acceptodds
Under review as a conference paper at ICLR 2027

Neopep: Reinforcement Learning From Machine-Learned Force Field Energy for Non-Canonical Peptide Structure Prediction

Abstract

Peptide therapeutics occupy a compelling middle ground between small molecules and larger proteins, and in practice, most approved peptide drugs incorporate non-canonical amino acids (ncAAs) or other non-peptidic chemistries to elicit improved physiochemical properties (Yaseen Khan et al. (2026)). The chemical space of known ncAAs numbers in the thousands, yet current state-of-theart sequence-to-structure prediction models handle them poorly, because fewer than 0.02% of residues deposited in the Protein Data Bank (PDB) (Berman et al.(2003)), the main source of training data, correspond to ncAAs. Our central idea is that this data bottleneck can be overcome by coupling generative structure prediction with reinforcement learning (RL) from machine-learned force field (MLFF)-derived energy. We present Neopep, a sequence-conditioned diffusion model for ncAA-containing peptide structure prediction, first pretrained on PDB structures to establish a structural prior and then fine-tuned via RL: for ncAA containing peptide sequences, the model proposes structures that are scored by a frozen MLFF, and this energy signal is backpropagated directly into the model to progressively lower the energy of its predictions. MLFFs have demonstrated near-quantum accuracy across much of chemical space, approacing DFT-level accuracy at a fraction of the computational cost (Kabylda et al. (2025)). This makes them substantially more accurate than classical, empirically derived force fields, which are parametrised for the canonical 20 amino acids and do not generalise to arbitrary chemistries. While Neopep is trained on a fixed set of 189 ncAAs, the underlying architecture operates at atom-level granularity and is in principle generalisable to any chemistry. We show that our model is able to produce structures of ncAA-containing peptides that achieve lower RMSD values on average to the ground truth structures, in comparison to without finetuning and other state-of-the-art structure prediction models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
machine learning for biology, ai for proteins, ML for protein design, peptide design, ML for peptide therapeutics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_H5szOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Neopep, a diffusion-based model for predicting the structures of peptides containing non-canonical amino acids (ncAAs). The model follows the AlphaFold3/Boltz architecture and is first pretrained on PDB-derived protein fragments. It is subsequently fine-tuned using DRaFT-K to minimize energies predicted by the SO3LR machine-learned force field. A differentiable hydrogen-placement module enables energy gradients to propagate through the predicted heavy-atom coordinates. The method is evaluated on 34 non-canonical and 24 canonical peptides. On the non-canonical set, fine-tuning improves minimum Cα-RMSD from 2.271 Å to 1.565 Å and average Cα-RMSD from 3.236 Å to 2.546 Å compared with the pretrained model. Additional comparisons are provided against Boltz-2 and Protenix-v1.

Critical strengths

1. Relevant problem with a practical motivation.

Accurate structure prediction for peptides containing ncAAs is challenging because experimental structural data are limited. Using a machine-learned force field to provide additional supervision is a reasonable direction, and the differentiable hydrogen-placement implementation is a potentially useful engineering contribution.

2. Encouraging improvements over the pretrained model.

The reported reductions in both Cα-RMSD and heavy-atom RMSD are meaningful. The improvements are not restricted to one example, and the pretrained-versus-fine-tuned comparison provides evidence that energy-guided optimization can improve the model's predictions on the proposed benchmark.

Critical weaknesses

1. Questionable physical assumptions. The model minimizes isolated-peptide energy without accounting for solvent or receptor interactions, although some test structures are receptor-bound. It is unclear whether this objective reliably favors experimentally observed conformations.

2. Potential error in hydrogen placement. Equation (12) appears inconsistent with the stated methylene geometry and may produce incorrect hydrogen bond angles. Since hydrogen coordinates directly affect the force-field gradients, this raises concerns about the reliability of the training signal.

3. Limited methodological novelty and missing baselines. The method primarily combines existing Boltz/AF3 architectures, DRaFT-K, and a pretrained force field. Comparisons against direct energy minimization or inference-time energy guidance are missing, making it difficult to establish the benefit of the proposed fine-tuning strategy.

4. Insufficient experimental validation. The benchmark contains only 34 non-canonical peptides, with Protenix evaluated on just 29. The emphasis on best-of-25 RMSD, lack of statistical significance tests, and absence of evaluation on unseen ncAA types weaken the claims of accuracy and generalization.

Questions for authors

Q1 Can the authors clarify Equation (12) and confirm whether the implementation follows the stated formula? What are the distributions of hydrogen bond-length and bond-angle errors after protonation, particularly for ncAAs?

