README.md

---
pretty_name: Dayhoff FASTA and MMseqs2 Databases
---

# Dayhoff FASTA and MMseqs2 databases

This dataset contains the original [Dayhoff Atlas](https://huggingface.co/datasets/microsoft/Dayhoff) GigaRef and UniRef50 datasets, in formats amenable to MMSeqs2 CPU and GPU utilities.

The train, validation, and test sets from the original atlas were combined and the following datasets available:

* GigaRef No Singletons - The GigaRef dataset, with no singleton clusters.
* GigaRef Singletons - The GigaRef dataset, with only singleton clusters.
* GigaRef Full - Every sequence contained in both the no-singletons and singletons subsets.
* UniRef50 - UniProt clustered at 50% sequence identity.

Each dataset is or will be available in the following formats:

* FASTA - Canonical sequence storage format, usable with many bioinformatics tools.
* MMSeqs2-CPU - Converted folder of unindexed database files compatible with MMSeqs2-CPU. Searches can be tuned to splits that accommodate your system RAM.
* MMSeqs2-GPU - Converted folder of padded sequence databases for MMSeqs2-GPU search. Requires 1+ GPU(s) on your machine to run.

## Current Repo Organization

```text
fastas/
├── gigaref-full.fasta.gz
├── gigaref-singletons.fasta.gz
├── gigaref-no-singletons.fasta.gz
└── uniref50.fasta.gz

mmseqs-cpu/
├── gigaref-singletons/db/
├── gigaref-no-singletons/db/
└── uniref50/db/

mmseqs-gpu/
├── gigaref-singletons/db_gpu/
├── gigaref-no-singletons/db_gpu/
└── uniref50/db_gpu/
```

MMseqs can read the `.fasta.gz` files directly.

| Artifact | Download | Working disk | Host RAM | GPU |
| --- | ---: | ---: | --- | --- |
| GigaRef full FASTA | 358.90 GB | 358.90 GB compressed | Not applicable | None |
| GigaRef singleton FASTA | 147.28 GB | 147.28 GB compressed | Not applicable | None |
| GigaRef no-singleton FASTA | 211.62 GB | 211.62 GB compressed | Not applicable | None |
| UniRef50 FASTA | 13.27 GB | 13.27 GB compressed | Not applicable | None |
| UniRef50 CPU MMseqs | 14.97 GB | approximately 28 GB extracted | 32 GB recommended; lower RAM works with splitting | None |
| UniRef50 GPU MMseqs | 15.29 GB | approximately 28 GB extracted | 32 GB recommended; lower RAM works with splitting | At least one MMseqs2-GPU-compatible NVIDIA GPU |
| GigaRef singleton CPU MMseqs | 182.49 GB | approximately 387 GB extracted | 64 GB starting point with splitting; 400+ GB maximizes throughput | None |
| GigaRef singleton GPU MMseqs | 189.12 GB | approximately 395 GB extracted | 64 GB starting point with splitting; 400+ GB maximizes throughput | At least one MMseqs2-GPU-compatible NVIDIA GPU; the database need not fit VRAM |
| GigaRef no-singleton CPU MMseqs | 279.08 GB | 572.73 GB extracted; allow 647 GB while extracting | 64 GB is a practical starting point with splitting; 600+ GB maximizes  throughput | None |
| GigaRef no-singleton GPU MMseqs | 292.13 GB | 586.94 GB extracted; allow 660 GB while extracting | 64 GB is a practical starting point with splitting; 600+ GB maximizes throughput | At least one MMseqs2-GPU-compatible NVIDIA GPU; the database need not fit VRAM |

Put the extracted MMseqs database and temporary search directory on the
fastest local SSD or NVMe available. I/O speed, RAM, and GPUs improve throughput;
slow storage substantially increases search time.

## Hugging Face download example

```bash
export REPO=microsoft/Dayhoff-MMseqs2
export REV=main
export DEST=/data/Dayhoff-MMseqs2
export TARGET=gigaref-no-singletons

# FASTA
hf download "$REPO" "fastas/$TARGET.fasta.gz" \
  --repo-type dataset --revision "$REV" --local-dir "$DEST"

# CPU MMseqs
hf download "$REPO" --repo-type dataset --revision "$REV" \
  --include "mmseqs-cpu/$TARGET/**" --local-dir "$DEST"

# GPU MMseqs
hf download "$REPO" --repo-type dataset --revision "$REV" \
  --include "mmseqs-gpu/$TARGET/**" --local-dir "$DEST"
```

Valid FASTA targets are `gigaref-full`, `gigaref-singletons`,
`gigaref-no-singletons`, and `uniref50`. CPU and GPU MMseqs targets are
`gigaref-singletons`, `gigaref-no-singletons`, and `uniref50`.

## Extract and search

The FASTA needs no extraction for MMseqs. To create an uncompressed FASTA:

```bash
pigz -dc "fastas/$TARGET.fasta.gz" > "fastas/$TARGET.fasta"
```

Extract either MMseqs representation once:

```bash
find "mmseqs-cpu/$TARGET" -type f -name '*.gz' -print0 |
  xargs -0 -n1 pigz -d

find "mmseqs-gpu/$TARGET" -type f -name '*.gz' -print0 |
  xargs -0 -n1 pigz -d
```

The resulting target prefixes are:

```text
mmseqs-cpu/$TARGET/db/db
mmseqs-gpu/$TARGET/db_gpu/db_gpu
```

For a 64 GB host, use native MMseqs target splitting:

```bash
mmseqs search queryDB TARGET_DB resultDB tmp \
  --gpu 1 \
  --split-mode 0 \
  --split-memory-limit 48G \
```

Omit `--gpu 1` for CPU search.

The UniRef50 and singleton databases were built directly from their published
combined FASTAs, so their sequence identifiers match. The no-singletons FASTA
and MMseqs database contain the same 1.8B-sequence corpus, but the FASTA uses
`gr_<source-index>` identifiers while the existing MMseqs database retains
older `g<part>_<row>` identifiers. This difference does not affect inference.