Use Dayhoff Atlas Online

Dayhoff Atlas: A Unified Protein Sequence Resource and Protein Language Models

The Dayhoff Atlas is an open resource that centralizes protein sequence data and protein language models (PLMs), combining metagenomic sequences, structure-derived synthetic sequences and homolog sets. The authors release the datasets, model weights and code under a permissive license.

PLMs depend on their training data, but existing sources (UniProt/UniRef, metagenomic catalogs, PDB and OpenProteinSet MSAs) are not unified. The authors also report that common in silico metrics may not predict whether designed proteins express in the lab.

How the Dayhoff Atlas and Models Work

• GigaRef: clusters 5.7B raw sequences in multiple steps; the non-singleton set has 1.89B sequences in 243M clusters.

• BackboneRef: 72,243,300 production sequences at 300 per backbone; 57.3% of backbones have max TM-score below 0.5 to natural structures.

• Architecture: one attention layer per seven state-space layers, with mixture-of-experts in every other layer.

• Training: autoregressive in both N-to-C and C-to-N directions, enabling any-order infilling, with MSAs unrolled into homolog sequences.

Dayhoff Results: Protein Generation, Scaffolding and Expression

The models were tested on zero-shot ProteinGym fitness prediction, motif scaffolding (MotifBench and the RFdiffusion benchmark), and homolog-conditioned generation, where Dayhoff outperformed EvoDiff-seq and EvoDiff-MSA on the stated measures. In Cas9 generation, 27 of 100 sequences had pLDDT above 0.7, and three shortened candidates (906 to 924 aa) had 51 to 71% identity to natural Cas9s.

Overall, 40 of 144 sequences (27.8%) expressed in the head-to-head test. Perplexity, pLDDT and BLAST homology did not predict expression.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with Dayhoff on Tamarind Bio

• Unconditional generation: sample new protein sequences.

• Family-conditioned design: condition on homologs to generate family members, as with the shortened Cas9s.

• Motif scaffolding: build sequences around a functional motif.

• Variant scoring: score sequence likelihoods and mutation effects, including indels.

How to Use Dayhoff on Tamarind Bio

1. Open the tool: log in to tamarind.bio and open Dayhoff.

2. Pick a mode: generation, conditioned generation or scoring, if exposed.

3. Provide inputs: sequences, homolog sets or motif residues as the paper defines.

4. Choose a model: the paper describes 170M and 3B variants, if exposed.

5. Run: submit the job.

6. Download: retrieve generated sequences or scores.

7. Filter and validate: predict structures, but note in silico metrics did not predict expression.

Things to Keep in Mind

• Family-conditioned generation, motif inpainting and thermostability need more validation.

• Singletons were excluded from training.

• Expression was tested only in E. coli.

• Gains from stacking all data sources in one model were not tested.

Source: bioRxiv preprint describing the Dayhoff Atlas, https://www.biorxiv.org/content/10.1101/2025.07.21.665991v1.