Q2 Since several test peptides are extracted from receptor-bound complexes, how do the authors justify optimizing isolated-peptide energy to recover their bound conformations? Can the results be separated into bound versus unbound and cyclic versus linear peptides?

Q3 How does Neopep compare with applying SO3LR-based energy minimization or inference-time guidance directly to the pretrained model, under comparable computational budgets?

Q4 Can the authors provide paired statistical comparisons on the common 29-peptide subset, results across independent training seeds, and evaluations on genuinely unseen ncAA types? Can they also report stereochemical validity and cyclic bond-closure accuracy?

15 comments
AuthorsOpenReview ↗

Response to Reviewer H5sz

We thank the reviewer for taking the time to read our manuscript and for the constructive feedback. Below, we respond to the reviewers first question, and address the related weaknesses (Part 1 of 2):

Can the authors clarify Equation (12) and confirm whether the implementation follows the stated formula? Related Weaknesses: Equation (12) appears inconsistent with the stated methylene geometry and may produce incorrect hydrogen bond angles. Since hydrogen coordinates directly affect the force-field gradients, this raises concerns about the reliability of the training signal.

We thank the reviewer for catching this error in the manuscript. The previous form of Eq. 12,

h1,2=C+d(sin⁡αu±cos⁡αv),\mathbf{h}_{1,2} = \mathbf{C} + d(\sin\alpha\mathbf{u} \pm \cos\alpha\mathbf{v}),

did not match the implementation. Because u\mathbf{u} points toward the two heavy-atom substituents, the positive u\mathbf{u} coefficient placed both hydrogens on the wrong side of the carbon. With α\alpha defined as the half-opening angle, the sine and cosine terms were also swapped, which does not give tetrahedral angles. We have corrected Eq. 12 in the paper to

h1,2=C+d(−cos⁡αu±sin⁡αv),\mathbf{h}_{1,2} = \mathbf{C} + d(-\cos\alpha\mathbf{u} \pm \sin\alpha\mathbf{v}),

where α\alpha is still the half-opening angle derived from the ideal H–C–H angle. We confirm that the implementation in the code is correct (no changes needed), and only correct the formula in the manuscript. This updated equation places the geminal hydrogens opposite the substituents, symmetric about the −u-\mathbf{u} axis, and recovers tetrahedral geometry.

AuthorsOpenReview ↗

Response to Reviewer H5sz

We thank the reviewer for taking the time to read our manuscript and for the constructive feedback. Below, we respond to the reviewers first question (Part 2 of 2):

What are the distributions of hydrogen bond-length and bond-angle errors after protonation, particularly for ncAAs?

The SO3LR reward is evaluated on all-atom structures, so every prediction is protonated by the differentiable rule-based protonator of Section C.4 before its energy is computed. In the appendix section E.2 (Hydrogen Bond-Length and Bond-Angle Error of the Protonator), we quantify the geometric error of that protonation directly against the CCD, for every non-canonical residue type that occurs in the test set.

For each of the 95 non-canonical residue types present in the 34 non-canonical test peptides, we take the residue’s ideal CCD conformer, which carries CCD-assigned hydrogen positions, discard its hydrogens, and run the protonator on the remaining heavy atoms exactly as it is run during fine-tuning.

Each hydrogen placed by the protonator is then compared with the corresponding CCD hydrogen. Both hydrogen sets are built on identical heavy-atom coordinates, so any difference reflects the protonator’s placement rules alone and not the structure predictor. We compare (i) the X–H bond length to the parent heavy atom X and (ii) the H–X–Y bond angle for every heavy neighbour Y of X. Leaving atoms (OXT, HXT, and the second amine hydrogen of a free N-terminus) are not placed by the protonator and are excluded, as our current protonator protonates residues as if they are within the peptide chain. Figure 9 in appendix section E.2 of the updated manuscript shows the distribution of angle errors. 98.2% of angles are within 5◦. The bond-length error is identically zero. The protonator takes every X–H distance from the CCD and determines only the direction of the bond, so this comparison is satisfied by construction.

AuthorsOpenReview ↗

Response to Reviewer H5sz

We thank the reviewer for taking the time to read our manuscript and for the constructive feedback. Below, we respond to the reviewers second question:

Since several test peptides are extracted from receptor-bound complexes, how do the authors justify optimizing isolated-peptide energy to recover their bound conformations?

