Molecular String-based Bamba Encoder-Decoder (STR-Bamba)

This repository provides PyTorch source code associated with our publication, "STR-Bamba: Multimodal Molecular Textual Representation Encoder-Decoder Foundation Model".

Paper: OpenReview Link

GitHub: GitHub Link

For more information contact: vshirasuna@ibm.com or evital@br.ibm.com.

Introduction

We present a large encoder-decoder chemical foundation model based on the IBM Bamba architecture, a hybrid of Transformers and Mamba-2 layers, designed to support multi-representational molecular string inputs. The model is pre-trained in a BERT-style on 588 million samples, resulting in a corpus of approximately 29 billion molecular tokens. These models serve as a foundation for language chemical research in supporting different complex tasks, including molecular properties prediction, classification, and molecular translation. Additionally, the STR-Bamba architecture allows for the aggregation of multiple representations in a single text input, as it does not contain any token length limitation, except for hardware limitations. Our experiments across multiple benchmark datasets demonstrate state-of-the-art performance for various tasks. Code details are available at: GitHub Link.

The STR-Bamba model supports the following molecular representations:

• SMILES
• SELFIES
• Molecular Formula
• InChI
• IUPAC Name
• Polymer SMILES in SPG notation
• Formulations

Table of Contents

1. Getting Started
   1. Pretrained Models and Training Logs
   2. Replicating Conda Environment
2. Pretraining
3. Finetuning
4. Feature Extraction
5. Citations

Getting Started

This code and environment have been tested on Nvidia V100s and Nvidia A100s

Pretrained Models and Training Logs

We provide checkpoints of the STR-Bamba model pre-trained on a dataset of ~118M small molecules, ~2M polymer structures, and 258 formulations. The pre-trained model shows competitive performance on classification and regression benchmarks across small and polymer molecules, and electrolyte formulations. For code details: GitHub Link

Add the STR-Bamba pre-trained weights.pt to the inference/ or finetune/ directory according to your needs. The directory structure should look like the following:

inference/
└── str_bamba/
    ├── config/
    ├── checkpoints/
    │   └── STR-Bamba_8.pt
    └── tokenizer/

and/or:

finetune/
└── str_bamba/
    ├── config/
    ├── checkpoints/
    │   └── STR-Bamba_8.pt
    └── tokenizer/

Replicating Conda Environment

Follow these steps to replicate our Conda environment and install the necessary libraries:

Create and Activate Conda Environment

mamba create -n strbamba python=3.10.13
mamba activate strbamba

PyTorch 2.4.0 and CUDA 12.4

pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 --index-url https://download.pytorch.org/whl/cu124

Mamba2 dependencies:

Install the following packages in this order and with a GPU, because mamba depends on causal-conv1d to be installed.

# causal-conv1d
git clone https://github.com/Dao-AILab/causal-conv1d.git
cd causal-conv1d && git checkout v1.5.0.post8 && pip install . && cd .. && rm -rf causal-conv1d

# mamba
git clone https://github.com/state-spaces/mamba.git
cd mamba && git checkout v2.2.4 && pip install --no-build-isolation . && cd .. && rm -rf mamba

# flash-attn
pip install flash-attn==2.6.1 --no-build-isolation

Install Packages with Pip

pip install -r requirements.txt

Troubleshooting

pip install mamba-ssm==2.2.4
MAX_JOBS=2 pip install flash-attn==2.6.1 --no-build-isolation --verbose

Pretraining

For pretraining, we use two strategies: the masked language model method to train the encoder part and a next token prediction strategy to train the decoder in order to refine molecular representation reconstruction and generation conditioned from the encoder.

The pretraining code provides examples of data processing and model training on a smaller dataset, requiring a A100 GPU.

To pre-train the two stages of the STR-Bamba model, run:

bash training/run_model_encoder_training.sh

or

bash training/run_model_decoder_training.sh

Finetuning

The finetuning datasets and environment can be found in the finetune directory. After setting up the environment, you can run a finetuning task with:

bash finetune/runs/esol/run_finetune_esol.sh

Finetuning training/checkpointing resources will be available in directories named checkpoint_<measure_name>.

Feature Extraction

To load STR-Bamba, you can simply use:

model = load_strbamba('STR-Bamba_8.pt')

To encode SMILES, SELFIES, InChI or other supported molecular representations into embeddings, you can use:

with torch.no_grad():
    encoded_embeddings = model.encode(df['SMILES'], return_torch=True)

For decoder, you can use the following code, so you can generate new molecular representations conditioned from the encoder:

with torch.no_grad():
    # encoder and decoder inputs
    encoder_input = '<smiles>CCO'
    decoder_input = '<smiles>'
    decoder_target = '<smiles>CCO'

    # tokenization
    encoder_input_ids = model.tokenizer(encoder_input, 
                                        padding=True, 
                                        truncation=True, 
                                        return_tensors='pt')['input_ids'].to(device)
    decoder_input_ids = model.tokenizer(decoder_input, 
                                        padding=True, 
                                        truncation=True, 
                                        return_tensors='pt')['input_ids'][:, :-1].to(device)
    decoder_target_ids = model.tokenizer(decoder_target, 
                                         padding=True, 
                                         truncation=True, 
                                         return_tensors='pt')['input_ids'].to(device)

    # visualize input texts
    print('Encoder input:', model.tokenizer.batch_decode(encoder_input_ids))
    print('Decoder input:', model.tokenizer.batch_decode(decoder_input_ids))
    print('Decoder target:', model.tokenizer.batch_decode(decoder_target_ids))
    print('Target:', decoder_target_ids)

    # encoder forward
    encoder_hidden_states = model.encoder(encoder_input_ids).hidden_states

    # model generation
    output = model.decoder.generate(
        input_ids=decoder_input_ids,
        encoder_hidden_states=encoder_hidden_states,
        max_length=decoder_target_ids.shape[1],
        cg=True,
        return_dict_in_generate=True,
        output_scores=True,
        enable_timing=False,
        temperature=1,
        top_k=1,
        top_p=1.0,
        min_p=0.,
        repetition_penalty=1,
    )

    # visualize model output
    generated_text = ''.join(
        ''.join(
            model.tokenizer.batch_decode(
                output.sequences, 
                clean_up_tokenization_spaces=True, 
                skip_special_tokens=False
            )
        ).split(' ')
    )
    print(generated_text)

Citations