it is an important point that the way PLMs are used in the tree reconstruction benchmark is by simply computing Euclidean distances in embedding space (previously only mentioned in the appendix, line 607). We will move this detail to the main text to make the evaluation setting clearer. To address the concern that this setup may be unfairly harsh on current PLMs, we performed an additional experiment with the best-performing PLM on the benchmark, ESM-650M, by training a network on top of its embeddings using the same tree-loss employed by PHYLA. We froze ESM-650M during training, as it was not computationally feasible to store gradients for larger trees containing over 100 proteins; this limitation highlights a key advantage of PHYLA’s architecture, which is designed to jointly process all sequences efficiently.

We evaluated two variants: (i) a simple feed-forward network (FNN) head and (ii) a transformer-based head over the protein embeddings. In both cases, ESM-650M was unable to match PHYLA’s performance. On the TreeBase benchmark, the corrected ESM-650M models achieved normalized RF scores starting at 0.78 and plateauing at 0.77/0.76, while on the TreeFam benchmark performance degraded to 0.79/0.80. These experiments show that PHYLA’s superior accuracy stems from its architecture’s ability to reason jointly over entire sequence sets, not from the PLMs being evaluated in zero‑shot settings. We have added these results to the appendix.

Weakness 2

We thank the reviewer for highlighting this crucial point. Our intent was not to claim that normalized Hamming distance (HD) is less biased than tree‑based algorithms and distances, but rather to avoid injecting the hyper‑parameter choices and heuristic assumptions of a particular tree‑building algorithm into the supervision signal. We have clarified this wording in the paper and explicitly acknowledge that FastTree/IQ‑Tree (along with other algorithms, RaxML, etc) are state‑of‑the‑art in this space.

Why HD was used originally.

Definition & clarity: To be clear, the (normalized) HDs are computed column‑wise on each multiple‑sequence alignment (MSA) not on raw, unaligned sequences.

Practicality: MSAs were already computed on the OpenProteinSet and HDs can be obtained from MSAs in O(L × N²) time, whereas building trees even with FastTree with a large number of trees is computationally expensive.

Focus on topology: PHYLA’s quartet loss optimizes which branches merge, independent of branch length. FastTree likewise derives its initial topology from the pairwise similarity in the underlying MSA, then refines topology and branch lengths. We therefore treated HD as a proxy for topological signal, not a perfect metric of evolutionary time.

New experiment requested by the reviewer.

We reconstructed FastTree trees for all training MSAs, extracted pairwise tree distances, and re‑trained PHYLA with these distances in place of HD. Test results:

The near‑identical scores indicate that PHYLA is not merely copying HD; it is learning to compare sequences in a way that generalizes across distance definitions.

Take‑away.

Using HD does not handicap PHYLA on the evaluated topological tasks, while avoiding dependence on specific FastTree/IQ‑Tree settings. We have added this explanation, the table above, and a note that incorporating richer substitution models for branch‑length prediction is valuable future work.

Weakness 3

We thank the reviewer for raising this point. The current wording in Section A 1.2 is confusing, and we have revised it. To clarify, PHYLA is always fed raw, unaligned amino‑acid sequences. The only role of the multiple‑sequence alignments (MSAs) in training is to provide a target distance matrix—we compute normalized Hamming distances on the MSA, then discard the alignment. Thus the model never “sees” gap symbols and is explicitly trained on unaligned inputs.

Because the supervision signal is derived from MSAs while the inputs are raw sequences, PHYLA must implicitly learn to align (or otherwise compare) sequences in a differentiable manner to minimise the quartet loss. This built‑in alignment‑free processing is precisely why the model generalises to novel, unaligned sequences at test time.

Weakness 4

We thank the reviewer for pointing out the need to quantify any overlap between the training corpus (OpenProteinSet) and the evaluation sets (TreeBase, TreeFam). We ran BLAST between each tree in TreeBase and TreeFam and all the sequences in the training corpus using an e-value cutoff of 1e-5 and bitscore cutoff of 50. We found in TreeBase only 0.56% of evaluation trees had any match to the pretraining set. The overall mean percent identity of sequences in TreeBase is 0.38%. In TreeFam only 3.38% of evaluation trees had any match to the pretraining set. The overall mean percent identity of sequences in TreeFam is 0.23%. These results indicate minimal leakage and highlights Phyla’s ability to generalize. We will add this to the appendix and cite it in the main text.