We agree with the reviewer that solvent and receptor interactions (bound vs. unbound states) are important things our model should take into account. We acknowledge that simply training on isolated peptide energies might not be able to model bound/unbound states, and do not claim that it does so. What we do hope, is that the model has learned a rough prior on what structures of peptides in solution, or bound/unbound conformations look like during phase-1 training on the PDB, and is able to refine this prior with the additional energy training (and not stray away too much from it, but just enough so that the predicted structures with ncAAs are plausible). Of course, this is only testable once we curate a more comprehensive benchmark with the limited data that we have. We would also like to note that one could, in principle, replace the phase-1 prior with a more SOTA structure predictor like Protenix or AF3.

Can the results be separated into bound versus unbound and cyclic versus linear peptides?

In appendix section E.3 (Results Split by Binding State and Cyclisation) of the updated manuscript, we split the main-text results for the ncAA peptides (from Section 4.4, tables 1 and 2) along both axes. We also provide statistical tests for this in appendix section F.2 (tables 16 and 17). Cyclisation is taken from the inference input (17 cyclic, 17 linear); binding state from the deposited entry, where 28 of the 34 peptides are resolved bound to a receptor and 6 are unbound. We share the tables below:

Table 6. CA-RMSD (Å) on the non-canonical test set, split by cyclic versus linear peptides. Mean ± std across peptides. Bold indicates the best value in each column. Protenix does not support 5 of the 34 peptides and is evaluated on fewer peptides in each group (nn given in the row).

Cyclic (n=17n{=}17)Linear (n=17n{=}17)
ModelMin CA-RMSDAvg CA-RMSDMin CA-RMSDAvg CA-RMSD
Boltz-24.677 ± 1.6808.302 ± 3.8184.219 ± 1.5209.793 ± 6.865
Protenix-v1 †1.964 ± 1.2992.529 ± 1.3752.573 ± 1.3163.183 ± 1.474
Neopep Pretrained1.889 ± 0.5032.747 ± 0.4782.653 ± 0.9383.725 ± 1.075
  + SO3LR relaxation1.900 ± 0.5152.770 ± 0.4922.687 ± 0.9503.757 ± 1.092
Neopep Finetuned1.228 ± 0.7461.942 ± 0.7571.901 ± 1.0463.150 ± 1.195

† Evaluated on n=15n=15 cyclic and n=14n=14 linear peptides.

Table 7. CA-RMSD (Å) on the non-canonical test set, split by bound versus unbound peptides. Mean ± std across peptides. Bold indicates the best value in each column. Protenix does not support 5 of the 34 peptides and is evaluated on fewer peptides in each group (nn given in the row).

Bound (n=28n{=}28)Unbound (n=6n{=}6)
ModelMin CA-RMSDAvg CA-RMSDMin CA-RMSDAvg CA-RMSD
Boltz-24.366 ± 1.4959.513 ± 5.8974.831 ± 2.1226.877 ± 2.570
Protenix-v1 †2.261 ± 1.3012.892 ± 1.3862.248 ± 1.5222.665 ± 1.750
Neopep Pretrained2.094 ± 0.7152.990 ± 0.6443.099 ± 0.9294.384 ± 1.377
  + SO3LR relaxation2.112 ± 0.7243.014 ± 0.6603.140 ± 0.9564.428 ± 1.385
Neopep Finetuned1.319 ± 0.8062.248 ± 0.9342.711 ± 0.7933.934 ± 1.170

† Evaluated on n=23n=23 bound and n=6n=6 unbound peptides.

Table 8. Minimum heavy-atom RMSD (FA-RMSD, Å) on the non-canonical test set, split by cyclic versus linear peptides. Mean ± std across peptides. Bold indicates the best value in each column.

ModelCyclic (n=17n{=}17)Linear (n=17n{=}17)
Boltz-25.971 ± 1.5435.666 ± 1.398
Protenix-v1 †3.412 ± 1.4203.725 ± 1.722
Neopep Pretrained3.434 ± 0.6664.035 ± 0.835
  + SO3LR relaxation3.485 ± 0.6774.094 ± 0.852
Neopep Finetuned2.918 ± 0.7253.470 ± 1.072

† Evaluated on n=15n=15 cyclic and n=14n=14 linear peptides.

Table 9. Minimum heavy-atom RMSD (FA-RMSD, Å) on the non-canonical test set, split by bound versus unbound peptides. Mean ± std across peptides. Bold indicates the best value in each column.

