Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute
Authors: Kieran Didi, Zuobai Zhang, Guoqing Zhou, Danny Reidenbach, Zhonglin Cao, Sooyoung Cha, Tomas Geffner, Christian Dallago, +6 more
Organizations: NVIDIA · University of Oxford · Mila - Québec AI Institute · Université de Montréal · School of Biological Sciences, Seoul National University · Interdisciplinary Program in Bioinformatics, Seoul National University · HEC Montréal · CIFAR AI Chair · AITHYRA · Institute of Molecular Biology and Genetics, Seoul National University · Artificial Intelligence Institute, Seoul National University
Protein interaction modeling is central to protein design, which has been transformed by machine learning with applications in drug discovery and beyond. In this landscape, structure-based de novo binder design is cast as either conditional generative modeling or sequence optimization via structure predictors ("hallucination"). We argue that this is a false dichotomy and propose Proteina-Complexa, a novel fully atomistic binder generation method unifying both paradigms. We extend recent flow-based latent protein generation architectures and leverage the domain-domain interactions of monomeric computationally predicted protein structures to construct Teddymer, a new large-scale dataset of synthetic binder-target pairs for pretraining. Combined with high-quality experimental multimers, this enables training a strong base model. We then perform inference-time optimization with this generative prior, unifying the strengths of previously distinct generative and hallucination methods. Proteina-Complexa sets a new state of the art in computational binder design benchmarks: it delivers markedly higher in-silico success rates than existing generative approaches, and our novel test-time optimization strategies greatly outperform previous hallucination methods under normalized compute budgets. We also demonstrate interface hydrogen bond optimization, fold class-guided binder generation, and extensions to small molecule targets and enzyme design tasks, again surpassing prior methods. Code, models and new data will be publicly released.
Figures & tables
Figure 1: (Top) Proteína-Complexa’s target-conditioned generation process. (Bottom) Scaling test-time compute, we use Complexa’s generative prior for more efficient optimization than previous hallucination methods ( Sec. 3.4 ). We depict beam search, which steers stochastic generation toward high-quality binders, guided by structure prediction models’ interface scores or hydrogen bond energies. Intermediate candidate states (blue) are scored via rollouts (blue, dotted), promising candidates are kept, and new trajectories are launched (orange).
Figure 2: Binders generated by Complexa, passing in-silico success criteria (more visualizations in Appendix K ). (a) TNF- α three-chain target. (b) Claudin-1 target, red interface hydrogen bonds. (c) OQO small molecule target.
Figure 3: Teddymer dimers resemble realistic binder-target structures, including interface hydrogen bonding (zoom-in). Also see Appendix C .
Figure 4: Filtered training datasets used by Complexa.
Figure 5: Complexa’s latent target conditioning . When training the conditional denoiser, the encoder and decoder are frozen.
Table 1: Complexa’s generative performance for small molecule targets without optimization. RFDiffusion-AllAtom uses LigandMPNN, we evaluate sequences produced by Complexa.
Model
# Unique Successes ↑
# Times Best Method ↑
Time [s] ↓
Novelty ↓
Self
MPNN-FI
MPNN
Self
MPNN-FI
MPNN
RFDiffusion ( Watson et al., 2023 )
–
–
4.68
–
–
3
70.8
0.87
Protpardelle-1c ( Lu et al., 2025 )
–
–
0.73
–
–
0
8.13
0.77
APM ( Chen et al., 2025 )
0.31
1.52
3.15
1
0
1
73.1
0.86
Complexa (ours)
9.10
13.6
14.4
14
14
14
15.6
0.80
Table 2: Complexa’s generative performance for protein targets without optimization vs. baselines. Self denotes model sequence evaluation, MPNN full backbone-based re-design, and MPNN-FI the same with fixed interface amino acids. Note that RFDiffusion and Protpardelle only generate backbones and not their own sequences. Complete results in Appendix I . If methods tie on unique and absolute successes, we do not count this, see Appendix F .
Figure 7: Scaling analysis of inference-time optimization across different target difficulties. Average unique success rate on 12 easy and 7 hard targets against test-time GPU hours. Five inference-time optimization algorithms from Sec. 3.4 are applied to Complexa and compared with hallucination baselines BoltzDesign and BindCraft.
