Organizations: BioGeometry · School of Computer Science, Peking University · Mila - Québec AI Institute · Université de Montréal · HEC Montréal · CIFAR AI Research Chair
Protein sequence design must balance designability, defined as the ability to recover a target backbone, with multiple, often competing, developability properties such as solubility, thermostability, and expression. Existing approaches address these properties through post hoc mutation, inference-time biasing, or retraining on property-specific subsets, yet they are target dependent and demand substantial domain expertise or careful hyperparameter tuning. In this paper, we introduce ProtAlign, a multi-objective preference alignment framework that fine-tunes pretrained inverse folding models to satisfy diverse developability objectives while preserving structural fidelity. ProtAlign employs a semi-online Direct Preference Optimization strategy with a flexible preference margin to mitigate conflicts among competing objectives and constructs preference pairs using in silico property predictors. Applied to the widely used ProteinMPNN backbone, the resulting model MoMPNN enhances developability without compromising designability across tasks including sequence design for CATH 4.3 crystal structures, de novo generated backbones, and real-world binder design scenarios, making it an appealing framework for practical protein sequence design.
Figures & tables
Figure 1: The ProtAlign framework. ProtAlign optimizes the policy model in a semi-online regime composed of alternating rollout and training stages. In the rollout stage, protein backbones are sampled from the training set, and the current policy model generates rollouts at a higher temperature. These rollouts are evaluated with property predictors, and pairwise preference datasets are constructed for each property. During training, pairwise entries are drawn evenly across the datasets, and an adaptive preference margin is introduced to resolve conflicts among multiple objectives.
Method
Designability Metrics
Developability Metrics
AAR ↑
RMSD ↓
TM score ↑
PLDDT ↑
EP ↓
Sol ↑
Thermo ↑
Test Dataset
3.97
0.761
80.8
5.80
0.620
0.246
1.000
ESM-IF
4.36
0.737
78.4
6.11
0.733
0.719
0.464
InstructPLM 1 1 1 InstructPLM was trained on the 4.2 version of the CATH dataset. (default)
6.81
0.628
73.4
7.97
0.653
0.396
0.574
InstructPLM (T=0.1)
6.96
0.632
74.4
7.31
0.657
0.455
0.584
ProteinMPNN
4.30
0.740
79.1
6.70
0.719
0.769
0.389
Table 1: Comparison of protein sequence design methods on the CATH 4.3 test set across various metrics. Results for our RL-based MoMPNN trained with different annotator combinations (TM, IG, EP, Sol, and Thermo) are shown. The best and second-best values are highlighted in bold and underlined , respectively.
Method
Designability Metrics
Developability Metrics
RMSD ↓
TM score ↑
PLDDT ↑
EP ↓
Sol ↑
Thermo ↑
ESM-IF
13.51
0.461
57.6
7.27
0.616
0.806
InstructPLM (default)
22.44
0.134
32.6
3.33
0.539
0.278
InstructPLM (T=0.1)
22.58
0.132
33.5
2.73
0.538
0.367
ProteinMPNN
6.86
0.718
70.0
8.32
0.731
0.978
ProteinDPO
16.77
0.296
43.5
14.70
0.596
0.145
Table 2: Comparison of protein sequence design methods across different evaluation metrics on de novo backbone structures from RFDiffusion. The best results and the second-best results are marked bold and underlined . Note that AAR is not evaluated in this setting, as alignment to reference sequences is not applicable in de novo design.
Figure 2: The result for ProteinMPNN, SolubleMPNN and MoMPNN on the binder design benchmark.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Quantative Analysis of ProteinMPNN, SolubleMPNN and MoMPNN generated sequences on hydrophilic-related metrics.
Figure 4: Differences in amino acid composition of proteins from ProteinMPNN, HyperMPNN and MoMPNN.
Method
Designability Metrics
Evo. ppl ↓
AAR ↑
RMSD ↓
TM score ↑
PLDDT ↑
ESM-IF
4.033
0.805
80.55
6.538
0.509
InstructPLM
7.196
0.683
74.77
6.946
0.607
ProteinMPNN
3.658
0.823
82.09
6.843
0.441
Weighted-score DPO
3.783
0.811
80.74
5.843
0.410
MoMPNN [IG+ESM]
3.706
0.825
82.28
6.205
0.431
Appendix
Table 3: Comparison of protein sequence design methods across different evaluation metrics on de novo backbone structures from RFDiffusion. The best results and the second-best results are marked bold and bold .
Figure 5: Analysis of MoMPNN [IG+ESM] and Weighted-score DPO. Initial Guess, Evo. ppl and recover rate changes across each round of iterative refinement.
Design Target
Input PDB
Target chain and residue numbers
Target Hotspot
Binder length(for benchmarks)
PD-L1
5O45
A18-132
A56, A115, A123
50-120
SC2RBD
6M0J
E333-526
E485, E489, E494, E500, E505
50-120
BHRF1
2WH6
A2-158
A65, A74, A77, A82, A85, A93
80-120
PD-1
AF2 prediction
A32-146
A64, A126, A129, A133
80-150
CLN1-14
AF2 prediction
A1-188
A31, A46, A55, A152
80-175
Appendix
Table 4: Input structures, hotspot settings and binder lengths for benchmarks.