[
{
"path": ".gitignore",
"content": "# Byte-compiled / optimized / DLL files\n__pycache__/\n*.py[cod]\n*$py.class\n\n# C extensions\n*.so\n\n# Distribution / packaging\n.Python\nbuild/\ndevelop-eggs/\ndist/\ndownloads/\neggs/\n.eggs/\nlib/\nlib64/\nparts/\nsdist/\nvar/\nwheels/\npip-wheel-metadata/\nshare/python-wheels/\n*.egg-info/\n.installed.cfg\n*.egg\nMANIFEST\n\n# PyInstaller\n# Usually these files are written by a python script from a template\n# before PyInstaller builds the exe, so as to inject date/other infos into it.\n*.manifest\n*.spec\n\n# Installer logs\npip-log.txt\npip-delete-this-directory.txt\n\n# Unit test / coverage reports\nhtmlcov/\n.tox/\n.nox/\n.coverage\n.coverage.*\n.cache\nnosetests.xml\ncoverage.xml\n*.cover\n*.py,cover\n.hypothesis/\n.pytest_cache/\n\n# Translations\n*.mo\n*.pot\n\n# Django stuff:\n*.log\nlocal_settings.py\ndb.sqlite3\ndb.sqlite3-journal\n\n# Flask stuff:\ninstance/\n.webassets-cache\n\n# Scrapy stuff:\n.scrapy\n\n# Sphinx documentation\ndocs/_build/\n\n# PyBuilder\ntarget/\n\n# Jupyter Notebook\n.ipynb_checkpoints\n\n# IPython\nprofile_default/\nipython_config.py\n\n# pyenv\n.python-version\n\n# pipenv\n# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.\n# However, in case of collaboration, if having platform-specific dependencies or dependencies\n# having no cross-platform support, pipenv may install dependencies that don't work, or not\n# install all needed dependencies.\n#Pipfile.lock\n\n# PEP 582; used by e.g. github.com/David-OConnor/pyflow\n__pypackages__/\n\n# Celery stuff\ncelerybeat-schedule\ncelerybeat.pid\n\n# SageMath parsed files\n*.sage.py\n\n# Environments\n.venv\nenv/\nvenv/\nENV/\nenv.bak/\nvenv.bak/\n\n# Spyder project settings\n.spyderproject\n.spyproject\n\n# Rope project settings\n.ropeproject\n\n# mkdocs documentation\n/site\n\n# mypy\n.mypy_cache/\n.dmypy.json\ndmypy.json\n\n# Pyre type checker\n.pyre/\n\n### VisualStudioCode\n.vscode/*\n!.vscode/settings.json\n!.vscode/tasks.json\n!.vscode/launch.json\n!.vscode/extensions.json\n*.code-workspace\n**/.vscode\n\n# JetBrains\n.idea/\n\n# Lightning-Hydra-Template\nconfigs/local/default.yaml\nconfigs/local/*\n!*/data\n/data/\nlogs/\nwandb/\n.env\n.autoenv\nworkspace.ipynb\nrun/logs\n\n# model weight\n*.ckpt\n\n# pdb\n*.pdb\n\nbyprot-checkpoints/\ngeneration-results/\ndata-bin/scaffolding-pdbs/*.pdb\ntemp.ipynb\n# run/\n"
},
{
"path": ".pre-commit-config.yaml",
"content": "default_language_version:\n python: python3.9\n\nexclude: ^(docs/|build/|node_modules/|venv/|\\.venv/|vendor/)\nrepos:\n - repo: https://github.com/pre-commit/pre-commit-hooks\n rev: v4.3.0\n hooks:\n # list of supported hooks: https://pre-commit.com/hooks.html\n - id: trailing-whitespace\n - id: end-of-file-fixer\n - id: check-docstring-first\n - id: check-yaml\n - id: debug-statements\n - id: detect-private-key\n - id: check-toml\n - id: check-case-conflict\n - id: check-added-large-files\n args: [--maxkb=1000]\n - id: check-json\n - id: check-merge-conflict\n - id: check-shebang-scripts-are-executable\n - id: fix-byte-order-marker\n - id: fix-encoding-pragma\n args: [--remove]\n - id: mixed-line-ending\n args: [--fix=lf]\n\n # python code formatting\n - repo: https://github.com/psf/black\n rev: 22.6.0\n hooks:\n - id: black\n args: [--line-length, \"79\"]\n language_version: python3.9\n\n # python import sorting\n - repo: https://github.com/PyCQA/isort\n rev: 5.12.0\n hooks:\n - id: isort\n args:\n [\n \"--line-length=79\",\n \"--multi-line=3\",\n \"--profile=black\",\n \"--filter-files\",\n ]\n language_version: python3.9\n\n # python upgrading syntax to newer version\n # - repo: https://github.com/asottile/pyupgrade\n # rev: v2.32.1\n # hooks:\n # - id: pyupgrade\n # args: [--py38-plus]\n\n # python docstring formatting\n - repo: https://github.com/myint/docformatter\n rev: v1.4\n hooks:\n - id: docformatter\n args: [--in-place, --wrap-summaries=79, --wrap-descriptions=79]\n\n # python check (PEP8), programming errors and code complexity\n - repo: https://github.com/PyCQA/flake8\n rev: 4.0.1\n hooks:\n - id: flake8\n args:\n [\n \"--extend-ignore\",\n \"E203,E402,E501,F401,F841\",\n \"--exclude\",\n \"logs/*,data/*\",\n ]\n\n # python security linter\n - repo: https://github.com/PyCQA/bandit\n rev: \"1.7.1\"\n hooks:\n - id: bandit\n args: [\"-s\", \"B101\"]\n\n # yaml formatting\n - repo: https://github.com/pre-commit/mirrors-prettier\n rev: v2.7.1\n hooks:\n - id: prettier\n types: [yaml]\n\n # shell scripts linter\n - repo: https://github.com/shellcheck-py/shellcheck-py\n rev: v0.8.0.4\n hooks:\n - id: shellcheck\n\n # # md formatting\n # - repo: https://github.com/executablebooks/mdformat\n # rev: 0.7.14\n # hooks:\n # - id: mdformat\n # args: [\"--number\"]\n # additional_dependencies:\n # - mdformat-gfm\n # - mdformat-tables\n # - mdformat_frontmatter\n # # - mdformat-toc\n # # - mdformat-black\n\n # jupyter notebook cell output clearing\n # - repo: https://github.com/kynan/nbstripout\n # rev: 0.5.0\n # hooks:\n # - id: nbstripout\n\n # jupyter notebook linting\n # - repo: https://github.com/nbQA-dev/nbQA\n # rev: 1.4.0\n # hooks:\n # - id: nbqa-black\n # args: [\"--line-length=79\"]\n # - id: nbqa-isort\n # args: [\"--profile=black\"]\n # - id: nbqa-flake8\n # args:\n # [\n # \"--extend-ignore=E203,E402,E501,F401,F841\",\n # \"--exclude=logs/*,data/*\",\n # ]\n"
},
{
"path": "LICENSE",
"content": "\n Apache License\n Version 2.0, January 2004\n http://www.apache.org/licenses/\n\n TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION\n\n 1. Definitions.\n\n \"License\" shall mean the terms and conditions for use, reproduction,\n and distribution as defined by Sections 1 through 9 of this document.\n\n \"Licensor\" shall mean the copyright owner or entity authorized by\n the copyright owner that is granting the License.\n\n \"Legal Entity\" shall mean the union of the acting entity and all\n other entities that control, are controlled by, or are under common\n control with that entity. For the purposes of this definition,\n \"control\" means (i) the power, direct or indirect, to cause the\n direction or management of such entity, whether by contract or\n otherwise, or (ii) ownership of fifty percent (50%) or more of the\n outstanding shares, or (iii) beneficial ownership of such entity.