Dayhoff is an Atlas of both protein sequence data and generative language models — a centralized resource that brings together 3.34 billion protein sequences across 1.7 billion clusters of metagenomic and natural protein sequences (GigaRef), 46 million structure-derived synthetic sequences (BackboneRef), and 16 million multiple sequence alignments (OpenProteinSet). These models can natively predict zero-shot mutation effects on fitness, scaffold structural motifs by conditioning on evolutionary or structural context, and perform guided generation of novel proteins within specified families. Learning from metagenomic and structure-based synthetic data from the Dayhoff Atlas increased the cellular expression rates of generated proteins, highlighting the real-world value of expanding the scale, diversity, and novelty of protein sequence data.
The Dayhoff architecture is a hybrid of state-space Mamba layers and Transformer self-attention, interleaved with Mixture-of-Experts modules to maximize capacity while preserving efficiency. It natively handles long contexts, allowing both single sequences and unrolled MSAs to be modeled. Trained with an autoregressive objective in both N→C and C→N directions, Dayhoff supports order-agnostic infilling and scales to billions of parameters.
Model Details
Model Description
Developed by: Kevin K. Yang, Sarah Alamdari, Alex J. Lee, Kaeli Kaymak-Loveless, Samir Char, Garyk Brixi, Carles Domingo-Enrich, Chentong Wang, Suyue Lyu, Nicolo Fusi, Neil Tenenholtz, Ava P. Amini
Model type: Hybrid state-space-model transformer architecture with mixture-of-experts
Dayhoff is intended for broad research use on protein language modeling. The model has been used and assessed on the following capabilities:
Unconditional design of protein sequences
Zero-shot mutation effect prediction on ProteinGym
Designing scaffolds for structural motifs in sequence space on RFDiffusion and MotifBench
Homolog conditioning with Dayhoff-3b-GR-HM and Dayhoff-3b-GR-HM-c
Bias, Risks, and Limitations
This model should not be used to generate anything that is not a protein sequence or a set of homologuous protein sequences. It is not meant for natural language or other biological sequences, such as DNA sequences. Not all sequences are guaranteed to be realistic. It remains difficult to generate high-quality sequences with no sequence homology to any natural sequence.
How to Get Started with the Model
The simplest way to use these models and datasets is via the HuggingFace interface. You will need PyTorch, mamba=ssm, causal-conv1d, and flash-attn.
Requirements:
PyTorch: 2.7.1
CUDA 12.8 and above
We recommend using uv and creating a clean environment.
For detailed instructions on package usage, please refer to the README in model repo.
Evaluation
Results
See the preprint for the latest benchmark results and evaluations.
Model perplexity on held-out test sequences for Dayhoff models.
Model
UniRef50
GigaRef
Aligned homologs
Unaligned homologs
170m-UR50
11.62
11.88
170m-UR90
11.52
11.85
170m-GR
13.67
9.36
170m-UR50-BRn
11.78
12.03
170m-UR50-BRq
11.67
11.91
170m-UR50-BRu
11.66
11.87
3b-UR90
8.95
9.64
3b-GR-HM
11.95
6.68
4.34
4.60
3b-GR-HM-c
10.11
9.21
3.57
3.56
Quality of generated sequences as measured by ESMFold pLDDT and scPerplexity. Dataset statistics are for 1024 randomly-sampled sequences. Model statistics are for 1024 generations at T=1 in the N-to-C direction.
Model or dataset
pLDDT (mean ± s.d.)
scPerplexity (mean ± s.d.)
Natural sequences
UniRef50
0.653 ± 0.196
9.45 ± 2.89
GigaRef-clusters
0.619 ± 0.199
9.69 ± 2.83
GigaRef-singletons
0.561 ± 0.201
10.07 ± 2.88
Generated sequences
170m-UR50
0.421 ± 0.132
11.97 ± 2.14
170m-UR90
0.407 ± 0.125
12.12 ± 2.14
170m-GR
0.422 ± 0.129
11.83 ± 2.12
170m-UR50-BRu
0.441 ± 0.157
11.71 ± 2.18
170m-UR50-BRq
0.434 ± 0.152
11.72 ± 2.24
170m-UR50-BRn
0.432 ± 0.131
11.77 ± 2.24
3b-UR90
0.454 ± 0.150
11.79 ± 2.38
3b-GR-HM
0.406 ± 0.126
11.50 ± 2.16
3b-GR-HM-c
0.423 ± 0.132
11.91 ± 2.18
ProteinGym zero-shot performance Spearman’s correlation coefficient on ProteinGym substitutions and indels.
Input
Model
Parameters
Substitutions
Indels
Single sequence
170m-UR50
170M
0.353
0.479
170m-UR90
170M
0.354
0.483
170m-GR
170M
0.199
0.292
170m-UR50-BRu
170M
0.341
0.476
170m-UR50-BRq
170M
0.356
0.477
170m-UR50-BRn
170M
0.341
0.478
3b-UR90
3B
0.394
0.497
3b-GR-HM
3B
0.328
0.423
3b-GR-HM-c
3B
0.417
0.466
Aligned homologs
3b-GR-HM-c
3B
0.368
NA
Unaligned homologs
3b-GR-HM-c
3B
0.372
0.401
RFDiffusion Benchmark Performance Motif scaffolding performance, problems solved, successes out of 100, and MotifBench score.