ModelBound (n=28n{=}28)Unbound (n=6n{=}6)
Boltz-25.855 ± 1.3995.648 ± 1.847
Protenix-v1 †3.633 ± 1.4553.293 ± 2.018
Neopep Pretrained3.583 ± 0.7394.441 ± 0.767
  + SO3LR relaxation3.636 ± 0.7474.505 ± 0.812
Neopep Finetuned2.974 ± 0.8454.219 ± 0.702

† Evaluated on n=23n=23 bound and n=6n=6 unbound peptides.

Reviewer_FGmiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Neopep addresses structure prediction for peptides containing non-canonical amino acids by pretraining a reduced AF3/Boltz-style diffusion model on PDB fragments, then fine-tuning it with SO3LR force-field energy through DRaFT-K. A differentiable hydrogen-placement module supports 189 non-canonical residue types and connects the heavy-atom predictor to the all-atom energy. On 34 temporally and sequence-filtered non-canonical peptides, fine-tuning improves mean best-of-25 Cα RMSD from 2.271 to 1.565 Å and mean Cα RMSD from 3.236 to 2.546 Å; a separate 24-peptide canonical set also shows improved best-of-25 accuracy.

Critical strengths
  • The improvement over the same pretrained model is substantial and addresses a useful data-scarce setting without requiring additional experimental structures, especially in best-of-25 Cα RMSD and best-of-25 heavy-atom RMSD.
  • Reporting average Cα RMSD as well as oracle-selected minima strengthens the performance of the proposed method, and the canonical control provides evidence that the adaptation does not simply sacrifice standard peptide prediction.
  • The problem of structure prediction for biomolecules with non-canonical amino acids is important and under explored.
Critical weaknesses
  • The released fine-tuning code appears to leave linear peptides with incomplete terminal atoms. In phase2_finetuning/train.py, the hydrogen builder receives is_cyclic=True even for linear samples. The separate linear_n_term option changes the first N–H reference frame but does not restore terminal atoms; the parser and builder omit the terminal oxygen OXT, and the sampler strips terminal caps.
  • A minimal execution of this path for four alanines produces C12H20N4O4, rather than C12H22N4O5 for an uncapped linear tetrapeptide. These atoms are passed directly to SO3LR, so the omission could affect the energies and gradients used for training. This check used a public CCD record; the authors' exact CCD cache and historical code version were unavailable. It establishes a concern in the supplied code, but not the impact on the reported results. The intended terminal chemistry and actual reward inputs need validation.
  • The experiments do not establish how much fine-tuning adds beyond directly relaxing the pretrained predictions with the same SO3LR potential. This is consequential because the qualitative examples emphasize repairing structural breaks, while the energy evaluation uses the same potential as the training reward. Those results do not distinguish the benefits of learned adaptation from improvements obtainable through direct geometry refinement.
  • Figure 2 adds little independent validation when the method is fine-tuned on SO3LR and is being evaluated on SO3LR. In this way, this part of the evaluation is self-fulfilling.
  • The paper credits DRaFT-K, so its contribution rests on demonstrating the value of this adaptation for non-canonical peptides. Comparing fine-tuning with pretrained-plus-SO3LR relaxation, in both structural accuracy and computational cost, would clarify that value.
Questions for authors
  1. For linear peptides, what terminal chemistry was used in the reported runs, and where are the missing terminal atoms restored? If the released forced-cyclic builder was used, how do energy, gradients, and structural accuracy change after completing the intended termini? Separately, please correct Eq. 12 to match the implemented methylene rule, which uses signed CCD coefficients and recovers tetrahedral geometry.

  2. On the non-canonical test set, how does directly relaxing the same pretrained predictions with SO3LR compare with fine-tuning in Cα/heavy-atom RMSD? Please compare both the fine-tuning cost and the per-prediction inference/refinement cost to clarify when updating the model is preferable to refining its outputs.

14 comments
AuthorsOpenReview ↗

Response to Reviewer FGmi

We thank the reviewer for taking the time to read our manuscript and for the constructive feedback. Below, we respond to the reviewers first question, and address the related weaknesses:

For linear peptides, what terminal chemistry was used in the reported runs, and where are the missing terminal atoms restored? If the released forced-cyclic builder was used, how do energy, gradients, and structural accuracy change after completing the intended termini?

