← AI Tools
Bio AIBeginner

ESMFold2

Released in May 2026 by the Chan Zuckerberg Biohub, ESMFold2 is a state-of-the-art bioinformatics model that accurately predicts the 3D spatial structure of a protein using only single amino acid sequence information, without the need to generate a large number of multiple sequence alignments. Previous generations of protein folding models required tens of minutes to search and align homologous sequences from gigabyte-sized databases in order to calculate the evolutionary relationships and distance constraints between amino acids.

ESMFold2, released by Chan Zuckerberg Biohub in May 2026, is a state-of-the-art bioinformatics model that accurately predicts the 3D conformation of a protein using only single amino acid sequence information, without the need to generate large-scale multiple sequence alignments. Previous generations of protein folding models required significant time, often tens of minutes, to query and align homologous sequences from massive, gigabyte-sized databases to calculate the evolutionary relationships and distance constraints between amino acids. In contrast, this model bypasses the complex pre-processing query process entirely and restores accurate 3D structures within seconds by leveraging ESMC, a 6B parameter protein language model that deeply learns the evolutionary context and structural patterns inherent in protein amino acid sequence information. This is similar to how a large language model like GPT learns from billions of sentence data to generate the most contextually natural flow of words; ESMFold2 recognizes the flow of residue arrangements within a protein domain and infers the structural plausibility of a protein's 3D morphology that can exist in nature in real-time.

Previously widely used methods, such as AlphaFold, required significant time and the use of dozens of high-performance computer nodes and vast database disks to analyze a single protein structure, which posed a serious limitation in terms of speed for real-time protein design research or large-scale genomic analysis involving the screening of tens of thousands of variants. ESMFold2 dramatically reduces processing speed by not relying on these high-cost sequence alignment pipelines and instead using single-sequence folding, ensuring overwhelming efficiency by allowing the conversion of tens of sequences into structural space per second. Furthermore, by incorporating a sophisticated diffusion model decoder as the structure generation component, it overcomes the limitations of conventional single-sequence models in predicting multi-domain structures and maintains a high level of prediction accuracy even for complex structures such as protein complexes or viral capsids.

Thanks to these computational innovations, modern biotechnology researchers can now design and experimentally validate novel bio-binders or single-chain variable region (scFv) fragments optimized for disease target proteins, such as EGFR, a target for cancer therapy, or PD-L1, a key component of immune checkpoint inhibitors, in just a few days. In addition, it can rapidly decode the folding patterns of previously unknown proteins from environmental sources belonging to unknown evolutionary lineages, which was virtually impossible with conventional methods, accelerating the development of novel enzymes and the discovery of biological resources for climate change mitigation by a factor of hundreds. In short, ESMFold2 is more than just a tool for generating 3D coordinates of proteins; it is becoming a powerful microscope that rapidly projects the entire genomic sequencing data into the world of 3D structures and a key indicator for molecular engineering.

💻 System Requirements

🧠RAM

Minimum 12GB (not required for API inference), NVIDIA GPU 24GB+ recommended for local execution (A100, RTX 3090/4090 or higher)

💾Storage

Download of model weight files and inclusion of dependency packages; at least 20GB or more of free space is recommended

⚡ Installation

4-1. Quick Start

pip install esm@git+https://github.com/Biohub/esm.git@main

4-2. Detailed installation

# After installing PyTorch and the necessary build tools, proceed with the installation of the ESM SDK.
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install esm@git+https://github.com/Biohub/esm.git@main

🧬 Bio Use Cases

🔬

De Novo Minibinder Design

Rapidly model the 3D interaction structure of ultra-small binding proteins that specifically bind to a target protein (e.g., EGFR, PD-L1) and perform binding affinity optimization simulations.

🧬

Large-Scale Genomic Variation Structural Impact Analysis

Simulate at high speed the impact of single nucleotide polymorphisms (SNPs) or amino acid variations on the 3D conformation of a protein and its ligand-binding pocket, contributing to the interpretation of clinical variation information.

💊

Ultra-Fast Protein Complex Screening

In protein-protein interaction (PPI) prediction, rapidly screen a large library of candidate partners in seconds without MSA searching, enabling large-scale structural interaction mapping.