\n\n \"You\" (or \"Your\") shall mean an individual or Legal Entity\n exercising permissions granted by this License.\n\n \"Source\" form shall mean the preferred form for making modifications,\n including but not limited to software source code, documentation\n source, and configuration files.\n\n \"Object\" form shall mean any form resulting from mechanical\n transformation or translation of a Source form, including but\n not limited to compiled object code, generated documentation,\n and conversions to other media types.\n\n \"Work\" shall mean the work of authorship, whether in Source or\n Object form, made available under the License, as indicated by a\n copyright notice that is included in or attached to the work\n (an example is provided in the Appendix below).\n\n \"Derivative Works\" shall mean any work, whether in Source or Object\n form, that is based on (or derived from) the Work and for which the\n editorial revisions, annotations, elaborations, or other modifications\n represent, as a whole, an original work of authorship. For the purposes\n of this License, Derivative Works shall not include works that remain\n separable from, or merely link (or bind by name) to the interfaces of,\n the Work and Derivative Works thereof.\n\n \"Contribution\" shall mean any work of authorship, including\n the original version of the Work and any modifications or additions\n to that Work or Derivative Works thereof, that is intentionally\n submitted to Licensor for inclusion in the Work by the copyright owner\n or by an individual or Legal Entity authorized to submit on behalf of\n the copyright owner. For the purposes of this definition, \"submitted\"\n means any form of electronic, verbal, or written communication sent\n to the Licensor or its representatives, including but not limited to\n communication on electronic mailing lists, source code control systems,\n and issue tracking systems that are managed by, or on behalf of, the\n Licensor for the purpose of discussing and improving the Work, but\n excluding communication that is conspicuously marked or otherwise\n designated in writing by the copyright owner as \"Not a Contribution.\"\n\n \"Contributor\" shall mean Licensor and any individual or Legal Entity\n on behalf of whom a Contribution has been received by Licensor and\n subsequently incorporated within the Work.\n\n 2. Grant of Copyright License. Subject to the terms and conditions of\n this License, each Contributor hereby grants to You a perpetual,\n worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n copyright license to reproduce, prepare Derivative Works of,\n publicly display, publicly perform, sublicense, and distribute the\n Work and such Derivative Works in Source or Object form.\n\n 3. Grant of Patent License. Subject to the terms and conditions of\n this License, each Contributor hereby grants to You a perpetual,\n worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n (except as stated in this section) patent license to make, have made,\n use, offer to sell, sell, import, and otherwise transfer the Work,\n where such license applies only to those patent claims licensable\n by such Contributor that are necessarily infringed by their\n Contribution(s) alone or by combination of their Contribution(s)\n with the Work to which such Contribution(s) was submitted. If You\n institute patent litigation against any entity (including a\n cross-claim or counterclaim in a lawsuit) alleging that the Work\n or a Contribution incorporated within the Work constitutes direct\n or contributory patent infringement, then any patent licenses\n granted to You under this License for that Work shall terminate\n as of the date such litigation is filed.\n\n 4. Redistribution. You may reproduce and distribute copies of the\n Work or Derivative Works thereof in any medium, with or without\n modifications, and in Source or Object form, provided that You\n meet the following conditions:\n\n (a) You must give any other recipients of the Work or\n Derivative Works a copy of this License; and\n\n (b) You must cause any modified files to carry prominent notices\n stating that You changed the files; and\n\n (c) You must retain, in the Source form of any Derivative Works\n that You distribute, all copyright, patent, trademark, and\n attribution notices from the Source form of the Work,\n excluding those notices that do not pertain to any part of\n the Derivative Works; and\n\n (d) If the Work includes a \"NOTICE\" text file as part of its\n distribution, then any Derivative Works that You distribute must\n include a readable copy of the attribution notices contained\n within such NOTICE file, excluding those notices that do not\n pertain to any part of the Derivative Works, in at least one\n of the following places: within a NOTICE text file distributed\n as part of the Derivative Works; within the Source form or\n documentation, if provided along with the Derivative Works; or,\n within a display generated by the Derivative Works, if and\n wherever such third-party notices normally appear. The contents\n of the NOTICE file are for informational purposes only and\n do not modify the License. You may add Your own attribution\n notices within Derivative Works that You distribute, alongside\n or as an addendum to the NOTICE text from the Work, provided\n that such additional attribution notices cannot be construed\n as modifying the License.\n\n You may add Your own copyright statement to Your modifications and\n may provide additional or different license terms and conditions\n for use, reproduction, or distribution of Your modifications, or\n for any such Derivative Works as a whole, provided Your use,\n reproduction, and distribution of the Work otherwise complies with\n the conditions stated in this License.\n\n 5. Submission of Contributions. Unless You explicitly state otherwise,\n any Contribution intentionally submitted for inclusion in the Work\n by You to the Licensor shall be under the terms and conditions of\n this License, without any additional terms or conditions.\n Notwithstanding the above, nothing herein shall supersede or modify\n the terms of any separate license agreement you may have executed\n with Licensor regarding such Contributions.