Figure 8: Inference-time scaling for VEGFA multi-chain target (hard).
Figure 9: Inference-time optimization for small molecule targets.
Model
# Unique
# H-Bonds
Successes ↑
(avg.) ↑
Complexa (no reward)
77.00
5.271
Complexa w/ fipAE
83.36
5.524
Complexa w/ fH-Bond
82.36
7.154
Complexa w/ fipAE + fH-Bond
86.26
6.518
Table 3: Inference-time optimization using beam search with different combinations of folding and hydrogen bond rewards. Details in Sec. I.6 .
Figure 10: Enzyme Design Benchmark. Complexa significantly outperforms RFDiffusion2 in 38/41 AME benchmark tasks, both with re-designed and self-generated sequences. Extended results in Sec. I.9 .
Figure 11: (a) Interface hydrogen bond optimization can create binders with extended interacting regions, forming strong hydrogen bonding (red, zoom-in). (b) Regular binders, even though passing success criteria, can be smaller with less interactions. Visualization shows successful binders for TrkA target.
Appendix figures & tables33 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 12: Experimental vs. synthetic interfaces. Comparison between Teddymer and PDB multimer on various protein-protein interface metrics.
Target Name
PDB ID
Target Chain & Residues
Hotspot Residues
Binder Length
Source
BBF-14
9HAG
A1-112
–
70-250
BindCraft
Betv1
AF2 prediction
A1-159
A24
70-185
BindCraft
CbAgo (PAZ)
AF2 prediction
A180-276
A268
70-160
BindCraft
CD45
AF2 prediction
d1 domain
A1
80-200
BindCraft
Claudin1
AF2 prediction
A1-188
A31, A46, A55, A152
80-175
BindCraft
CrSAS6
AF2 prediction
A1-145
–
90-200
BindCraft
Appendix
Table 4: Selected protein targets with their structural information, binding specifications, and data source.
Target Name
PDB ID
Target Chain
Binder Length
SAM
A
A
100
FAD
B
A
100
IAI
C
A
100
OQO
D
A
100
Appendix
Table 5: Selected small molecule targets with their structural information, binding specifications, and data source.
p(c)=∑c′∈Cexp{βR(c′)}exp{βR(c)}
Appendix
Algorithm 3 Protein Binder Feynman–Kac Steering with Reward Approximation
Model
# Unique Successes ↑
# Times Best Method ↑
Self
MPNN-FI
MPNN
Self
MPNN-FI
MPNN
Complexa
9.10
13.5
14.4
19
13
14
Complexa w/o Teddymer
0.15
1.68
3.84
0
0
0
Complexa w/o Translation Noise
1.47
3.89
3.73
0
6
5
Appendix
Table 6: Translation noise and Teddymer ablation studies. average unique successes and the number of times each method ranks best across 19 targets. Results are shown for Complexa, and variants without Teddymer data and without translation noise.
Target
Complexa
Complexa w/o Teddymer
Complexa w/o Translation Noise
MPNN
MPNN-FI
Self
MPNN
MPNN-FI
Self
MPNN
MPNN-FI
Self
PD-L1
16
14
9
8
4
0
1
4
1
BHRF1
26
29
21
18
7
2
28
28
14
IFNAR2
51
52
39
5
2
0
4
2
0
BBF14
24
25
20
12
10
0
4
4
0
DerF7
24
16
10
7
0
0
1
2
2
Appendix
Table 7: Translation noise and Teddymer ablation studies (detailed results). Unique successes across 19 targets for Complexa and its variants without Teddymer data or translation noise, evaluated with different sequence redesign methods. Results are visualized in Fig. 13 .
Figure 13: Translation noise and Teddymer ablation studies. Unique successes of different Complexa variants (base, without Teddymer data, and without translation noise) on easy targets (left of the dashed line) and hard targets (right of the dashed line).
Figure 14: Teddymer ablation study with Boltz-2. Boltz-2-based evaluation of the unique successes of different Complexa variants (trained with vs. without Teddymer data) on easy targets (left of the dashed line) and hard targets (right of the dashed line).
