macwiatrak
/

bacformer-causal-MAG

@@ -1,199 +1,183 @@
 ---
 library_name: transformers
-tags: []
 ---
-# Model Card for Model ID
-<!-- Provide a quick summary of what the model is/does. -->
-## Model Details
-### Model Description
-<!-- Provide a longer summary of what this model is. -->
-This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
-- **Developed by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Model type:** [More Information Needed]
-- **Language(s) (NLP):** [More Information Needed]
-- **License:** [More Information Needed]
-- **Finetuned from model [optional]:** [More Information Needed]
 ### Model Sources [optional]
-<!-- Provide the basic links for the model. -->
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
-## Uses
-<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
-### Direct Use
-<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
-[More Information Needed]
-### Downstream Use [optional]
-<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
-[More Information Needed]
-### Out-of-Scope Use
-<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
-[More Information Needed]
-## Bias, Risks, and Limitations
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
-[More Information Needed]
-### Recommendations
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
-Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
-## How to Get Started with the Model
-Use the code below to get started with the model.
-[More Information Needed]
 ## Training Details
 ### Training Data
-<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
-[More Information Needed]
 ### Training Procedure
-<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
-#### Preprocessing [optional]
-[More Information Needed]
-#### Training Hyperparameters
-- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
-#### Speeds, Sizes, Times [optional]
-<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
-[More Information Needed]
-## Evaluation
-<!-- This section describes the evaluation protocols and provides the results. -->
-### Testing Data, Factors & Metrics
-#### Testing Data
-<!-- This should link to a Dataset Card if possible. -->
-[More Information Needed]
-#### Factors
-<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
-[More Information Needed]
-#### Metrics
-<!-- These are the evaluation metrics being used, ideally with a description of why. -->
-[More Information Needed]
-### Results
-[More Information Needed]
-#### Summary
-## Model Examination [optional]
-<!-- Relevant interpretability work for the model goes here -->
-[More Information Needed]
-## Environmental Impact
-<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
-Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
-- **Hardware Type:** [More Information Needed]
-- **Hours used:** [More Information Needed]
-- **Cloud Provider:** [More Information Needed]
-- **Compute Region:** [More Information Needed]
-- **Carbon Emitted:** [More Information Needed]
-## Technical Specifications [optional]
-### Model Architecture and Objective
-[More Information Needed]
-### Compute Infrastructure
-[More Information Needed]
-#### Hardware
-[More Information Needed]
-#### Software
-[More Information Needed]
-## Citation [optional]
-<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
 **BibTeX:**
-[More Information Needed]
-**APA:**
-[More Information Needed]
-## Glossary [optional]
-<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
-[More Information Needed]
-## More Information [optional]
-[More Information Needed]
-## Model Card Authors [optional]
-[More Information Needed]
-## Model Card Contact
-[More Information Needed]

 ---
 library_name: transformers
+tags:
+- biology
+- bacteria
+- prokaryotes
+- genomics
+- protein
+- plm
+- cplm
+license: apache-2.0
+metrics:
+- accuracy
 ---
+# bacformer-causal-MAG
+Bacformer is a foundational model for bacterial genomics, modeling the whole bacterial genome as a sequence of proteins.
+Bacformer takes as input the set of proteins present in a genome, **ordered** by their positions on the chromosome and plasmid(s), and
+computes contextual protein representations with a transformer that conditions each protein on every other protein in the genome.
+Each protein is treated as a token; thus, Bacformer may be viewed as a **contextualized protein language model** that captures the protein–protein interactions
+underlying an organism’s phenotype.
+Bacformer is pretrained to predict a masked (or next) protein *family* given the remaining proteins in the genome.
+The base model was trained on ~1.3 M diverse bacterial genomes comprising ~3 B protein sequences and can be adapted to a wide range of downstream tasks.
+See the [Bacformer models](https://huggingface.co/collections/macwiatrak/bacformer-681a17d6a77a928a1531def2) collection for all available checkpoints.
+A key member of this collection is **bacformer-causal-MAG**, a 27M-parameter base transformer model trained on ~1.3M
+diverse metagenome-assembled genomes across more than 20k bacterial species and multiple environments.