\n\n 6. Trademarks. This License does not grant permission to use the trade\n names, trademarks, service marks, or product names of the Licensor,\n except as required for reasonable and customary use in describing the\n origin of the Work and reproducing the content of the NOTICE file.\n\n 7. Disclaimer of Warranty. Unless required by applicable law or\n agreed to in writing, Licensor provides the Work (and each\n Contributor provides its Contributions) on an \"AS IS\" BASIS,\n WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or\n implied, including, without limitation, any warranties or conditions\n of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A\n PARTICULAR PURPOSE. You are solely responsible for determining the\n appropriateness of using or redistributing the Work and assume any\n risks associated with Your exercise of permissions under this License.\n\n 8. Limitation of Liability. In no event and under no legal theory,\n whether in tort (including negligence), contract, or otherwise,\n unless required by applicable law (such as deliberate and grossly\n negligent acts) or agreed to in writing, shall any Contributor be\n liable to You for damages, including any direct, indirect, special,\n incidental, or consequential damages of any character arising as a\n result of this License or out of the use or inability to use the\n Work (including but not limited to damages for loss of goodwill,\n work stoppage, computer failure or malfunction, or any and all\n other commercial damages or losses), even if such Contributor\n has been advised of the possibility of such damages.\n\n 9. Accepting Warranty or Additional Liability. While redistributing\n the Work or Derivative Works thereof, You may choose to offer,\n and charge a fee for, acceptance of support, warranty, indemnity,\n or other liability obligations and/or rights consistent with this\n License. However, in accepting such obligations, You may act only\n on Your own behalf and on Your sole responsibility, not on behalf\n of any other Contributor, and only if You agree to indemnify,\n defend, and hold each Contributor harmless for any liability\n incurred by, or claims asserted against, such Contributor by reason\n of your accepting any such warranty or additional liability.\n\n END OF TERMS AND CONDITIONS\n\n APPENDIX: How to apply the Apache License to your work.\n\n To apply the Apache License to your work, attach the following\n boilerplate notice, with the fields enclosed by brackets \"[]\"\n replaced with your own identifying information. (Don't include\n the brackets!) The text should be enclosed in the appropriate\n comment syntax for the file format. We also recommend that a\n file or class name and description of purpose be included on the\n same \"printed page\" as the copyright notice for easier\n identification within third-party archives.\n\n Copyright [yyyy] [name of copyright owner]\n\n Licensed under the Apache License, Version 2.0 (the \"License\");\n you may not use this file except in compliance with the License.\n You may obtain a copy of the License at\n\n http://www.apache.org/licenses/LICENSE-2.0\n\n Unless required by applicable law or agreed to in writing, software\n distributed under the License is distributed on an \"AS IS\" BASIS,\n WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n See the License for the specific language governing permissions and\n limitations under the License.\n"
},
{
"path": "README.md",
"content": "\n\n# The Family of Diffusion Protein Language Models (DPLM)\n![\"PyTorch\"](\"https://img.shields.io/badge/PyTorch-ee4c2c?logo=pytorch&logoColor=white\")
\n![\"Lightning\"](\"https://img.shields.io/badge/-Lightning-792ee5?logo=pytorchlightning&logoColor=white\")
\n![\"Config:](\"https://img.shields.io/badge/Config-Hydra-89b8cd\")
\n![\"Template\"](\"https://img.shields.io/badge/-Lightning--Hydra--Template-017F2F?style=flat&logo=github&labelColor=gray\")
\n\n## Overview 🌟\n\nThis repository contains the official implementation of training and inference as well as the pre-trained weights for the Family of Diffusion Protein Language Models (DPLM), including:\n- `DPLM` from ICML'24 paper [\"Diffusion Language Models Are Versatile Protein Learners\"](https://arxiv.org/abs/2402.18567), which introduces **d**iffusion **p**rotein **l**anguage **m**odel (DPLM), a versatile protein language model that demonstrates strong generative and predictive capabilities for protein sequences.\n- `DPLM-2` from ICLR'25 paper [\"DPLM-2: A Multimodal Diffusion Protein Language Model\"](https://arxiv.org/abs/2410.13782), a multimodal protein foundation model that extends discrete diffusion protein language model to accommodate both sequences and structures.\n- ICML'25 spotlight paper [\"Elucidating the Design Space of Multimodal Protein Language Models\"](https://arxiv.org/abs/2504.11454), where we elucidate the challenges of structure modeling of multimodal protein language models (e.g., DPLM-2 and ESM3) and propose advanced designs for better structure modeling. We have released the finer-grained bit-based generative modeling (`DPLM-2 Bit`). The full implementation of the paper will be released soon.\n\n## Key Features 🔑\nSpecifically, the DPLM family exhibits impressive performance in protein (structure and sequence) co-generation, any-to-any conditional generation (e.g., folding, inverse folding, and motif scaffolding), and representation learning.\nWe develop DPLM based on the [ByProt](https://github.com/BytedProtein/ByProt). This repository contains pretraining scripts for DPLM and running scripts for various protein generation and understanding tasks, as detailed below:\n- **Unconditional protein generation**:\n **DPLM** is capable of unconditionally generating protein sequences with reasonable predicted structures. **DPLM-2** can generate diverse and highly plausible proteins through simultaneous structure-sequence co-generation.\n- **Sequence-conditioned generation (forward folding)**:\n DPLM-2 can generate reasonable protein structure given the input protein sequence, achieving close performance with the strong folding model (e.g., ESMFold).\n- **Structure-conditioned generation (inverse folding)**:\n DPLM and DPLM-2 can produce sequences that can confidently fold into the given backbone structure.\n- **Motif scaffolding**:\n DPLM can generate reasonable scaffold sequences given specific functional motifs. DPLM-2 achieves more successful motif scaffolding through multimodal motif conditioning.