Questions

We thank the reviewer for raising these important questions, we believe our detailed responses above (the new experiments and clarifications) fully resolve these points.

Limitations

We thank the reviewer for emphasizing that several limitations were in the appendix. Space constraints forced difficult trade‑offs, but we agree these details belong in the main narrative. In the revised draft we have moved all limitations from the appendix to the discussion and added more training details from the appendix to the main text. These changes make the scope and limitations explicit while respecting the page limit.

Weakness 1

We thank the reviewer for pointing out the need to clarify the evaluation protocol.

How PLM baselines were evaluated

1. Input: each raw (unaligned) protein sequence is fed through the frozen PLM
2. Embedding extraction: we extracted protein embeddings as detailed by each PLM
3. Distance computation: Pairwise Euclidean distance is computed on these vectors

Tree scoring: We ran a neighbor-joining algorithm on these distances to construct a tree from which we calculated the norm-RF. This description is in the appendix, but as other reviewers have also pointed out, belongs in the main text and we have moved it there.

Why evaluate PLMs this way

We agree there is no a‑priori guarantee that these embedding distances encode phylogeny. Yet several influential papers and benchmarks (ESM2 [1], ESM3 [2], DGEB Benchmark [3], ESMC [4], EVO2[5]) argue that PLMs capture evolutionary relationships. Our study rigorously tests that claim and, as the reviewer notes, finds little evidence in support.

Contribution clarified

The key contribution is therefore two‑fold:

Empirical: we show systematically that state‑of‑the‑art PLMs do not recover phylogenetic structure under standard usage.

Methodological: we introduce PHYLA, whose quartet loss and architecture enables it to reason over evolutionary splits, achieving state of the art performance on our benchmark

These results highlight a gap in current protein representation learning and propose a concrete path forward.

---

Weakness 2

We thank the author for this important comment. We define evolutionary reasoning as the following:

Let $S = \{s_1, \ldots, s_n\}$ be a set of protein sequences drawn from an (unknown) evolutionary process that generated a true, strictly-binary phylogenetic tree $T^\star$ with leaves in bijection with $S$.

Denote by $d_{\text{evol}}(i, j)$ the patristic distance between $s_i$ and $s_j$ on $T^\star$.

A model $f_\theta : \Sigma^\star \rightarrow \mathbb{R}^d$ is said to perform evolutionary reasoning on $S$ if there exists a computable mapping $h$ (e.g., a distance-based tree builder) such that the tree

$$\hat{T} = h(D_{\text{pred}}), \quad \text{where } D_{\text{pred}}[i, j] = \|f_\theta(s_i) - f_\theta(s_j)\|,$$

minimizes a tree-level loss $\mathcal{L}_{\text{tree}}(\hat{T}, T^\star)$.

Where in our case $h$ is the neighbor joining algorithm and $\mathcal{L}_{\text{tree}}$ is the normalized Robinson–Foulds distance. We have added this definition to the problem definition portion of the manuscript.

---

Weakness 3

We thank the reviewer for raising this point. Our original focus was to test whether general‑purpose PLMs capture phylogenetic structure, but we agree that including a classical pipeline provides an essential upper bound. We have therefore run the standard MAFFT + FastTree workflow on TreeBase and TreeFam. It achieves the best tree‑reconstruction scores (normRF ≈ 0.65 on TreeBase and ≈ 0.32 on TreeFam), while PHYLA attains 0.73 and 0.58 respectively—closing roughly half of the gap that separates PLMs from classical methods. Importantly, the classical run required ~2 hours on TreeBase and ~66 hours on TreeFam across multiple CPU nodes, whereas PHYLA completed both benchmarks in under an hour on a single H100 GPU. These results, now reported in the appendix, clarify the landscape: classical align‑then‑tree pipelines remain the accuracy benchmark, but PHYLA offers a substantial speed advantage and demonstrates that models can approach SOTA accuracy without an explicit alignment step. We believe this addition resolves the fairness concern and sets a clear, measurable target for future improvements.