Model
# Unique Successes ↑
MPNN
MPNN-FI
Self
Complexa
34.4
31.4
14.8
Complexa w/o Teddymer
16.2
13.1
2.2
Appendix
Table 8: Teddymer ablation study with Boltz-2. Boltz-2-based evaluation of the average unique successes across 19 targets. Results are shown for Complexa and the variant trained without Teddymer data.
Figure 15: Teddymer ablation study with RF3. RF3-based evaluation of the unique successes of different Complexa variants (trained with vs. without Teddymer data) on easy targets (left of the dashed line) and hard targets (right of the dashed line).
Model
# Unique Successes ↑
MPNN
MPNN-FI
Self
Complexa
4.9
4.3
2.4
Complexa w/o Teddymer
1.2
0.5
0.1
Appendix
Table 9: Teddymer ablation study with RF3. RF3-based evaluation of the average unique successes across 19 targets. Results are shown for Complexa and the variant trained without Teddymer data.
Figure 16: Generate & Hallucinate ablation study. Numbers of unique successes against inference-time GPU hours, for Generate & Hallucinate, its variant with logit optimization, and BindCraft, averaged over 12 easy and 7 hard targets.
Figure 17: Generate & Hallucinate ablation study (easy targets). Number of unique successes over 12 easy targets against inference-time GPU hours, for Generate & Hallucinate, its variant with logit optimization, and BindCraft. Successes are measured with self-generated sequences and inference-time compute includes both generation and evaluation time.
Figure 18: Generate & Hallucinate ablation study (hard targets). Number of unique successes over 7 hard targets against inference-time GPU hours, for Generate & Hallucinate, its variant with logit optimization, and BindCraft. Successes are measured with self-generated sequences and inference-time compute includes both generation and evaluation time.
Figure 19: H1, TNF- α , and IL17A targets. Inference-time optimization compute scaling plots for three very hard targets. See Sec. I.4 for details.
Target
RFDiffusion
Protpardelle-1c
APM
Complexa (Ours)
MPNN
MPNN
MPNN
MPNN-FI
Self
MPNN
MPNN-FI
Self
PD-L1
8
2
1
0
0
16
14
9
BHRF1
5
4
20
10
5
26
29
21
IFNAR2
11
2
3
3
1
51
52
39
BBF14
2
0
0
0
0
24
25
20
DerF7
5
0
0
0
0
24
16
10
Appendix
Table 10: De novo binder design performance for all protein targets . Unique successes for different models and three sequence redesign methods per target.
Target
Model
# Unique Successes ↑
Self
MPNN-FI
MPNN
SAM
RFdiffusionAA
-
-
2
Complexa
10
8
15
OQO
RFdiffusionAA
-
-
3
Complexa
6
12
9
FAD
RFdiffusionAA
-
-
5
Appendix
Table 11: De novo binder design performance of models for all small molecule targets . RFdiffusionAA only produces backbones and as a result can only be evaluated under full LigandMPNN re-design. We report the number of FoldSeek clusters out of the successful subset of 200 samples of length 100.
Figure 20: Unique successes for different models with three sequence redesign methods on easy targets (left of the dashed line) and hard targets (right of the dashed line). Detailed results are in Tab. 10
Figure 21: Unique successes against inference-time GPU hours on 12 easy targets. Successes are measured with self-generated sequences and inference-time compute includes both generation and evaluation time.
Figure 22: Unique successes against inference-time GPU hours on 7 hard targets. Successes are measured with self-generated sequences and inference-time compute includes both generation and evaluation time.
Figure 23: Unique successes against inference-time GPU hours on 4 small molecule targets. Successes are measured with self-generated sequences and inference-time compute includes both generation and evaluation time.
Figure 24: Unique successes and number of interface hydrogen bonds on 19 targets using Beam Search with different combinations of folding and hydrogen bond rewards. Successes are measured with self-generated sequences and hydrogen bond counts are computed on refolded structures over successful designs.