FAQ

What is ESMFold2?

ESMFold2, released by Chan Zuckerberg Biohub in May 2026, is a state-of-the-art bioinformatics model that accurately predicts the 3D conformation of a protein using only single amino acid sequence information, without the need to generate large-scale multiple sequence alignments. Previous generations of protein folding models required significant time, often tens of minutes, to query and align homologous sequences from massive, gigabyte-sized databases to calculate the evolutionary relationships and distance constraints between amino acids. In contrast, this model bypasses the complex pre-processing query process entirely and restores accurate 3D structures within seconds by leveraging ESMC, a 6B parameter protein language model that deeply learns the evolutionary context and structural patterns inherent in protein amino acid sequence information. This is similar to how a large language model like GPT learns from billions of sentence data to generate the most contextually natural flow of words; ESMFold2 recognizes the flow of residue arrangements within a protein domain and infers the structural plausibility of a protein's 3D morphology that can exist in nature in real-time. Previously widely used methods, such as AlphaFold, required significant time and the use of dozens of high-performance computer nodes and vast database disks to analyze a single protein structure, which posed a serious limitation in terms of speed for real-time protein design research or large-scale genomic analysis involving the screening of tens of thousands of variants. ESMFold2 dramatically reduces processing speed by not relying on these high-cost sequence alignment pipelines and instead using single-sequence folding, ensuring overwhelming efficiency by allowing the conversion of tens of sequences into structural space per second. Furthermore, by incorporating a sophisticated diffusion model decoder as the structure generation component, it overcomes the limitations of conventional single-sequence models in predicting multi-domain structures and maintains a high level of prediction accuracy even for complex structures such as protein complexes or viral capsids. Thanks to these computational innovations, modern biotechnology researchers can now design and experimentally validate novel bio-binders or single-chain variable region (scFv) fragments optimized for disease target proteins, such as EGFR, a target for cancer therapy, or PD-L1, a key component of immune checkpoint inhibitors, in just a few days. In addition, it can rapidly decode the folding patterns of previously unknown proteins from environmental sources belonging to unknown evolutionary lineages, which was virtually impossible with conventional methods, accelerating the development of novel enzymes and the discovery of biological resources for climate change mitigation by a factor of hundreds. In short, ESMFold2 is more than just a tool for generating 3D coordinates of proteins; it is becoming a powerful microscope that rapidly projects the entire genomic sequencing data into the world of 3D structures and a key indicator for molecular engineering.

When should I use ESMFold2?

Released in May 2026 by the Chan Zuckerberg Biohub, ESMFold2 is a state-of-the-art bioinformatics model that accurately predicts the 3D spatial structure of a protein using only single amino acid sequence information, without the need to generate a large number of multiple sequence alignments. Previous generations of protein folding models required tens of minutes to search and align homologous sequences from gigabyte-sized databases in order to calculate the evolutionary relationships and distance constraints between amino acids.

What is a biomedical use case for ESMFold2?

De Novo Minibinder Design: Rapidly model the 3D interaction structure of ultra-small binding proteins that specifically bind to a target protein (e.g., EGFR, PD-L1) and perform binding affinity optimization simulations.

📄 Official Docs🐙 GitHub

📝 Update Notes

  1. vv3.5.011/21/2026

    ESMFold2 v3.5.0 업데이트를 통해 단백질 구조 예측 속도가 비약적으로 향상되었습니다. 새로운 'fast' 모드는 기존 대비 약 7배, 'fused' 모드는 2배 이상 빨라져 대규모 단백질 스크리닝 작업을 훨씬 효율적으로 수행할 수 있습니다. 또한 'big' 모드가 도입되어 4개의 H100 GPU 사용 시 최대 12,000개 아미노산 길이의 긴 서열도 예측할 수 있게 되었습니다. 이를 통해 기존에 메모리 한계로 다루기 어려웠던 거대 단백질이나 복합 구조 연구의 가능성이 크게 열렸습니다.