---

Weakness 4

We thank the reviewer for raising this important concern. We evaluated PLMs in a zero‑shot setting because a substantial body of work [1, 2, 4, 5] claims that their pretrained embeddings already encode evolutionary relationships. To ensure this choice did not disadvantage the PLMs, we fine‑tuned the strongest model, ESM‑650 M, with two lightweight heads (a FNN head and transformer head) trained using the same quartet loss as PHYLA while freezing the backbone (back‑propagating through 650 M parameters for >100‑leaf trees is infeasible). Despite this targeted correction, the best model reached a normRF 0.77 on TreeBase and 0.79 on TreeFam, whereas PHYLA—trained end‑to‑end to reason over sequence sets—achieves 0.73 and 0.58 respectively. These results confirm that the performance gap is not simply a consequence of evaluating PLMs in a zero‑shot setting. We have added these results to the appendix.

---

Weakness 5

We thank the reviewer for catching this. We have inserted explicit citations where we state that PLMs are often described as capturing evolutionary signals (lines 27–30, ESM2 [1], ESM3 [2], DGEB Benchmark [3], ESMC [4], EVO2[5]). We also re‑scanned the manuscript and added citations wherever similar assertions were previously unsupported.

---

Weakness 6

We thank the reviewer for raising this important point. To meet the page requirement we moved model architecture details to the appendix, but after concerns raised by other reviewers as well we will move it back to the main text. From the appendix:

The PHYLA architecture alternates between inter-sequence and intra-sequence reasoning blocks.Inter-sequence reasoning is performed using BiMamba layers, while intra-sequence reasoning is handled by sparsified attention layers. The current 24M-parameter PHYLA model comprises three inter-sequence blocks, each containing 16 BiMamba layers followed by a single sparsified attention layer. Each layer operates with a hidden dimension of 256.

We believe this satisfies the reviewer’s request and makes the work reproducible without exceeding the page limit.

---

Weakness 7

We thank the reviewer for pointing this out. We have moved the essential training data description into the main text. The paragraph now states that PHYLA is trained on the 3k‑MSA subset of OpenProteinSet, with supervision coming from normalized Hamming distances computed on those MSAs, while the model itself sees only the raw, unaligned sequences. We believe this resolves the concern without inflating the manuscript.

---

Weakness 8

We thank the reviewer for flagging the ambiguity. The intended point is that PHYLA’s inputs are always raw, unaligned sequences; it never ingests gap‑padded MSAs at inference time. MSAs are used only to compute target distances (Hamming) during training. We have re‑phrased line 129 to:

“Unlike most prior methods, our approach does not rely on a multiple sequence alignment (MSAs are used solely to derive supervision distances, not as model inputs), …”

This clarification is also echoed in appendix information pulled into the main text in response to earlier comments.

---

Question 1

We thank the reviewer for pointing this out. PHYLA processes hundreds of unaligned sequences in a single forward pass. During training we then sample many quartets from that batch and apply the quartet loss, which rewards the model for making the correct topological decisions. This sequence context is enabled by the BiMamba inter‑sequence layers, a capability not present in prior PLMs. The new “Model architecture” and “Training details” paragraphs in the main text (moved from the appendix) now explain this step‑by‑step, so the distinction should be clear.

---

Question 2

We thank the reviewer for highlighting this. For every evaluation tree (TreeBase or TreeFam) we feed PHYLA the entire set of its sequences in one batch; the sequence order is arbitrary. Re‑ordering the sequences in the batch simply permutes the output rows, and pairwise distances changes minimally. We believe this happens because in training we sample arbitrary sub-trees of any order, so PHYLA in training sees that arbitrary ordering of the same sequences leads to the same quartet loss. We confirmed this empirically by shuffling the sequences of 250 random trees from Treebase fives times and taking the average standard deviation in resulting normRF for each tree and got 0.01. These details and numbers are now noted in the main text and appendix. Thus the embeddings and the resulting distance matrix are effectively deterministic for the purposes of evaluation.

---

Question 3

We thank the reviewer for highlighting this point. See response to earlier comment with ESM2 (650M) finetuning.

