- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
AbEvolve introduces a framework for aligning protein language models with evolutionary selection signals extracted from natural antibody repertoires. The approach uses clonal abundance as a proxy for fitness and constructs mutation trajectories to train models through reward-weighted maximum likelihood and GFlowNet alignment. The authors demonstrate improved zero-shot prediction of antibody binding and thermostability across multiple benchmarks and model architectures.
Critical strengths
The central strength is the clever use of clonal abundance as an essentially free and massively scalable source of functional supervision. Although abundance is an imperfect proxy for fitness, the authors demonstrate that these noisy signals can meaningfully improve protein language models. The consistent improvements across multiple benchmarks and architectures, together with the ablation studies, provide encouraging support for the approach.
Critical weaknesses
My main concern is the interpretation of the reconstructed evolutionary trajectories. The inferred mutation order does not necessarily reflect actual evolutionary history, and it remains unclear how much additional information the GFlowNet approach captures beyond abundance-weighted sequence preferences. However, I do not consider this a fundamental limitation of the central contribution.
Questions for authors
Can the authors further clarify what additional information is captured by GFlowNet alignment compared with simpler preference-learning approaches? In particular, would a pairwise preference-learning objective using the same abundance-derived signals achieve similar improvements?
We thank the reviewer. The concern and the question raise three points: whether the reconstructed order is an evolutionary history, what GFlowNet alignment adds beyond an abundance-weighted sequence preference, and whether a pairwise objective on the same abundance signal gives a similar improvement. The comparison of Trajectory Balance, path-level RW-MLE, and endpoint RW-MLE is in the comment Trajectory Balance versus a reward-weighted path likelihood. The pairwise result is the comment Response for DPO alignment.
1. The reconstructed order is not an evolutionary history.
FGMO orders substitutions by how often they occur in the lineage and inserts a virtual node when an intermediate is unobserved. On the 2.6K-lineage ablation, with the GFlowNet objective held fixed, FGMO is stronger than reverse-FGMO. We will describe it as a frequency heuristic of the construction, and we will explore better ways of building these trajectories in future work.
2. What GFlowNet adds beyond an abundance-weighted sequence preference.
The setup and the table are in the comment cited above. Path-level RW-MLE and endpoint RW-MLE use the same softmax weight. The gain comes from modeling the edited positions: the path loss applies that weight only to the edit steps, whereas the endpoint loss applies it to every token, including the background shared with the parent. The means are 0.362 and 0.327. Trajectory Balance, on the same fixed paths, reaches a similar mean to path-level RW-MLE (0.367 vs. 0.362). The flow-matching terms do not add a consistent further gain.
3. Would a pairwise objective give a similar improvement?
We thank the reviewer for suggesting DPO. The setting and the table are in the comment “Response for DPO alignment” . DPO can also emphasize the edited positions. The loss depends on the log-probability gap between the chosen and rejected sequences, so positions shared by the pair contribute little, and the update concentrates on the substitutions that distinguish them. This is a pairwise contrast rather than an explicit path, but it is the same reason the path objectives help: the abundance signal is applied to the edits instead of to the shared background. DPO reaches a similar eight-benchmark mean to the path objectives (0.362, versus 0.367 for TB and 0.362 for path-level RW-MLE). The gain is concentrated on binding, where DPO is the best of these objectives (mean 0.287, versus 0.263 for TB and 0.254 for endpoint RW-MLE).