+All Bacformer variants embed protein sequences with a base protein language model by averaging the amino-acid token embeddings across each sequence.
+Bacformer uses [ESM-2 t12 35 M](https://huggingface.co/facebook/esm2_t12_35M_UR50D) as its base model.
+- **Developed by:** University of Cambridge (Floto Lab) & EPFL (Brbić Lab), led by Maciej Wiatrak
+- **License:** Apache 2.0
 ### Model Sources [optional]
+- **Repository:** [https://github.com/macwiatrak/Bacformer](https://github.com/macwiatrak/Bacformer)
+- **Paper:** TBA
+## Usage
+Install the `bacformer` package (see [https://github.com/macwiatrak/Bacformer](https://github.com/macwiatrak/Bacformer)). An end-to-end Python example demonstrating how to embed a genome with Bacformer is provided in the *tutorials* folder.
+Below snippet shows how you can embed multiple protein sequences with `Bacformer`.
+```python
+import torch
+from transformers import AutoModel
+from bacformer.pp import protein_seqs_to_bacformer_inputs
+device = "cuda:0"
+model = AutoModel.from_pretrained("macwiatrak/bacformer-causal-MAG", trust_remote_code=True).to(device).eval().to(torch.bfloat16)
+# Example input: a sequence of protein sequences
+# in this case: 4 toy protein sequences
+# Bacformer was trained with a maximum nr of proteins of 6000.
+protein_sequences = [
+    "MGYDLVAGFQKNVRTI",
+    "MKAILVVLLG",
+    "MQLIESRFYKDPWGNVHATC",
+    "MSTNPKPQRFAWL",
+]
+# embed the proteins with ESM-2 to get average protein embeddings
+inputs = protein_seqs_to_bacformer_inputs(
+    protein_sequences,
+    device=device,
+    batch_size=128,  # the batch size for computing the protein embeddings
+    max_n_proteins=6000,  # the maximum number of proteins Bacformer was trained with
+)
+# move the inputs to the device
+inputs = {k: v.to(device) for k, v in inputs.items()}
+# compute contextualized protein embeddings with Bacformer
+with torch.no_grad():
+    outputs = model(**inputs, return_dict=True)
+print('last hidden state shape:', outputs["last_hidden_state"].shape)  # (batch_size, max_length, hidden_size)
+print('genome embedding:', outputs.last_hidden_state.mean(dim=1).shape)  # (batch_size, hidden_size)
+```
+### Tutorials
+We include a number of tutorials, see the [tutorials](https://github.com/macwiatrak/Bacformer) folder on the github repository.
 ## Training Details
 ### Training Data
+**bacformer-causal-MAG** was pretrained on the **MAG corpus**
+(≈1.3 M metagenome-assembled genomes).
+The MAG set maximises environmental and taxonomic diversity and contains roughly 3B protein sequences.
 ### Training Procedure
+#### Preprocessing
+Each genome is represented as an ordered list of proteins.
+ranslated protein sequences were obtained from annotations or translated de novo when necessary.
+Proteins were ordered by genomic coordinates; for MAGs, proteins within each contig were ordered, and the contig order was randomised in every epoch.
+The set of protein sequences in the genome are embedded with the base protein language model ([ESM-2 t12 35M](https://huggingface.co/facebook/esm2_t12_35M_UR50D)).
+The protein embeddings computed by averaging the amino acid tokens in a protein sequence are used as protein tokens.
+Finally the protein tokens are fed into a transformer model.
+During pretraining, we limit the maximum number of proteins in a genome to `6,000`, which covers whole bacterial genome in `>98%` of genomes present in out training corpus.
+### Pretraining
+#### Training objective
+The model was optimised to predict the next protein given previous proteins present in the genome. As the number of protein is effectively unbound, we assigned each protein a discrete protein family index by performing
+unsupervised clustering on a set of proteins, resulting in `50k` distinct protein family clusters.
+Importantly, the input to Bacformer are exact protein sequence and we only use the discrete protein family label in a final classification layer where predicting
+the protein family of masked proteins. This allows the model to work on amino acid level tasks where even single mutations can change the phenotype of a genome.
