DNA FM 7B

DNA FM 7B is DNA foundation model trained on 10.6 billion nucleotides from 796 species, enabling genome mining, in silico mutagenesis studies, gene expression prediction, and directed sequence generation.

By scaling model depth while maintaining a short context length of 4000 nucleotides, DNA FM shows substantial improvements across a breadth of tasks in functional genomics using transfer learning, sequence generation, and unsupervised annotation of functional elements. Notably, DNA FM outperforms prior encoder-only architectures without new data, suggesting that new scaling laws are needed to achieve compute-optimal DNA language models.

Model Architectural Details

DNA FM 7B is based on the bidirectional transformer encoder (BERT) architecture with single-nucleotide tokenization, and is optimized using a masked language modeling (MLM) training objective.

To learn semantically meaningful representations, we employed an BERT-style encoder-only dense transformer architecture. We make minor updates to this architecture to align with current best practices, including using SwiGLU and LayerNorms. Additionally, we use Rotary Positional Embeddings (RoPE), given that DNA syntax does not function based on absolute nucleotide positions but nucleotides interact in highly local and context-specific ways. Below are more detailes about the model architecture:

Model Arch Component	Value
Num Attention Heads	32
Num Hidden Layers	32
Hidden Size	4352
Intermediate Size	11584
Vocab Size	16
Context Length	4000

Pre-training of DNA FM 7B

Here we briefly introduce the details of pre-training of DNA FM 7B. For more detailed information, please refer to our paper.

Data

To test whether representation capacity has limited the development of DNA language models in previous studies, we utilize the data set and splits from the Nucleotide Transformer. Starting from a total of 812 genomes with 712 for training, 50 for validation, and 50 for testing, we removed 17 entries which had been deleted from NCBI since the original dataset’s publication on Hugging Face. One of these was the important model organism Rattus norvegigus, which we replaced with the current reference genome. This resulted in 696 genomes for training, 50 for validation, and 50 for testing. We pre-trained DNA FM 7B With a total of 10.6 billion training tokens.

Training Details

The weights of our seven billion parameter model occupy over 200GB of memory in 32 bit precision. To train a model of this size, we use model parallelism to split training across 256 H100 GPUs using the Megatron-LM framework. We also employed bfloat16 mixed precision training and FlashAttention-2 to allow for training with large context length at scale. With this configuration, DNA FM 7B took 8 days to train.

Hyper-params	Value
Global Batch Size	1024
Per Device Micro Batch Size	2
Precision	Mixed FP32-BF16
Total Iters	100000

Tokenization

To minimize bias and learn high-resolution single-nucleotide dependencies, we opted to align closely with the real data and use character-level tokenization with a 5-letter vocabulary: A, T, C, G, N, where N is commonly used in gene sequencing to denote uncertain elements. Sequences were also prefixed with a [CLS] token and suffixed with a [EOS] token as hooks for downstream tasks. We chose a context length of 4,000 nucleotides as the longest context which would fit within DNA FM 7B during pretraining, and chunked our dataset of 796 genomes into non-overlapping segments.

Evaluation of DNA FM 7B

We evaluate the benefits of pretraining DNA FM 7B by conducting a comprehensive series of experiments related to functional genomics, genome mining, metabolic engineering, synthetic biology, and therapeutics design, covering supervised, unsupervised, and generative objectives. Unless otherwise stated, hyperparameters were determined by optimizing model performance on a 10% validation split of the training data, and models were tested using the checkpoint with the lowest validation loss. For more detailed information, please refer to our paper.

Results

TODO (@Caleb), we will need to see what results we want to put here.

How to Use

Build any downstream models from this backbone

Embedding

from genbio_finetune.tasks import Embed
model = Embed.from_config({"model.backbone": "dnafm"})
collated_batch = model.collate({"sequences": ["ACGT", "ACGT"]})
embedding = model(collated_batch)
print(embedding.shape)
print(embedding)

Sequence Level Classification

import torch
from genbio_finetune.tasks import SequenceClassification
model = SequenceClassification.from_config({"model.backbone": "dnafm", "model.n_classes": 2})
collated_batch = model.collate({"sequences": ["ACGT", "ACGT"]})
logits = model(collated_batch)
print(logits)
print(torch.argmax(logits, dim=-1))

Token Level Classification

import torch
from genbio_finetune.tasks import TokenClassification
model = TokenClassification.from_config({"model.backbone": "dnafm", "model.n_classes": 3})
collated_batch = model.collate({"sequences": ["ACGT", "ACGT"]})
logits = model(collated_batch)
print(logits)
print(torch.argmax(logits, dim=-1))

Regression

from genbio_finetune.tasks import SequenceRegression
model = SequenceRegression.from_config({"model.backbone": "dnafm"})
collated_batch = model.collate({"sequences": ["ACGT", "ACGT"]})
logits = model(collated_batch)
print(logits)

Or use our one-liner CLI to finetune or evaluate any of the above!

gbft fit --model SequenceClassification --model.backbone dnafm --data SequenceClassification --data.path <hf_or_local_path_to_your_dataset>
gbft test --model SequenceClassification --model.backbone dnafm --data SequenceClassification --data.path <hf_or_local_path_to_your_dataset>

For more information, visit: Model Generator

Citation

Please cite DNA FM using the following BibTeX code:

@inproceedings{ellington2024accurate,
title={Accurate and General {DNA} Representations Emerge from Genome Foundation Models at Scale},
author={Caleb Ellington, Ning Sun, Nicholas Ho, Tianhua Tao, Sazan Mahbub, Yonghao Zhuang, Hongyi Wang, Eric P. Xing, Le Song},
booktitle={NeurIPS 2024 Workshop on AI for New Drug Modalities},
year={2024}
}