Related weaknesses: The released fine-tuning code appears to leave linear peptides with incomplete terminal atoms. In phase2_finetuning/train.py, the hydrogen builder receives is_cyclic=True even for linear samples. The separate linear_n_term option changes the first N–H reference frame but does not restore terminal atoms; the parser and builder omit the terminal oxygen OXT, and the sampler strips terminal caps. A minimal execution of this path for four alanines produces C12H20N4O4, rather than C12H22N4O5 for an uncapped linear tetrapeptide. These atoms are passed directly to SO3LR, so the omission could affect the energies and gradients used for training. This check used a public CCD record; the authors' exact CCD cache and historical code version were unavailable. It establishes a concern in the supplied code, but not the impact on the reported results. The intended terminal chemistry and actual reward inputs need validation.

Terminal atoms: We follow the atom representation of Boltz-2, on which our model is built. Boltz-2 does not predict leaving atoms (Passaro et al., 2025): the C-terminal OXT (oxygen leaving atom) is not part of its output. Our model inherits this convention, so predicted coordinates for linear peptides in the reported runs contain no OXT, and after protonation before SO3LR evaluation, does not have the HXT (hydrogen on OXT) and the leaving H atom on the N-terminal nitrogen atom. I.e., the terminal residues (on the C- and N- terminii) are represented as if they are within a peptide chain. We definitely agree that terminal atoms are important for computing an accurate energy of the system, however, we'd like to offer some reasons as to why we kept this decision/didn't feel the need to change it: (1) this convention of excluding the terminal OXT, HXT, and N terminal H is applied identically to all predictions of all the baselines, in all energy/RMSD comparisons reported in the paper, so those comparisons are like-for-like. (2) We like to clarify that the main contribution of this work is the Phase-2 fine-tuning scheme (which, from our results, shows an improvement in RMSD to ground truth despite the exclusion of the terminal atoms), so in principle, one could replace the Phase-1 model with any state-of-the-art model that actually predicts leaving atoms, and we hypothesize that a similar decrease in RMSD would be observed.

Terminal caps: The reviewer is also correct that the sequence sampler removes N-terminal ACE and C-terminal NH2 caps during fine tuning (and peptides in the test set are thus stripped of both, if any). This was a deliberate scoping choice- we wanted to train and evaluate on the peptide backbone and its non-canonical residues alone, without caps that would otherwise appear on some sequences and not others, so that differences in accuracy across peptides reflect the residues themselves. We agree that caps contribute to the energy of the structure, and argue that they (as well as other chemical groups in the CCD) can be brought within scope the same way as the other ncAAs by adding their own hydrogen builder during protonation before SO3LR energy evaluation during finetuning.

Separately, please correct Eq. 12 to match the implemented methylene rule, which uses signed CCD coefficients and recovers tetrahedral geometry.

We thank the reviewer for catching this error in the manuscript. The previous form of Eq. 12,

h1,2=C+d(sin⁡αu±cos⁡αv),\mathbf{h}_{1,2} = \mathbf{C} + d(\sin\alpha\mathbf{u} \pm \cos\alpha\mathbf{v}),

did not match the implementation. Because u\mathbf{u} points toward the two heavy-atom substituents, the positive u\mathbf{u} coefficient placed both hydrogens on the wrong side of the carbon. With α\alpha defined as the half-opening angle, the sine and cosine terms were also swapped, which does not give tetrahedral angles. We have corrected Eq. 12 in the paper to

h1,2=C+d(−cos⁡αu±sin⁡αv),\mathbf{h}_{1,2} = \mathbf{C} + d(-\cos\alpha\mathbf{u} \pm \sin\alpha\mathbf{v}),

where α\alpha is still the half-opening angle derived from the ideal H–C–H angle, and confirm that the code implementation used for all experiments/training in the paper is correct, and did not need any changes. This updated equation places the geminal hydrogens opposite the substituents, symmetric about the −u-\mathbf{u} axis, and recovers tetrahedral geometry.

AuthorsOpenReview ↗

Response to Reviewer FGmi

We thank the reviewer for taking the time to read our manuscript and for the constructive feedback. Below, we respond to the reviewers second question (Part 1 of 2):

On the non-canonical test set, how does directly relaxing the same pretrained predictions with SO3LR compare with fine-tuning in Cα/heavy-atom RMSD?

Related Weaknesses: The experiments do not establish how much fine-tuning adds beyond directly relaxing the pretrained predictions with the same SO3LR potential. This is consequential because the qualitative examples emphasize repairing structural breaks, while the energy evaluation uses the same potential as the training reward. Those results do not distinguish the benefits of learned adaptation from improvements obtainable through direct geometry refinement. Figure 2 adds little independent validation when the method is fine-tuned on SO3LR and is being evaluated on SO3LR. In this way, this part of the evaluation is self-fulfilling. The paper credits DRaFT-K, so its contribution rests on demonstrating the value of this adaptation for non-canonical peptides. Comparing fine-tuning with pretrained-plus-SO3LR relaxation, in both structural accuracy and computational cost, would clarify that value.