---

Question 4

We thank the reviewer for their suggestion. Existing benchmark suites (e.g., EC/GO function prediction, remote-homology classification) were built for models whose primary signal comes from residue-level co-occurrence patterns within individual sequences. They do not assess whether a model can reconstruct or reason over the evolutionary structure across sequences, which is the central question we address. To reflect this shift in emphasis, we introduced the Evolutionary-Reasoning Benchmark and related tasks tailored to multi-sequence inference. We believe the field benefits from recognizing this emerging class of evolution-focused PLMs and from developing corresponding evaluation protocols, just as earlier work developed function-oriented tests for classical PLMs. Pre-trained weights and code are public, so future studies can certainly explore hybrid objectives or broader task coverage; however, such experiments would not alter the paper’s key claims and contributions.

---

References

[1] Lin et al. Science 2023

[2] Hayes et al. Science 2025

[3] West-Roberts et al. BioArxiv 2024

[4] ESM Team, 2024

[5] Brixi et al. BioArxiv 2025

Weakness 1

We thank the reviewer for their suggestion. PHYLA is deliberately positioned as the start of a new class of protein language models, one that prioritises evolutionary reasoning (recovering accurate phylogenetic relationships) over traditional single-sequence objectives such as masked-token recovery. Existing benchmark suites (e.g., EC/GO function prediction, remote-homology classification, generative design) were built for models whose primary signal comes from residue-level co-occurrence patterns within individual sequences. They do not assess whether a model can reconstruct or reason over the evolutionary structure among sequences, which is the central question we address.

To reflect this shift in emphasis, we introduced the Evolutionary-Reasoning Benchmark and related tasks tailored to multi-sequence inference. We believe the field benefits from recognising this emerging class of evolution-focused PLMs and from developing corresponding evaluation protocols, just as earlier work developed function-oriented tests for classical PLMs. Pre-trained weights and code are public, so future studies can certainly explore hybrid objectives or broader task coverage; however, such experiments would not alter the paper’s key claims.

Weakness 2

We agree. At present no dataset, or set of reference trees, exhaustively represents evolutionary dynamics for every protein family. Because evolution is both open-ended and unevenly sampled, any benchmark necessarily covers only a subset of sequence space. Our benchmark is therefore a practical, but incomplete, proxy, an inherent limitation shared with all current phylogenetic evaluations. We will make this explicit in the manuscript’s limitations section, noting that expanding the benchmark as new data become available is an important direction for future work.

Weakness 3

We thank the reviewer for highlighting this point. In the Tree-of-Life case study we show that PHYLA’s output, when paired with a standard multiple-sequence alignment (MSA), highlights specific sequence features that drive the inferred topology. In that example, the alignment revealed a protein deletion that separates two clades. This workflow—tree → alignment → feature inspection—mirrors how phylogeneticists routinely move from a topology to biological hypotheses (e.g., gene loss, domain rearrangement) for downstream validation. We will clarify this interpretability pipeline in the results and discussion to make explicit that PHYLA, like traditional trees, serves as a starting point for generating testable biological insights beyond the topology itself.

Weakness 4

We thank the reviewer for raising this important point. PHYLA does not require expert-curated trees. To train PHYLA we need high-quality MSAs which can be generated by a slew of existing algorithms. We agree that these algorithms are computationally costly, however given a new domain with new sequencing data, it is possible to construct MSAs. Moreover, at inference time PHYLA operates on unaligned sequences; MSAs are needed only for training. We have also shown PHYLA can generalize to unseen sequences, so given a domain previously lacking sequencing data, PHYLA can be deployed effectively.

Weakness 5

We thank the reviewer for pointing out the missing benchmark. We have now evaluated PHYLA on masked-token prediction (MTP) on our training set. PHYLA trained only with MTP attains 30% top-1 accuracy with reduced evolutionary reasoning performance; training with tree loss decreases accuracy to 11% but improves evolutionary-reasoning performance (See Table 2 for evolutionary reasoning performance). Conventional PLMs (ESM-2 650 M, ProGen2) achieve 45–55%. These results confirm a trade-off: per-token objectives favour single-sequence reconstruction, whereas the tree loss preserves cross-sequence signals essential for evolutionary reasoning. We will incorporate this benchmark, training MTP curves, and a discussion in the manuscript.