\n- **Representation learning**:\n DPLM is a superior protein sequence representation learner, while DPLM-2 offers structure-aware protein represenrations, demonstrating impressive performance across a variety of protein predictive tasks.\n- **Controllable generation**:\n DPLM enjoys plug-and-play programmability, generating samples satisfying provided secondary structure annotations.\n\n**TODOs**\n\n- [ ] Controllable/guided generation with discrete diffusion classifier guidance.\n- [ ] Representation learning of DPLM-2\n\n\n\n## DPLM\n> [\"Diffusion Language Models Are Versatile Protein Learners.\" Wang et al., In ICML 2024](https://arxiv.org/abs/2402.18567)\n\n\n\n## DPLM-2\n> [\"DPLM-2: A Multimodal Diffusion Protein Language Model.\" Wang et al., In ICLR 2025](https://arxiv.org/abs/2410.13782)\n\n\n\n## Updates 📢\n- **[2025-07]** We update the default sampling strategy of **DPLM-2** to `annealing@2.0:0.1`.\n- **[2025-04]** Our latest work **DPLM-2.1**, which focuses on analysis and better protein structure modeling of multimodal protein language models, is accepted to ICML'25 Spotlight! Check [Elucidating the Design Space of Multimodal Protein Language Models](https://arxiv.org/abs/2504.11454). We have release the implementation of finer-grained and better structure modeling (**DPLM-2 Bit**). The full implementation will be released soon.\n- **[2024-10]** Check out our new work [DPLM-2](https://arxiv.org/abs/2410.13782), a multimodal protein foundation model that extends DPLM to simultaneously model, understand, and generate both sequences and structures!\n- **[2024-03]** We release [DPLM](https://arxiv.org/abs/2402.18567), a versatile protein language model that demonstrates strong generative and predictive capabilities for protein sequences!\n\n## Table of Contents 📚\n\n- [Quick Start](#quick-start)\n - [Installation](#installation)\n - [Load Pretrained Models](#load-pretrained-models)\n - [Generation Examples](#generation-examples)\n - [Model Checkpoints](#model-checkpoints)\n- [Advanced Usage](#advanced-usage)\n - [Training](#training)\n - [Unconditional protein (co-)generation](#unconditional-protein-co-generation)\n - [Protein sequence generation (DPLM)](#protein-sequence-generation-dplm)\n - [Protein sequence-structure co-generation (DPLM-2 & DPLM-2-Bit)](#protein-sequence-structure-co-generation-dplm-2--dplm-2-bit)\n - [Sequence-conditioned Generation: Forward Folding](#sequence-conditioned-generation-forward-folding)\n - [Structure-conditioned generation: inverse folding](#structure-conditioned-generation-inverse-folding)\n - [Motif scaffolding](#motif-scaffolding)\n - [Representation Learning](#representation-learning)\n- [Acknowledgements](#acknowledgements)\n- [Citation](#citation)\n\n\n# Quick Start\n\n## Installation\n\n```bash\n# clone project\ngit clone --recursive https://url/to/this/repo/dplm.git\ncd dplm\n\n# create conda virtual environment\nenv_name=dplm\n\nconda create -n ${env_name} python=3.9 pip\nconda activate ${env_name}\n\n# automatically install everything else\nbash scripts/install.sh\n```\n## Load Pretrained Models\nUsers can load DPLM/DPLM-2 checkpoint by:\n\n```python\nfrom byprot.models.dplm import DiffusionProteinLanguageModel as DPLM\nfrom byprot.models.dplm2 import MultimodalDiffusionProteinLanguageModel as DPLM2\nfrom byprot.models.dplm2 import DPLM2Bit\n\ndplm = DPLM.from_pretrained(\"airkingbd/dplm_650m\").cuda()\ndplm2 = DPLM2.from_pretrained(\"airkingbd/dplm2_650m\").cuda()\ndplm2_bit = DPLM2Bit.from_pretrained(\"airkingbd/dplm2_bit_650m\").cuda()\n```\n\n## Generation Examples\n**Protein sequence generation**\n\n```python\nfrom generate_dplm import initialize_generation\n\ninput_tokens = initialize_generation(\n length=200,\n num_seqs=5,\n tokenizer=dplm.tokenizer,\n device=next(dplm.parameters()).device\n)\nsamples = dplm.generate(\n input_tokens=input_tokens,\n max_iter=500,\n)\nprint([''.join(seq.split(' ')) for seq in dplm.tokenizer.batch_decode(samples, skip_special_tokens=True)])\n```\n\n**Protein sequence-structure co-generation**\n\nUser can check the generated sequence and structure in the `./generation-results` folder.\n```python\nfrom generate_dplm2 import initialize_generation, save_results\n\ninput_tokens = initialize_generation(\n task=\"co_generation\",\n length=200,\n num_seqs=5,\n tokenizer=dplm2.tokenizer,\n device=next(dplm2.parameters()).device\n)[0]\n\nsamples = dplm2.generate(\n input_tokens=input_tokens,\n max_iter=500,\n)\nsave_results(\n outputs=samples,\n task=\"co_generation\",\n save_dir=\"./generation-results/dplm2_generation\",\n tokenizer=dplm2.tokenizer,\n struct_tokenizer=dplm2.struct_tokenizer, save_pdb=True\n)\n\nsamples = dplm2_bit.generate(\n input_tokens=input_tokens,\n max_iter=500,\n)\nsave_results(\n outputs=samples,\n task=\"co_generation\",\n save_dir=\"./generation-results/dplm2_bit_generation\",\n tokenizer=dplm2_bit.tokenizer,\n struct_tokenizer=dplm2_bit.struct_tokenizer\n)\n```\n\n## Model Checkpoints\n\nAccess pretrained models in varying sizes:\n\n| Model name | Model size |\n| ------------------------------------------------------------ | --------------- |\n| [dplm-150m](https://huggingface.co/airkingbd/dplm_150m/tree/main) | 150M parameters |\n| [dplm-650m](https://huggingface.co/airkingbd/dplm_650m/tree/main) | 650M parameters |\n| [dplm-3b](https://huggingface.co/airkingbd/dplm_3b/tree/main) | 3B parameters |\n| [dplm2-150m](https://huggingface.co/airkingbd/dplm2_150m/tree/main) | 150M parameters |\n| [dplm2-650m](https://huggingface.co/airkingbd/dplm2_650m/tree/main) | 650M parameters |\n| [dplm2-3b](https://huggingface.co/airkingbd/dplm2_3b/tree/main) | 3B parameters |\n| [dplm2-bit-650m](https://huggingface.co/airkingbd/dplm2_bit_650m/tree/main) | 650M parameters |\n\n# Advanced Usage\n\n## Training\n\n\n### DPLM\n\n\n#### Dataset\n\nWe pretrain DPLM on the UniRef50 dataset, which contains about 42 million protein sequences. We obtain the preprocessed UniRef50 dataset provided by [EvoDiff (Alamdari et al, 2023)](https://www.biorxiv.org/content/10.1101/2023.09.11.556673v1), which can be downloaded from this [link](https://zenodo.org/record/6564798). After downloading, please place the dataset in the `./data-bin/uniref50` folder.