We ran exactly this comparison (directly relaxing the same pretrained predictions with SO3LR) only on the non-canonical test set, and have updated the manuscript to include it: the tables in the main text (Section 4.4 Results tables 1 and 2) contain an additional row for this relaxation baseline, reporting the CA-RMSD and FA-RMSD averaged over all non-canonical test peptides (mean ± std across peptides, min CA-RMSD is computed over the average of all the min CA-RMSDs across the test peptides, same for min FA-RMSD). More details on the exact procedure are included in appendix section E.1 of the updated manuscript (Additional Results: Inference-time SO3LR Relaxation), and statistical tests are reported in table 14 of appendix section F.1 (Statistical Tests: RMSD Comparisons), but we summarize the procedure and key findings here:

Procedure: We took the same 25 pretrained predictions for each of the 34 non-canonical test peptides, protonated each once with the protonator used during fine-tuning, and relaxed it for 200 steps of steepest descent on the SO3LR potential with all atoms free (step size 0.01, per-atom displacement capped at 0.1 Å).

Findings: We found that relaxation for 200 steps lowers energy (as intended), but does not improve accuracy: The SO3LR energy falls on every one of the 850 (34 x 25) structures, from −1.094 to −1.672 eV/atom on average (Appendix E.1, figure 8). But accuracy (in terms of RMSD) gets slightly worse. Statistical tests in table 14 of appendix section F.1 confirm the significance of these results, but we present the main min/avg CA-RMSD and min FA-RMSD tables here:

Table 1. CA-RMSD (Å) averaged over all test peptides (mean ± std across peptides; min CA-RMSD is computed over the average of all the min CA-RMSDs across the test peptides). Bold indicates the best value in each column. Protenix evaluated on n=29n=29 non-canonical peptides (5 excluded due to missing CCD support); all other models evaluated on n=34n=34.

Non-canonical (n=34n{=}34)Canonical (n=24n{=}24)
ModelMin CA-RMSDAvg CA-RMSDMin CA-RMSDAvg CA-RMSD
Boltz-24.448 ± 1.5959.047 ± 5.5222.744 ± 2.6053.438 ± 2.955
Protenix-v1 †2.258 ± 1.3202.845 ± 1.4373.005 ± 3.0223.738 ± 3.039
Neopep Pretrained2.271 ± 0.8373.236 ± 0.9583.046 ± 2.7653.847 ± 3.105
+ SO3LR relaxation2.293 ± 0.8523.263 ± 0.973——
Neopep Finetuned1.565 ± 0.9572.546 ± 1.1602.211 ± 1.9273.768 ± 3.078

† Evaluated on n=29n=29 non-canonical peptides.

Table 2. Minimum heavy-atom RMSD (FA-RMSD, Å) averaged over all test peptides (mean ± std across peptides). Bold indicates the best value in each column. Protenix evaluated on n=29n=29 non-canonical peptides; all other models on n=34n=34.

ModelMin FA-RMSD, Non-canonical (n=34n{=}34)Min FA-RMSD, Canonical (n=24n{=}24)
Boltz-25.818 ± 1.4584.035 ± 2.550
Protenix-v1 †3.563 ± 1.5534.331 ± 2.922
Neopep Pretrained3.735 ± 0.8044.037 ± 2.800
+ SO3LR relaxation3.789 ± 0.818—
Neopep Finetuned3.194 ± 0.9443.456 ± 1.917

† Evaluated on n=29n=29 non-canonical peptides.

We hypothesize that the lack of improvement from inference-time relaxation might be due to the fact that relaxation is a purely local procedure- gradient descent from a fixed starting conformation can only reach the nearest minimum, so it is able to repair local geometry like bond lengths, angles, clashes, etc but preserves larger errors that would potentially require crossing a barrier, such as backbone breaks or misplaced side-chains.

AuthorsOpenReview ↗

Response to Reviewer FGmi

We thank the reviewer for taking the time to read our manuscript and for the constructive feedback. Below, we respond to the reviewers second question (Part 2 of 2):

Please compare both the fine-tuning cost and the per-prediction inference/refinement cost to clarify when updating the model is preferable to refining its outputs.