AME Task
# residue islands
RFDiffusion2 LigandMPNN(8)
RFDiffusion2 LigandMPNN(1)
Complexa (ours) LigandMPNN(8)
Complexa (ours) LigandMPNN(1)
Complexa (ours) Self-Generated
All
Unique
All
Unique
All
Unique
All
Unique
All
Unique
M0630
1
78
37
55
22
88
38
76
32
70
27
M0663
1
1
1
0
0
62
38
36
26
31
24
M0710
2
72
42
39
22
91
40
67
34
64
31
M0717
2
55
28
27
12
65
26
39
16
28
11
M0664
2
72
38
36
19
88
49
70
47
60
42
Appendix
Table 12: Enzyme Design Benchmark—Detailed Quantitative Results. “All” indicates total number of successes produced by the model (we produce 100 samples per task), while “Unique” indicates number of unique successes, obtained by clustering all successes. The method with the most unique successes for 8 sequence re-designs (LigandMPNN(8)) is highlighted in blue and for single sequence methods is highlighted in red.
Figure 25: Enzyme Design Benchmark—Best-of-8 Re-designs. Complexa significantly outperforms RFDiffusion2 in 38/41 AME enzyme design benchmark tasks. Both methods use best-of-8 re-designed sequences.
Figure 26: Enzyme Design Benchmark—Aggregated Results. Complexa significantly outperforms RFDiffusion2 across all sequence modes.
Figure 27: Extensive Search Case Study. Number of unique successes on the target SpCas9 against inference-time GPU hours with different TM-score clustering thresholds used in the unique success calculation (green: 0.4 , red: 0.45 and blue: 0.5 ). Successes are measured with self-generated sequences and inference-time compute includes both generation and evaluation time.
Figure 28: Interface hydrogen bonding and interface predicted aligned error correlations. The number of interface hydrogen bonds increases with decreasing ipAE values of ColabDesign during inference-time search with negative ipAE as a reward ( Sec. I.11 ).
Target
Model
pK d ↑
# of Samples
Mean ± Std
Best
SAM
RFDiffusionAA ‡†
5.65 ± 0.15
5.76
2
Complexa (ours)
4.81 ± 0.49
5.49
10
Complexa (ours) †
4.94 ± 0.54
5.74
10
OQO
RFDiffusionAA ‡†
7.01 ± 0.31
7.35
3
Complexa (ours)
7.19 ± 0.14
7.37
6
Appendix
Table 13: Binding affinity comparison of unique small molecule binders. We report the average and best pK d in kcal/mol predicted by FLOWR-ROOT. ‡ denotes redesign of the binder sequence using LigandMPNN. † denotes co-folding with RosettaFold3, otherwise the model-generated structure is used. RFDiffusionAA requires both LigandMPNN redesign and co-folding as it generates a backbone without sequence.
Figure 29: Generated binder interfaces vs. PDB multimer inferfaces. Comparison of bioinformatic metrics between the PDB Multimer reference distribution and successful binders generated by Complexa.
Figure 30: Binders generated by Complexa for the challenging large multi-chain targets TNF- α (three-chain target), IL17A (two-chain target), and H1 (two-chain target; two different binders visualized). All shown binders meet the in-silico success criteria. Binders are visualized in cartoon representation, including side chains, multi-chain targets in surface representation. See Sec. I.4 .
Figure 31: Binders generated by Complexa for small molecule targets SAM, IAI, FAD and OQO (see Appendix E ). All shown binders meet the in-silico success criteria.
Figure 32: Fold class-conditioned binders generated by Complexa for five different targets. We control the secondary structure content through conditioning on C-level CAT labels “Mainly Alpha”, “Mainly Beta” or “Mixed Alpha Beta” ( Dawson et al., 2016 ; Geffner et al., 2025 ) . We can observe that the generated binders follow the conditioning, exhibiting primarily alpha helices, beta sheets, or both. All shown binders meet the in-silico success criteria. Binders are visualized in cartoon representation, including side chains, targets in surface representation.