\n\nWe also provide the preprocessed dataset in [HuggingFace datasets](https://huggingface.co/datasets/airkingbd/uniref50) format, which we recommend to use. User can download the HF dataset locally in advance for faster loading by:\n```bash\nbash scripts/download_uniref50_hf.sh\n```\n\n\n#### Example of training\n\nWe train DPLM with approximately 1 million tokens per batch for 100,000 training steps.\n\nThe following command is run on one node with 8 A100 GPUs. If you want to train on multiple nodes, you can adjust the total number of tokens by ensuring that `max_tokens` \\* `accumulate_grad_batches`\\*`#GPUs` is approximately 1 million.\n\n```bash\nexport CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7\n\nmax_tokens=8192\naccumulate_grad_batches=16\n# this means the effective batch size is #GPUs(8) * max_tokens(8192) * accumulate_grad_batches(16), resulting in approximately 1 million.\n\nexp=dplm/dplm_650m\nmodel_name=dplm_650m\n\npython train.py \\\n experiment=${exp} name=${model_name} \\\n datamodule.max_tokens=${max_tokens} \\\n trainer.accumulate_grad_batches=${accumulate_grad_batches}\n```\n\nYou can adjust the other training configurations in the `configs/experiment/dplm/dplm_650m.yaml` as needed.\n\n\n### DPLM-2\n\n\n#### Dataset\nWe use the experimental structures from [PDB](https://pubmed.ncbi.nlm.nih.gov/10592235/) and AF2-predicted structures from [SwissProt](https://academic.oup.com/nar/article/50/D1/D439/6430488) dataset as training data for DPLM-2. We provide a preprocessed [HuggingFace dataset](https://huggingface.co/datasets/airkingbd/pdb_swissprot) of PDB and SwissProt. User can download the HF dataset locally in advance for faster loading by:\n```bash\nbash scripts/download_pdb_swissprot.sh\n```\n\n\n#### Example of training\nAs noted in section 3.2 in [DPLM-2](https://arxiv.org/abs/2410.13782) paper, we propose an efficient warm-up training strategy to mitigate the scarcity of structure training data. During training, we initialize the DPLM-2 model with pretrained DPLM checkpoint, to leverage the evolutionary knowledge captured by sequence-based pLM during large-scale sequence pretraining, which is beneficial for structure modeling.\n\nWe train DPLM-2 with approximately 64,000 tokens per batch for 100,000 training steps. To preserve the evolutionary knowledge captured by DPLM, we use the [LoRA](github.com/peft) to prevent large parameter shifts. The training command is as follows:\n\n```bash\nexport CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7\n\nmax_tokens=8192\naccumulate_grad_batches=1\n# this means the effective batch size is #GPUs(8) * max_tokens(8192) * accumulate_grad_batches(1), resulting in approximately 64 thousand.\n\nexp=dplm2/dplm2_650m\nmodel_name=dplm2_650m\n\npython train.py \\\n experiment=${exp} name=${model_name} \\\n datamodule.max_tokens=${max_tokens} \\\n trainer.accumulate_grad_batches=${accumulate_grad_batches}\n```\n\n\n### DPLM-2 Bit-based Modeling\n\nIn our latest work [DPLM-2.1](https://arxiv.org/abs/2504.11454), we show that the index-based structure token is challenging for the model to predict. A finer-grained, bit-based modeling approach in the latent space (i.e., predicting each bit of the quantized structure feature instead of the index) leads to better structural modeling and generation performance.\n\nThe training dataset is same to DPLM-2, and the training command is as below:\n```bash\nexport CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7\n\nmax_tokens=8192\naccumulate_grad_batches=1\n# this means the effective batch size is #GPU(8) * max_tokens(8192) * accumulate_grad_batches(1), resulting in approximately 64 thousand.\n\nexp=dplm2/dplm2_bit_650m\nmodel_name=dplm2_bit_650m\n\npython train.py \\\n experiment=${exp} name=${model_name} \\\n datamodule.max_tokens=${max_tokens} \\\n trainer.accumulate_grad_batches=${accumulate_grad_batches}\n```\n\n\n## Unconditional protein (co-)generation\n\n\n### Protein sequence generation (DPLM)\n\nThe results of unconditional protein sequence generation of DPLM of different scales (150M, 650M, 3B) are shown in the table below. For more details, please refer to our paper.\n\n\n| Length | 100 | 200 | 300 | 400 | 500 | 600 | 700 | 800 | 900 | 1000 |\n| ------ | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | -------------- | -------------- |\n| 150M | 73.31 | 84.30 | 84.82 | 86.90 | 81.71 | 81.53 | 81.56 | 80.92 | 78.71 | 72.10 |\n| 650M | 74.00 (+0.69) | 85.61 (+1.31) | 85.91 (+1.09) | 88.16 (+1.26) | 82.58 (+0.87) | 84.38 (+2.85) | 83.87 (+2.31) | 83.00 (+2.08) | 84.92 (+6.21) | 81.51 (+9.41) |\n| 3B | 77.78 (+4.47) | 86.16 (+1.86) | 87.39 (+2.57) | 90.06 (+3.16) | 87.43 (+5.72) | 86.01 (+4.48) | 84.64 (+3.08) | 85.88 (+4.96) | 85.93 (+7.22) | 83.86 (+11.76) |\n\nTo generate new protein sequences using a pre-trained DPLM model:\n\n```bash\nmodel_name=dplm_650m # choose from dplm_150m, dplm_650m, dplm_3b\noutput_dir=generation-results/${model_name}/uncond_generation\n\nmkdir -p generation-results\n\npython generate_dplm.py --model_name airkingbd/${model_name} \\\n\t--seq_lens 100 200 300 400 500 \\\n\t--saveto ${output_dir}\n\n# Evaluation\nbash anylasis/plddt_calculate.sh ${output_dir} # compute pLDDT using ESMFold\n```\n\nWe also provide evaluation scripts in the `analysis` folder. Users can use the `analysis/uncond_analysis.ipynb` to obtain average pLDDT score of each length and draw the line chart of the pLDDT score.\n\n\n### Protein sequence-structure co-generation (DPLM-2 & DPLM-2-Bit)\n\nDPLM-2 can generate diverse and highly-plausible protein with simultaneous structure-sequence co-generation.\n\n![\"Descriptive](\"./assets/co_generation.png\")