Cost comparisons. Fine-tuning and relaxation differ in where their cost (in terms of # of gradient steps) sits. Relaxation carries no setup cost but a per-prediction one: each peptide is an independent optimization on the potential energy surface, and the compute spent on it is discarded afterwards. Total cost grows linearly in the number of predictions. Fine-tuning is a one-time cost paid during training, after which inference is identical to the pretrained model’s, with no force-field evaluations at all. Total cost is therefore effectively constant in the number of predictions. We reason that if one were to scale up both the number of predictions and the number of gradient steps taken for optimizing a prediction (for the case of inference-time optimization), fine-tuning is the more cost-effective option.

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

This paper introduces a model for prediction of structures containing non-canonical amino acids (ncAAs). The main contribution is the use of reinforcement learning with a reward function given by a frozen machine learning force field (MLFF) energy to finetune a pretrained trunk, a simplified Boltz2-style diffusion model. First, the trunk is pretrained on protein fragments of 3-50 residues from the Protein Data Bank (PDB). Then, in the fine-tuning step, the trunk is further trained to minimise the energy from the MLFF on a new synthetic dataset where peptide sequences from the pretraining step are randomly mutated to an ncAA. The resulting model, Neopep, is evaluated on a set of 34 ncAA and a set of 24 canonical peptides, both constructed from PDB entries after the training cutoff date. Neopep is compared with 3 baseline models: Boltz-2, which does not explicitly handle ncAAs; Protenix-v1, a state-of-the-art predictor that can handle ncAAs; and a version of Neopep that has been pretrained in the same way as the main model, without the MLFF finetuning. Neopep is shown to outperform all 3 baselines in terms of minimum and average RMSD.

Critical strengths

The idea of using a machine learning force field to provide the reward function for a reinforcement learning-based finetuning pipeline in the context of protein structure prediction is particularly relevant for non-canonical amino acids, where there is not enough data available. This has great potential as MLFFs improve in the future.

Critical weaknesses

None

Questions for authors
  1. The main concern is related to the particular MLFF chosen here, SO3LR. It is clear that the performance of the overall approach is dependent on the MLFF used, so I wonder what happens if the underlying MLFF has not seen in training structures similar to the non-canonical amino acids Neopep is evaluated on. Is this something that has been taken into account? Is there any analysis that can quantify the effect of such a mismatch?

  2. Since the authors mention AlphaFold3, it would be interesting to see if its performance on the evaluation datasets is different from Boltz2. Did the authors run it? If so, how did it perform?

  3. Protenix-v1 was only run on a subset of 29 of the full set of 34 ncAA, how does Neopep compare only on those 29? Currently the numbers in Table 1 are for the full set, which does not say anything about how Neopep performs on exactly the same ones compared to Protenix-v1. Does its improvement come mostly from the other 5 ncAAs, or also from the 29 ones?

Some minor issues:

  • Enumerated references look weird, I think they should be written as "\citep{ref1, ref2}" to appear "(ref 1, year1; ref2, ear2)"
  • In Algorithm 1, line 9: what is q in the SO3LR call?
  • Line 250: "minimize Eq. equation 2"
  • Line 253: "As shown in Figure 1b and Algorithm 1, we use DRaFT-K ..." I don't think Figure 1b shows that
  • Figure 2 legend and labels: What is "FT-9500"? Is that the main finetuned Neopep model?
6 comments
AuthorsOpenReview ↗

Response to Reviewer ZkuM

We thank the reviewer for taking the time to read our manuscript and for the constructive feedback. Below, we respond to the reviewers first question:

The main concern is related to the particular MLFF chosen here, SO3LR. It is clear that the performance of the overall approach is dependent on the MLFF used, so I wonder what happens if the underlying MLFF has not seen in training structures similar to the non-canonical amino acids Neopep is evaluated on. Is this something that has been taken into account? Is there any analysis that can quantify the effect of such a mismatch?

We thank the reviewer for this important point. We agree that Neopep's performance depends on the fidelity of the underlying MLFF. While peptides containing non-canonical amino acids (ncAAs) are not explicitly represented in SO3LR's training set, the specific chemical groups that make the ncAAs different from cAAs (canonical amino acids) are contained in its training set (and the cAAs are plentiful in the training set of the Phase-1 model), so we expect SO3LR to be quite accurate on them (Kabylda et al., 2025).

Additionally, before we began fine-tuning, we took the pretrained predictions (from phase-1) for a small number of validation-set peptides that had NMR structures, and compared their SO3LR energies. We were specifically checking whether the experimental conformers scored lower (in terms of energy) than our pretrained predictions, i.e. a relative comparison between the two sets rather than an absolute one. For the 3–4 peptides examined this was the case, indicating that the potential ranks the native conformation below the prior’s samples on ncAA-containing peptides, and we proceeded with fine-tuning on that basis. We report the results of this small test in our updated manuscript in appendix section E.6. We note that this was a small directional check rather than a thorough validation. As for an analysis that would quantify the effect of a mismatch more generally, we are unsure, but hope that as MLFFs continue to improve in the future, we take them as computable verifiers.