  2. vv3.4.311/9/2026

    ESMFold2의 v3.4.3 업데이트를 통해 더욱 안정적인 단백질 구조 예측 환경을 경험하실 수 있어요. 기존의 버그들이 수정되어 연구 과정에서 발생하던 예기치 않은 오류를 줄였으며, huggingface transformers 의존성 버전도 최신화되었습니다. 이를 통해 최신 AI 라이브러리와의 호환성이 높아져 더욱 원활한 연구 워크플로우를 구축할 수 있습니다.

  3. vv3.4.211/2/2026

    이번 업데이트에서는 대량의 단백질 구조를 한 번에 처리할 수 있는 Batch API와 최신 데이터를 반영한 2025-cutoff 모델이 새롭게 도입되었습니다. fp8 추론 기술 적용으로 예측 속도가 빨라졌을 뿐만 아니라, 설치 과정도 훨씬 간편해져 대규모 단백질 스크리닝 작업의 효율성이 크게 높아졌습니다. 또한 구조 예측의 정확도를 극대화한 Pro 기능과 안정적인 라이브러리 환경까지 갖추어, 더욱 빠르고 정밀한 단백질 구조 분석이 가능해졌습니다.

  4. vv3.4.1.post19/20/2026

    ESMFold2의 이번 업데이트는 다양한 구현 방식(HF, esm, mlx_lm) 간의 호환성을 개선하여, 하나의 모델 가중치만으로도 여러 환경에서 편리하게 사용할 수 있게 해줍니다. 특히 실험적인 기능의 안정성을 높이고 SAE(Sparse Autoencoder) 구현 오류를 해결하여, 더욱 정밀하고 신뢰할 수 있는 단백질 구조 분석이 가능해졌습니다. 연구자분들은 이제 하드웨어 환경에 구애받지 않고 더욱 효율적이고 정확한 연구 워크플로우를 구축할 수 있습니다.

  5. vv3.4.19/15/2026

    ESMFold2 v3.4.1 업데이트에서는 FoldCP 지원이 추가되어 더욱 폭넓은 단백질 구조 분석이 가능해졌습니다. 특히 Hugging Face 스타일의 체크포인트를 불러올 수 있게 되어, 기존 HF 생태계의 모델들을 연구 워크플로우에 훨씬 쉽고 빠르게 통합하여 활용할 수 있습니다. 다양한 버그 수정으로 안정성도 높아졌으니, 더욱 정밀한 단백질 구조 예측 연구를 위해 이번 업데이트를 적용해 보시길 추천드려요.

  6. vv3.4.08/24/2026

    이번 업데이트를 통해 최신 단백질 구조 예측 모델인 ESMC와 ESMFold2를 본격적으로 지원하여 더욱 정교한 분석이 가능해졌습니다. 특히 Hugging Face 모델과의 호환성 레이어가 추가되어, 기존에 사용하던 모델들을 더욱 편리하게 연동하여 연구에 활용할 수 있습니다. 기존 레거시 모델 및 가중치에 대한 하위 호환성도 유지되므로, 기존 연구 워크플로우를 변경하지 않고도 최신 기술을 안정적으로 도입해 보세요.

  7. vv3.2.2.post27/5/2026

    이번 업데이트에서는 단백질뿐만 아니라 DNA, RNA, 리간드를 모두 포함한 'All-Atom' 분자 복합체 구조 예측 기능이 새롭게 추가되었습니다. 이를 통해 단백질과 핵산, 혹은 화합물 간의 상호작용을 더욱 정밀하게 모델링할 수 있어 신약 개발 및 구조 생물학 연구의 범위를 크게 넓힐 수 있습니다. 또한 MSA 처리 효율을 높이고 구조 예측의 안정성을 개선하는 다양한 버그 수정이 이루어져, 더욱 빠르고 신뢰도 높은 연구 파이프라인 구축이 가능해졌습니다.

🧪 Related Code of Life

No related Code of Life posts yet.

BioPlayground

Reading, linking, and lawful quotation stay open; high-speed bulk collection and unauthorized redistribution do not.

Unless stated otherwise, content rights belong to BioPlayground or the relevant rights holder.