\n\nUser can co-generate sequence and structure simultaneously with the command below:\n\n\n```bash\n# choose from dplm2_150m, dplm2_650m, dplm2_3b\nmodel_name=dplm2_650m\n# About the default sampling strategy, annealing@2.0:0.1,\n# which anneals the temperature from 2.0 to 0.1. \n# It begins with high randomness to maximize diversity \n# and concludes with low randomness to ensure designability.\n# This achieves a better trade-off between the quality and diversity.\nsampling_strategy=annealing@2.0:0.1\n\noutput_dir=generation-results/${model_name}\ntask=co_generation\n\nmkdir -p ${output_dir}\n\npython generate_dplm2.py \\\n --model_name airkingbd/${model_name} \\\n --task ${task} \\\n --sampling_strategy ${sampling_strategy} \\\n --num_seqs 50 \\\n --max_iter 500 \\\n --seq_lens 100 200 300 400 500 \\\n --saveto ${output_dir}\n\n# Evaluation\ninput_fasta_dir=${output_dir}/co_generation\npython src/byprot/utils/protein/evaluator_dplm2.py -cn unconditional_codesign \\\n inference.input_fasta_dir=${input_fasta_dir}\n```\nUser can use `analysis/plot.ipynb` to plot the rmsd, tmscore distribution and diversity of each length.\n\nCo-generate sequence and structure with dplm-2 bit modeling variant:\n```bash\nmodel_name=dplm2_bit_650m\nsampling_strategy=annealing@1.1:0.1\n\noutput_dir=generation-results/${model_name}\ntask=co_generation\n\nmkdir -p ${output_dir}\n\npython generate_dplm2.py \\\n --model_name airkingbd/${model_name} \\\n --task ${task} \\\n --bit_model \\\n --sampling_strategy ${sampling_strategy} \\\n --num_seqs 50 \\\n --max_iter 500 \\\n --seq_lens 100 200 300 400 500 \\\n --saveto ${output_dir}\n```\n\n## Sequence-conditioned Generation: Forward Folding\nDPLM-2 spontaneously enables protein structure prediction given sequence (i.e., folding) in a zero-shot manner.\nWe use the [CAMEO 2022 (provided by EigenFold)](https://github.com/bjing2016/EigenFold) and a [PDB date split (provided by MultiFlow)](https://github.com/jasonkyuyim/multiflow) as testsets, and we provide our preprocessed dataset in this [link](https://zenodo.org/records/15424801), and can be downloaded by:\n```bash\nbash script/download_metadata.sh\n```\nPartial results are shown in the table below. For more details, please refer to [DPLM-2.1](https://arxiv.org/abs/2504.11454) paper.\n| Models | CAMEO 2022 | | PDB date | |\n|---|---|---|---|---|\n| | rmsd | tmscore | Rmsd | tmscore |\n| ESMFold | 3.99 | 0.85 | 2.84 | 0.93 |\n| DPLM-2 | 7.70 | 0.79 | 5.30 | 0.83 |\n| DPLM-2 Bit | 6.40 | 0.84 | 3.22 | 0.90 |\n\nThe folding generation and evaluation script is as follows.\nWe utilize RMSD and TMscore between the predicted and ground truth structures for evaluation. DPLM-2 adopts argmax decoding for 100 sampling iterations.\n\n```bash\nmodel_name=dplm2_650m\noutput_dir=generation-results/${model_name}\ntask=folding\n\nmkdir -p ${output_dir}\n\ninput_fasta_path=data-bin/cameo2022/aatype.fasta\npython generate_dplm2.py \\\n --model_name airkingbd/${model_name} \\\n --task ${task} \\\n --input_fasta_path ${input_fasta_path} \\\n --max_iter 100 \\\n --unmasking_strategy deterministic \\\n --sampling_strategy argmax \\\n --saveto ${output_dir}\n\n# Evaluation\ninput_fasta_dir=${output_dir}/folding\npython src/byprot/utils/protein/evaluator_dplm2.py -cn forward_folding inference.input_fasta_dir=${input_fasta_dir}\n```\n\nFor structure prediction conditioned on other customized sequences, users can input a FASTA file and modify the `input_fasta_path` variable to generate the predicted structure.\n\n## Structure-conditioned generation: inverse folding\n\nDPLM family can perform inverse folding in different ways according to DPLM variant. DPLM performs inverse folding by placing an adapter layer on the top of pLM, similar to [LM-Design](https://github.com/BytedProtein/ByProt). On the other hand, DPLM-2 directly conditions on the tokenized structure tokens to predict the sequence.\n\n\n### Inverse Folding with DPLM\n\nPartial results on the CATH 4.3 dataset are shown in the table below. For more details, please refer to our paper.\n\n| Models | Trainable Params. | AAR | scTM | pLDDT |\n|-----------|-------------------|-----------|----------|-----------|\n| LM-Design | 6.3M/650M | 56.49 | 0.85 | 74.89 |\n| DPLM-150M | 3.1M/150M | 53.27 | 0.85 | 75.31 |\n| DPLM-650M | 6.3M/650M | _56.61_ | _0.86_ | _76.78_ |\n| DPLM-3B | 68.2M/3.0B | **58.64** | **0.86** | **76.95** |\n\n\n#### Data\n\n**Download the preproceesd CATH datasets**\n\n- CATH 4.2 dataset provided by [Generative Models for Graph-Based Protein Design (Ingraham et al, NeurIPS'19)](https://papers.nips.cc/paper/2019/hash/f3a4ff4839c56a5f460c88cce3666a2b-Abstract.html)\n- CATH 4.3 dataset provided by [Learning inverse folding from millions of predicted structures (Hsu et al, ICML'22)](https://www.biorxiv.org/content/10.1101/2022.04.10.487779v1)\n\n```bash\nbash scripts/download_cath.sh\n```\n\n#### Training\n\nWe train structure-conditional DPLM based on the [LM-Design](https://github.com/BytedProtein/ByProt) framework, designating the pre-trained protein language model as DPLM. The training script is as below.\n\n```bash\nexp=dplm/dplm_650m_invfold\ndataset=cath_4.3\nname=${dataset}/dplm_650m/invfold\n\npython train.py \\\n experiment=${exp} datamodule=${dataset} name=${name} \\\n logger=tensorboard trainer=ddp_fp16\n```\n\n\n#### Evaluation on valid/test datasets\n\nUsers can set the `eval_sc` to `true` to calculate the self-consistency TMscore and pLDDT, which will result in a significant evaluation time overhead.\n\n```bash\ndataset=cath_4.3\nexp_path=${dataset}/dplm_650m/invfold\neval_sc=false\n# if set ${eval_sc} to true, the program will calculate the self-consistency\n# TMscore and pLDDT during generation,\n# thus siginificantly increase the evaluation time.\n\npython test.py \\\n experiment_path=${exp_path} \\\n data_split=test ckpt_path=best.ckpt mode=predict \\\n task.generator.max_iter=100 task.generator.eval_sc=${eval_sc}\n```\n\n\n### Inverse Folding with DPLM-2\nWe provide the CAMEO 2022 and PDB date test set split used in our paper, where the structure has been tokenized and saved to `data-bin/cameo2022/struct.fasta` and `data-bin/PDB_date/struct.fasta`.\nUser can use the following script to do the inverse folding and evaluation.\n\n```bash\nmodel_name=dplm2_650m\noutput_dir=generation-results/${model_name}\ntask=inverse_folding\n\nmkdir -p ${output_dir}\n\ninput_fasta_path=data-bin/cameo2022/struct.fasta\npython generate_dplm2.py \\\n --model_name airkingbd/${model_name} \\\n --task ${task} \\\n --input_fasta_path ${input_fasta_path} \\\n --max_iter 100 \\\n --unmasking_strategy deterministic \\\n --sampling_strategy argmax \\\n --saveto ${output_dir}\n\n# Evaluation\ninput_fasta_dir=${output_dir}/inverse_folding\npython src/byprot/utils/protein/evaluator_dplm2.py -cn inverse_folding inference.input_fasta_dir=${input_fasta_dir}\n```\nFor any customized input structure, user can first tokenize the structure with structure tokenizer and save it to a FASTA file using the following script:\n```bash\n# Tokenize\n# each protein is represented by a pdb file\ninput_pdb_folder=/path/to/your/input/structure\n# this will save two fasta files in the ${input_pdb_folder}/tokenized_protein folder:\n# 1) struct.fasta, containing the tokenized structure tokens\n# 2) aatype.fasta, containing the amino acid tokens.