AuthorsOpenReview ↗

Response to Reviewer ZkuM

We thank the reviewer for taking the time to read our manuscript and for the constructive feedback. Below, we respond to the reviewers second question:

Since the authors mention AlphaFold3, it would be interesting to see if its performance on the evaluation datasets is different from Boltz2. Did the authors run it? If so, how did it perform?

We did not run AlphaFold3 itself. As our main contribution is the fine-tuning framework, one could in principle replace the Phase-1 prior with a different model (like Protenix or AF3). Our pretrained model is deliberately reduced in size relative to Boltz-2 (due to compute constraints, training data provenance, etc) and is not intended as a competitive standalone predictor. The comparison that isolates our claim is therefore pretrained vs fine-tuned. We do still include Boltz-2 and Protenix-v1 to situate that result against current open predictors, and dedicate future efforts in replacing the Phase-1 prior with a more SOTA structure predictor.

AuthorsOpenReview ↗

Response to Reviewer ZkuM

We thank the reviewer for taking the time to read our manuscript and for the constructive feedback. Below, we respond to the reviewers third question:

Protenix-v1 was only run on a subset of 29 of the full set of 34 ncAA, how does Neopep compare only on those 29? Currently the numbers in Table 1 are for the full set, which does not say anything about how Neopep performs on exactly the same ones compared to Protenix-v1. Does its improvement come mostly from the other 5 ncAAs, or also from the 29 ones?

We thank the reviewer for the question- we have added paired statistical tests in the updated manuscript in Appendix section F.2 (tables 14 for the non-canonical set and 15 for the canonical set), with all Protenix comparisons computed on exactly the 29 peptides, for the non-canonical set. On that matched set, Neopep Finetuned attains significantly lower Min CA-RMSD (p = 0.002) and Min FA-RMSD (p = 0.030) than Protenix-v1, while Avg CA-RMSD is on the edge of statistical significance (p = 0.078). We would like to note that our main contribution is the fine-tuning framework, and one could in principle replace the Phase-1 prior with a different SOTA structure predictor (like Protenix itself, AF3, etc). We summarise the procedure and results below:

For each metric (Avg CA-RMSD, Min CA-RMSD, Min FA-RMSD), the Neopep finetuned model (FT-9500) is compared against each baseline with a one-sided Wilcoxon signed-rank test paired by test peptide, Holm-corrected across four baselines within each metric. Table below report the statistical tests for just the non-canonical test set:

Paired Wilcoxon signed-rank tests on the non-canonical test set, testing whether FT-9500 attains lower RMSD than each baseline (for each test peptide, we take the mean RMSD (or min, for min CA-RMSD and min FA-RMSD) over the 25 predictions). One-sided, Holm-corrected across the four baselines within each metric. * denotes p<0.05p<0.05. rrbr_{rb} is the matched-pairs rank-biserial correlation, where −1-1 indicates FT-9500 is lower on every peptide. PT + SO3LR is the pretrained model followed by inference-time relaxation: each of the 25 samples is protonated once and relaxed for 200 steps on the SO3LR potential-energy surface with all atoms able to be moved. Protenix does not support 5 non-canonical peptides, so those comparisons use n=29n=29.

MetricBaselinennpprrbr_{rb}
Min CA-RMSDBoltz-234<10−7<10^{-7} *−0.99-0.99
Pretrained34<10−7<10^{-7} *−0.94-0.94
PT + SO3LR34<10−7<10^{-7} *−0.96-0.96
Protenix290.0020.002 *−0.60-0.60
Avg CA-RMSDBoltz-234<10−7<10^{-7} *−1.00-1.00
Pretrained341.0×10−61.0\times10^{-6} *−0.87-0.87
PT + SO3LR345.6×10−75.6\times10^{-7} *−0.89-0.89
Protenix290.0780.078−0.31-0.31
Min FA-RMSDBoltz-234<10−7<10^{-7} *−0.97-0.97
Pretrained342.8×10−62.8\times10^{-6} *−0.84-0.84
PT + SO3LR341.5×10−61.5\times10^{-6} *−0.87-0.87
Protenix290.0300.030 *−0.40-0.40

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.