+The causal pretraining objective effectively predicts the protein family ID of the next protein given the embedding of the previous protein in the genome sequence.
+#### Pretraining details
+The initial pretraining on MAGs was trained on 4 A100 80GB NVIDIA GPUs, with an effective batch size of 32. The maximum sequence length of each protein
+was set to `1,024` and the maximum number of proteins in a genome was set to `6,000`, which covers whole bacterial genome in `>98%` of genomes present in our training corpus.
+The Adam optimizer [1] was used with a linear warmup learning rate schedule, with the number of warmup steps equal to `7,500`. The base learning rate of `0.00015` was used,
+scaled by square root of number of GPUs (`lr = args.lr * np.sqrt(max(n_gpus, 1))`). We monitor the loss on the validation set as the measure of performance during training.
+### Architecture
+#### Input embeddings
+The input embeddings are created by adding together 1) protein embeddings from a base protein language model (pLM) [ESM-2 t12 35M](https://huggingface.co/facebook/esm2_t12_35M_UR50D),
+2) contig (token type) embeddings. The 1) are created by embedding a protein sequence with the pretrained base pLM model and taking the average of all amino acid
+tokens in the sequence, resulting in a `D` dimensional vector. We embed all of the `N` proteins present in the whole bacterial genome resulting in a `N x D` matrix,
+where `D` is the dimension of the base pLM model, here `480`. The protein embeddings are added together with the 2) contig embeddings. The contig embeddings
+are learnable embeddings which represent the unique contigs present in the genome. As an example, if a genome is made up of `K` contigs, each containing a number of proteins,
+each protein within the same contig will have the same contig embedding, which is different from the embeddings of different contigs.
+Contig embeddings have been created to account for the fact that bacterial genomes are often made up of chromosome and plasmid(s) and are frequently collated by combining
+multiple contigs together (metagenome assembled genomes).
+The contig embeddings are initialised and train from scratch at the beginning of pretraining.
+The genomic organisation is highly important for bacteria, to model it we employ **rotary positional embeddings** [2].
+The input embeddings are ordered by their order on the contig/chromosome/plasmid. This [...]. Additionally, we include special tokens. Specifically, 1) `[CLS]` token
+at the start of the sequence, 2) `[SEP]` token between the contigs or chromosomes(s)/plasmid(s), 3) `[END]` token at the end of the genome. The example below show how does
+the genome representation look like for complete genomes and MAGs.
+**Complete genomes**
+```
+[CLS] [chromosome1_gene1] [chromosome2_gene2] ... [chromosomme1_geneN] [SEP] [plasmid1_gene1] ... [plasmid1_geneM] [END]
+```
+**Metagenome-assembled genome (MAG)**
+```
+[CLS] [contig1_gene1] ... [contig1_geneN] [SEP] [contig2_gene1] ... [contig2_geneM] [SEP] ... [contigZgeneV] [END]
+```
+#### Transformer backbone
+The input embeddings are fed into a transformer, which computes protein representations conditional on other proteins present in the genome
+by computing self-attention between them, resulting in **contextual protein embeddings**.
+The transformer is 12 layer transformer with `hidden_dim=480` and is trained from scratch. Bacformer leverages the flash attention available in `pytorch>=2.2`.
+#### Pretraining classification head
+Bacformer is pretrained to predict masked or next protein family based on the other proteins present in the genome. Given the embedding from the last hidden state of
+the transformer of masked or the previous protein, the classification head predicts the protein family. We predict protein family, rather than a protein, because the space of
+possible proteins is effectively unbound. To get a discrete vocabulary of proteins, we we assigned each protein a discrete protein family index by performing
+unsupervised clustering on a set of proteins, resulting in `50k` distinct protein family clusters. Importantly, the input to Bacformer are
+exact protein sequences present in the whole bacterial genome, rather than protein family tokens. This allows the model to work on amino acid level tasks where even single mutations
+can change the phenotype of a genome, while still allowing for pretraining.
+## Citation
 **BibTeX:**
+```
+TBA
+```
+## Contact
+In case of questions/issues or feature requests please raise an issue on github - https://github.com/macwiatrak/Bacformer.