\npython src/byprot/utils/protein/tokenize_pdb.py --input_pdb_folder ${input_pdb_folder} --output_dir ${input_pdb_folder}/tokenized_protein\n```\nThen user can specify the path of generated `struct.fasta` as input and predict the sequence.\n\n## Motif scaffolding\nDPLM and DPLM-2 can both perform motif scaffolding. DPLM can condition on the motif sequence and predict the scaffold sequence. DPLM-2 is able to condition on both the sequence and structure of the motif and simultaneously co-generate the sequence and structure of the scaffold part, which leads to better performance.\n\nWe examine on the benchmark, provided by [FrameFlow](https://github.com/microsoft/protein-frame-flow/blob/main/motif_scaffolding/benchmark.csv). We use the motif pdb files which are provided by [EvoDiff](https://github.com/microsoft/evodiff/tree/main/examples/scaffolding-pdbs), and we also provide the pdbs and the corresponding structure tokens in this [link](https://zenodo.org/records/15424801). You can download the dataset by\n```bash\nbash scripts/download_motif_scaffolds.sh\n```\nFor each motif-scaffolding problem, we sample 100 sequences and then calculate the success rate according to two aspects: motif part consistency and overall quality. For motif part consistency, we use the motif-RMSD < 1$\\AA$ as the success criterion. For overall quality, the assessment varies across different approaches: sequence-based method (DPLM) we use pLDDT > 70, while for co-generation method (DPLM-2) we use scTM > 0.8. For more details, please refer to our paper.\n\nThe success rate of each motif-scaffold problem is shown below.\n\n| | Pass rate | Avg. Success rate | 1BCF | 1PRW | 1QJG | 1YCR | 2KL8 | 3IXT | 4JHW | 4ZYP | 5IUS | 5TPN | 5TRV_long | 5TRV_med | 5TRV_short | 5WN9 | 5YUI | 6E6R_long | 6E6R_med | 6E6R_short | 6EXZ_long | 6EXZ_med | 6EXZ_short | 7MRX_long | 7MRX_med | 7MRX_short |\n|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| DPLM | 11/24 | 0.19 | 0.00 | 0.83 | 0.00 | 0.38 | 0.08 | 0.17 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.65 | 0.94 | 0.87 | 0.01 | 0.00 | 0.00 | 0.02 | 0.31 | 0.34 |\n| DPLM-2 | 18/24 | 0.29 | 0.01 | 0.84 | 0.02 | 0.53 | 0.57 | 0.41 | 0.00 | 0.10 | 0.00 | 0.00 | 0.00 | 0.02 | 0.03 | 0.00 | 0.00 | 0.78 | 0.77 | 0.64 | 0.44 | 0.55 | 0.58 | 0.20 | 0.22 | 0.24 |\n\n\n\n### DPLM\n\nWe provide the following script to sample sequences for each motif-scaffolding problem. Note that before generation, you should download the motif pdbs and place them in the `data-bin/scaffolding-pdbs` folder.\n\n```bash\nexport CUDA_VISIBLE_DEVICES=0\n\nmodel_name=dplm_650m\noutput_dir=./generation-results/${model_name}/motif_scaffold\n\nmkdir -p generation-results\n\n# Generate scaffold\npython run/scaffold_generate_dplm.py \\\n --model_name airkingbd/${model_name} \\\n --num_seqs 100 \\\n --saveto $output_dir\n\n# Predict structure by ESMFold\nmax_tokens=1024\npdb_path=$output_dir/scaffold_fasta/esmfold_pdb\n\n# folding\nmkdir -p $pdb_path\n\necho 'folding by ESMFold'\noutput_filename_list=$(ls ${output_dir}/scaffold_fasta)\necho $output_filename_list\n\npython analysis/cal_plddt_dir.py -i ${output_dir}/scaffold_fasta -o ${pdb_path} --max-tokens-per-batch ${max_tokens}\n```\n\nFor evaluation, users can use the `analysis/motif_analysis.ipynb` to obtain success rate of each problem.\n\n\n### DPLM-2\nBefore generation, the FASTA file of tokenized structure tokens and amino acid tokens of the motif should be in the `data-bin/scaffolding-pdbs` folder. Users can co-generate the scaffold sequence and structure, conditioning on the sequence and structure of the motif part.\n```bash\nexport CUDA_VISIBLE_DEVICES=0\n\nmodel_name=dplm2_650m\noutput_dir=./generation-results/${model_name}/motif_scaffold\n\nmkdir -p generation-results\n\n# Generate scaffold\npython run/scaffold_generate_dplm2.py \\\n --model_name airkingbd/${model_name} \\\n --num_seqs 100 \\\n --saveto ${output_dir}\n\n# Predict structure by ESMFold\nmax_tokens=1024\npython analysis/cal_plddt_dir.py -i ${output_dir}/scaffold_fasta --max-tokens-per-batch ${max_tokens}\n# Calculate sc-TMscore\npython src/byprot/utils/protein/evaluator_dplm2.py -cn unconditional_codesign \\\n inference.input_fasta_dir=${output_dir}/scaffold_fasta inference.calculate_diversity=false\n```\nFor evaluation, users can use the `analysis/motif_analysis.ipynb` to obtain success rate of each problem.\n\n## Representation Learning\n\nThe DPLM family excels in various downstream protein predictive tasks. DPLM is a superior protein sequence representation learner, while DPLM-2 can perform multimodal representation learning by leveraging both structure and sequence information, demonstrating its versatility and effectiveness. The following table summarizes the DPLM family performance, and the italic number means performance of DPLM-2, which offers structure-aware protein representations and outperforms sequence-based DPLM on most of the predictive tasks. Meanwhile, we also find the performance improves along with the model size.\n\n| Models | Thermostability | HumanPPI | Metal Ion Binding | EC | GO-MF | GO-BP | GO-CC | DeepLoc-Subcellular | DeepLoc-Binary |\n| --------------------- | --------------- | --------- | ----------------- | --------- | :-------: | :-------: | :-------: | ------------------- | -------------- |\n| ESM2 (650M) | 0.691 | 84.78 | 71.88 | 0.866 | 0.676 | 0.344 | 0.402 | 83.68 | 92.28 |\n| AR-LM | 0.638 | 68.48 | 61.66 | 0.691 | 0.566 | 0.258 | 0.287 | 68.53 | 88.31 |\n| DPLM (150M) | 0.687 | 80.98 | 72.17 | 0.822 | 0.662 | 0.328 | 0.379 | 82.41 | 92.63 |\n| DPLM (650M) | 0.695 | 86.41 | 75.15 | 0.875 | 0.680 | 0.357 | 0.409 | 84.56 | 93.09 |\n| DPLM-2 (650M) | **_0.714_** | _84.44_ | _74.28_ | _0.878_ | _0.680_ | _0.359_ | _0.411_ | 82.98 | _93.64_ |\n| *DPLM-2 (650M) | -- | _87.78_ | -- | --| --| -- | -- | _83.42_ | -- |\n| DPLM (3B) | 0.704 | **90.00** | **75.94** | **0.883** | **0.687** | **0.369** | **0.463** | **85.32** | **93.93** |\n\n> We find DPLM-2 demonstrates a performance degradation on some tasks (e.g., HumanPPI and DeepLoc-Subcellular), due to continue training on smaller magnitude of structure data and result in overfitting and degradation of the representations learned during large-scale sequence pretraining. \\* means training on the larger-scale [AFDB representative](https://www.nature.com/articles/s41586-023-06510-w) structure data, and we find that enlarging structure data is indeed a key factor for better multimodal protein representations. Please refer to DPLM-2 paper for more details about this.