Problem
170m-UR50
170m-UR90
170m-GR
170m-UR50-BRn
170m-UR50-BRq
170m-UR50-BRu
3b-UR90
3b-GR-HM
3b-GR-HM-c
EvoDiff-Seq
1PRW
62
72
81
95
91
90
94
81
79
82
1BCF
0
0
5
0
0
0
10
8
0
7
5TPN
0
0
0
0
0
0
0
0
0
0
5IUS
0
0
0
0
0
0
0
0
0
0
3IXT
12
17
12
14
18
12
18
11
14
20
5YUI
0
0
0
0
0
0
0
0
0
0
1QJG
0
0
0
0
0
0
0
0
0
0
1YCR
2
5
0
6
7
6
2
3
4
2
2KL8
0
1
0
1
0
1
1
1
1
1
7MRX_60
1
0
0
0
0
2
42
0
9
0
7MRX_85
0
0
0
0
0
0
19
1
1
0
7MRX_128
0
0
0
0
0
0
0
0
0
0
4JHW
0
0
0
0
0
0
0
0
0
0
4ZYP
0
0
0
0
0
1
0
0
0
0
5WN9
0
0
0
0
0
0
0
0
0
0
6VW1
1
1
1
0
0
1
0
0
0
0
5TRV_short
0
0
0
0
0
0
0
0
0
0
5TRV_med
0
0
0
0
0
0
0
0
0
0
5TRV_long
0
0
0
0
0
0
0
0
0
0
6E6R_short
2
2
1
3
3
2
14
7
8
6
6E6R_med
0
1
2
0
0
2
4
0
2
0
6E6R_long
0
1
0
0
0
1
3
0
1
0
6EXZ_short
0
0
0
0
0
0
0
0
0
0
6EXZ_med
0
0
0
0
0
0
0
0
0
0
6EXZ_long
0
0
0
0
0
0
0
0
0
0
Problems solved
6
8
6
5
4
10
10
7
9
6
Successes
80
100
102
119
119
118
207
112
119
118
Score
9.65
12.25
6.10
7.26
10.62
14.36
16.32
11.90
14.14
7.67
MotifBench Benchmark Performance Motif scaffolding performance, problems solved, successes out of 100, and MotifBench score.
Problem
170m-UR50
170m-UR90
170m-GR
170m-UR50-BRn
170m-UR50-BRq
170m-UR50-BRu
3b-UR90
3b-GR-HM
3b-GR-HM-c
EvoDiff-Seq
01_1LDB
1
1
3
0
0
1
20
2
12
0
02_1ITU
4
33
4
1
1
4
37
57
48
0
03_2CGA
0
0
0
0
0
0
0
0
0
0
04_5WN9
0
0
0
0
0
0
0
0
0
0
05_5ZE9
0
1
21
0
0
0
16
40
9
0
06_6E6R
1
1
1
1
2
1
6
3
1
2
07_6E6R
0
0
0
2
0
0
2
0
0
0
08_7AD5
0
0
0
0
0
0
0
0
0
0
09_7CG5
0
0
0
0
0
0
0
0
0
0
10_7WRK
0
0
0
0
0
0
0
0
0
0
11_3TQB
4
11
3
4
3
7
40
8
26
0
12_4JHW
0
0
0
0
0
0
0
0
0
0
13_4JHW
0
0
0
0
0
0
0
0
0
0
14_5IUS
0
0
0
0
0
0
0
0
0
0
15_7A8S
0
0
0
0
0
0
0
0
0
0
16_7BNY
0
0
0
0
0
0
0
0
0
0
17_7DGW
0
0
0
0
0
0
0
0
0
0
18_7MQQ
0
0
0
0
0
0
0
0
0
0
19_7MQQ
0
0
0
0
0
0
0
0
0
0
20_7UWL
0
0
0
0
0
0
0
0
0
0
21_1B73
0
0
0
0
0
0
0
0
0
0
22_1BCF
0
0
3
0
0
0
20
9
0
19
23_1MPY
0
0
0
0
0
0
0
0
0
0
24_1QY3
0
0
0
0
0
0
0
0
0
0
35_2RKX
0
0
0
0
0
0
0
0
0
0
36_3B5V
0
0
0
0
0
0
0
0
0
0
37_4XOJ
0
0
0
0
0
0
0
0
0
0
28_5YUI
0
0
0
0
0
0
0
0
0
0
29_6CPA
0
0
0
0
0
0
0
0
0
0
30_7UWL
0
0
0
0
0
0
0
0
0
0
Problems
4
5
6
4
3
4
7
6
5
2
Successes
10
47
35
8
6
13
141
119
96
21
Score
2.33
2.92
4.33
2.75
2.17
2.75
8.36
4.96
4.48
1.58
Technical Specifications
Compute Infrastructure
170M-parameter models: trained on 8 NVIDIA A100 or 8 NVIDIA H100 GPUs using Distributed Data Parallel.
3B-parameter models: trained on 176 NVIDIA H100 GPUs using Fully Sharded Data Parallel in hybrid-shard mode.
Responsible AI Considerations
The intended use of this model is to generate high-quality, realistic, protein sequences or sets of homologous protein sequences. Generations can be designed from scratch or conditioned on partial sequences in both N→C and C→N directions.
The code and datasets released in this repository are provided for research and development use only. They are not intended for use in clinical decision-making or for any other clinical use, and the performance of these models for clinical use has not been established. You bear sole responsibility for any use of these models, data and software, including incorporation into any product intended for clinical use.
Citation
If you use the code, data, models, or results. please cite our preprint.