Hyperparameters
Complexa
Protein-Protein
Protein-Protein CAT
Ligand-Protein
Model Architecture
Partially Latent Flow Matching
sequence repr dim
768
768
768
sequence cond dim
256
256
256
t sinusoidal enc dim
256
256
256
Appendix
Table 14: Hyperparameters for Complexa model training. We denote two versions of Complexa that specialized in generating binders for protein targets and small molecule ligand target as Protein-Protein and Ligand-Protein , respectively. Protein-Protein CAT refers to the model that can generate binders to protein target conditioned on CAT labels.
A central challenge in de novo protein design is generating proteins with desired structural and functional properties. Existing design models typically require a dedicated input representation for each property, limiting the design objectives they support. We develop Prompteina, the first all-atom structure generation model that supports new and broader design objectives specified in natural language, analogous to text-to-image generation. To train Prompteina, we construct Prompteina-Data, a paired text--protein structure dataset of 1.8M structures and 50M+ captions describing global and residue-specific structural and functional properties. We adapt Proteina-Complexa, a pretrained structure generation model, through text-conditioned pretraining and finetuning that aligns residues with their corresponding text. Prompteina outperforms baselines on 15 of 18 structural properties and generates proteins that more closely match descriptions of their intended functions. In binder design, it leads baselines on at least 13 of 20 targets and improves interface confidence by 14% over the strongest baseline. With iterative in silico feedback, Prompteina's gain over non-iterative generation is 57% larger than the next-best baseline on the hardest targets. Finally, we experimentally validate Prompteina designs against GM2A, a target with no reported de novo binders. We identify nine binders, with the strongest at an apparent KD of 107 nM. These results establish natural language as an effective and powerful conditioning interface for protein structure generation.
Joseph Lee, Yanjing Li, Bo Qiang +7
University of Washington · National University of Singapore · Allen Institute for AI (Ai2) +1
The rapid evolution of generative models has unlocked new potentials in protein binder design, a pivotal task in structural biology, by facilitating end-to-end generation via joint sequence-structure modeling or hallucination. However, existing approaches are predominantly implemented under a single-target, single-state assumption, limiting their ability to model multi-target or multi-state interactions required for advanced function-oriented protein design. Here, we introduce Chamaileon, which unifies multi-target and multi-state binder design by formulating the problem as cross-context binding landscape modeling. The framework is underpinned by a training paradigm termed In-Context Complex Co-Design (I3CD) for context-aware sequence-structure co-modeling. During inference, we employ Mixture-of-Paths Sampling (MoPS), a scalable strategy that optimizes a single sequence across contexts while alleviating the scarcity of high-quality multi-conformational paired data. Extensive evaluation on our newly constructed benchmark, CROSS, demonstrates that Chamaileon effectively generates sequences adaptable to diverse conformational landscapes and multi-target requirements. The code is available on https://github.com/caohengyuan/Chamaileon.
Hengyuan Cao, Shizhuo Cheng, Mingxuan Liu +5
Zhejiang University, Hangzhou, China · Fudan University, Shanghai, China · Shanghai Institute for Advanced Study Zhejiang University, Shanghai, China +1
The central challenge in de novo protein design is generating plausible, mutually compatible structures and sequences, such that each designed sequence folds into its intended structure and the structure accommodates that sequence. Compared to typical two-stage design methods, which decouple the modeling of the interdependent modalities, co-design models improve the cross-modal consistency by jointly generating sequences and structures. However, naively generating sequences and structures simultaneously does not ensure their consistency. To address this challenge, we propose CODesign framework. We improve data consistency by generating approximately 105,000 consistency-distilled dimers. We further promote consistency through a multimodal joint flow model that captures the joint distribution of sequences, backbone structures, and local atomic configurations, together with a consistency-aware joint resampling strategy that iteratively refines sequences and side chains. Experiments show that CODesign achieves state-of-the-art performance with the highest in silico success rates on both protein- and ligand-target binder design. Ablation studies also demonstrate our distilled dataset increases performance by 70.9%, which can be further improved by our proposed resampling mechanism with negligible additional computational cost. Code, model weights and the new dataset will be completely open-source.
Yuanle Mo, Bo Qiang, Haitao Lin +4
Valhalla Technology · The Chinese University of Hong Kong · University of Washington +2