\n\nThe training and evaluation pipeline is based on the [SaProt](https://github.com/westlake-repl/SaProt/tree/main) repository, and we slightly modify the code to support DPLM. Users can select the \"representationlearning\" branch for the evaluation of protein predictive tasks.\n\n# Acknowledgements\n\nDPLM extends its gratitude to the following projects and individuals.\n\nWe draw inspiration and leverages/modifies implementations from:\n- [microsoft/evodiff](https://github.com/microsoft/evodiff) for the preprocessed UniRef50 dataset, sequence sampling evaluation implementation and data pipeline.\n- [westlake-repl/SaProt](https://github.com/westlake-repl/SaProt/tree/main) for the representation learning evaluation pipeline.\n- [jingraham/neurips19-graph-protein-design](https://github.com/jingraham/neurips19-graph-protein-design) for the preprocessed CATH dataset.\n- [facebook/esm](https://github.com/facebookresearch/esm/) for their ESM implementations and pretrained model weights.\n- [jasonkyuyim/se3_diffusion](https://github.com/jasonkyuyim/se3_diffusion) for their self-consistency structural evaluation implementation.\n- [jasonkyuyim/multiflow](https://github.com/jasonkyuyim/multiflow) for their evaluation pipeline, structure data processing and preprocessed PDB dataset.\n- [bjing2016/EigenFold](https://github.com/bjing2016/EigenFold) for the CAMEO 2022 dataset.\n\nWe express our sincere appreciation to the authors of these repositories for their invaluable contributions to the development of DPLM family.\n\n# Citation\n\n```\n@inproceedings{wang2024dplm,\n title={Diffusion Language Models Are Versatile Protein Learners},\n author={Wang, Xinyou and Zheng, Zaixiang and Ye, Fei and Xue, Dongyu and Huang, Shujian and Gu, Quanquan},\n booktitle={International Conference on Machine Learning},\n year={2024}\n}\n@inproceedings{wang2025dplm2,\n title={DPLM-2: A Multimodal Diffusion Protein Language Model},\n author={Wang, Xinyou and Zheng, Zaixiang and Ye, Fei and Xue, Dongyu and Huang, Shujian and Gu, Quanquan},\n booktitle={International Conference on Learning Representations},\n year={2025}\n}\n@inproceedings{hsieh2025dplm2_1,\n title={Elucidating the Design Space of Multimodal Protein Language Models},\n author={Hsieh, Cheng-Yen and Wang, Xinyou and Zhang, Daiheng and Xue, Dongyu and Ye, Fei and Huang, Shujian and Zheng, Zaixiang and Gu, Quanquan},\n booktitle={International Conference on Machine Learning},\n year={2025}\n}\n```\n"
},
{
"path": "analysis/TMalign.cpp",
"content": "/* TM-align: sequence-independent structure alignment of monomer proteins by\r\n * TM-score superposition. Please report issues to yangzhanglab@umich.edu\r\n * \r\n * References to cite:\r\n * Y Zhang, J Skolnick. Nucl Acids Res 33, 2302-9 (2005)\r\n *\r\n * DISCLAIMER:\r\n * Permission to use, copy, modify, and distribute the Software for any\r\n * purpose, with or without fee, is hereby granted, provided that the\r\n * notices on the head, the reference information, and this copyright\r\n * notice appear in all copies or substantial portions of the Software.\r\n * It is provided \"as is\" without express or implied warranty.\r\n *\r\n * ==========================\r\n * How to install the program\r\n * ==========================\r\n * The following command compiles the program in your Linux computer:\r\n *\r\n * g++ -static -O3 -ffast-math -lm -o TMalign TMalign.cpp\r\n *\r\n * The '-static' flag should be removed on Mac OS, which does not support\r\n * building static executables.\r\n *\r\n * ======================\r\n * How to use the program\r\n * ======================\r\n * You can run the program without argument to obtain the document.\r\n * Briefly, you can compare two structures by:\r\n *\r\n * ./TMalign structure1.pdb structure2.pdb\r\n *\r\n * ==============\r\n * Update history\r\n * ==============\r\n * 2012/01/24: A C/C++ code of TM-align was constructed by Jianyi Yang\r\n * 2016/05/21: Several updates of this program were made by Jianji Wu:\r\n * (1) fixed several compiling bugs\r\n * (2) made I/O of C/C++ version consistent with the Fortran version\r\n * (3) added outputs including full-atom and ligand structures\r\n * (4) added options of '-i', '-I' and '-m'\r\n * 2016/05/25: Fixed a bug on PDB file reading\r\n * 2018/06/04: Several updates were made by Chengxin Zhang, including\r\n * (1) Fixed bug in reading PDB files with negative residue index,\r\n * (2) Implemented the fTM-align algorithm (by the '-fast' option)\r\n * as described in R Dong, S Pan, Z Peng, Y Zhang, J Yang\r\n * (2018) Nucleic acids research. gky430.\r\n * (3) Included option to perform TM-align against a whole \r\n * folder of PDB files. A full list of options not available\r\n * in the Fortran version can be explored by TMalign -h\r\n * 2018/07/27: Added the -byresi option for TM-score superposition without\r\n * re-alignment as in TMscore and TMscore -c\r\n * 2018/08/07: Added the -dir option\r\n * 2018/08/14: Added the -split option\r\n * 2018/08/16: Added the -infmt1, -infmt2 options.\r\n * 2019/01/07: Added support for PDBx/mmCIF format.\r\n * 2019/02/09: Fixed asymmetric alignment bug.\r\n * 2019/03/17: Added the -cp option for circular permutation\r\n * 2019/07/23: Supported RasMol output by '-o' option\r\n * 2019/07/24: Fixed bug on PyMOL format output by '-o' option with mmCIF input\r\n * 2019/08/18: Fixed bug on RasMol format output file *_atm. Removed excessive\r\n * circular permutation alignment by -cp\r\n * 2019/08/20: Clarified PyMOL syntax.\r\n * 2019/08/22: Added four additional PyMOL scripts.\r\n * 2020/12/12: Fixed bug in double precision coordinate cif file alignment.\r\n * 2021/02/24: Fixed file format issue for new incentive PyMOL.\r\n * 2022/04/12: Compatible with AlphaFold CIF\r\n */\r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include \r\n#include