AI Tools
Bio AIBeginner

ESMC / ESMFold2

Protein Biology World Model — A platform for understanding and designing proteins at an evolutionary scale.

ESMC and ESMFold2 are protein biology world models released by Biohub on May 27, 2026. Similar to how GPT learns the grammar and semantics of human language from a massive amount of text, ESMC is a 6 billion parameter foundational language model that has been pre-trained on approximately 2.8 billion protein sequences collected from all life forms on Earth, thereby acquiring the protein design grammar that evolution has accumulated over billions of years. ESMFold2 combines this ESMC embedding with a looped transformer and diffusion-based structure prediction architecture, creating an engine that predicts all-atom 3D structures from a single amino acid sequence and further designs therapeutic protein binders.

The AlphaFold family, which was previously the standard for protein structure prediction, heavily relies on multiple sequence alignment (MSA). While MSA-based approaches are powerful for evolutionarily conserved proteins, they have structural limitations in that the alignment signal is lost in rapidly mutating sequences, such as antibodies, leading to a significant drop in accuracy. ESMC-ESMFold2 adopts the so-called "The Bitter Lesson" philosophy, betting on the scaling law that training a sufficiently large and diverse BERT-based transformer without MSA will outperform specialized architectures. As a result, it achieves higher accuracy than AlphaFold 3 in antibody-antigen binding prediction and also outperforms competing models such as Chai-1 and Boltz-1. In single-sequence mode, it provides approximately a 10-fold increase in inference speed compared to MSA-based methods, making it suitable for large-scale screening.

From the perspective of a biotechnology researcher, the most noteworthy aspect is its ability to design therapeutic binders that have been validated in the laboratory. The Biohub team designed binders with nanomolar affinity for five cancer and immune targets (EGFR, PDGFRβ, PD-L1, CTLA-4, and CD45) with hit rates of 36-88% for compact minibinders and 15-29% for antibody-derived formats, and the PD-L1 binder has even demonstrated the ability to restore T-cell signaling in vivo. This reduces the target binder design cycle from months or years to just a few days. Furthermore, through ESM Atlas, it is possible to explore 6.8 billion sequences and 1.1 billion predicted structures, and to map the function of unannotated proteins using approximately 16,000 interpretable features based on a sparse autoencoder (SAE). The model and code are fully open-source under the MIT license, allowing both academic and industrial researchers to run it directly on their local GPUs or access it through the Biohub platform API.

💻 System Requirements

🧠RAM

NVIDIA GPU with 16GB+ of memory is recommended. Based on FP16, the ESMC-6B weights alone occupy approximately 12GB, and considering the memory required for inference activation, a 24GB GPU (e.g., A100, RTX 4090) is optimal. Running on CPU alone is possible but will be very slow.

💾Storage

The combined size of the ESMC-6B and ESMFold2 models is approximately 15-25GB. The ESM Atlas data is accessible online through the Biohub platform (local download requires several hundred GB or more).

Installation

4-1. Quick Start

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

Currently, only direct installation from GitHub is available (as of 2026-06-17). Official PyPI release is in preparation.

4-2. Basic Usage Example

from transformers import AutoModelForMaskedLM, AutoTokenizer

# Load ESMC-6B (HuggingFace, authentication required)
model = AutoModelForMaskedLM.from_pretrained("biohub/ESMC-6B", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("biohub/ESMC-6B")

# Extract protein sequence embeddings
sequence = "MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQAPILSRVGDGTQDNLSGAEKAVQVKVKALPDAQFEVVHSLAKWKRQQIA"
inputs = tokenizer(sequence, return_tensors="pt").to(model.device)
outputs = model(inputs, output_hidden_states=True)
embeddings = outputs.hidden_states[-1]

4-3. Accessing the Biohub Platform API

# Obtain an API token from the biohub.ai developer console and use it.
from esm import esmc_client

client = esmc_client(
    model="ESMC-6B",
    url="https://biohub.ai/api",
    token="YOUR_API_TOKEN"
)
result = client.predict(sequence="MKTAYIAK...")

🧬 Bio Use Cases

🔬

Design of Antibody Therapeutic Candidate Binders

Utilize ESMFold2's de novo binder design protocol to generate minibinders against immune checkpoint targets such as PD-L1 and CTLA-4. Achieve a hit rate of 36-88% and nanomolar affinity based on compact minibinder criteria. After design, visualize the 3D binding structure with PyMOL or ChimeraX to determine the priority for experimental validation. This significantly reduces the time to reach therapeutic candidates by shortening the design cycle from several months to 3-5 days.

🧬

Exploration of Unannotated Protein Functions and Discovery of Novel Targets

Leverage 16,000 interpretable features based on SAE from ESM Atlas (6.8 billion sequences, 1.1 billion structures) to classify proteins with unknown functions into functional clusters. Extract ESMC-6B embeddings and then perform dimensionality reduction and visualization using UMAP, allowing inference of the role of proteins with no existing annotations based on feature similarity to known proteins. This is used in the early stages of target expansion, such as discovering novel antimicrobial peptide candidates from environmental metagenomic samples.

💊

Large-Scale Variant Structure Screening

Utilize the single-sequence inference mode of ESMC-6B to predict the structures of tens of thousands of variant libraries in batch without MSA. With an inference speed approximately 10 times faster than AlphaFold 3, it can process tens of thousands of cases per day on a GPU server. Filter the top candidates based on the predicted confidence score and transfer them to precision docking simulations (HADDOCK, ClusPro, etc.) to maximize the efficiency of hit compound selection.

FAQ

What is ESMC / ESMFold2?

ESMC and ESMFold2 are protein biology world models released by Biohub on May 27, 2026. Similar to how GPT learns the grammar and semantics of human language from a massive amount of text, ESMC is a 6 billion parameter foundational language model that has been pre-trained on approximately 2.8 billion protein sequences collected from all life forms on Earth, thereby acquiring the protein design grammar that evolution has accumulated over billions of years. ESMFold2 combines this ESMC embedding with a looped transformer and diffusion-based structure prediction architecture, creating an engine that predicts all-atom 3D structures from a single amino acid sequence and further designs therapeutic protein binders. The AlphaFold family, which was previously the standard for protein structure prediction, heavily relies on multiple sequence alignment (MSA). While MSA-based approaches are powerful for evolutionarily conserved proteins, they have structural limitations in that the alignment signal is lost in rapidly mutating sequences, such as antibodies, leading to a significant drop in accuracy. ESMC-ESMFold2 adopts the so-called "The Bitter Lesson" philosophy, betting on the scaling law that training a sufficiently large and diverse BERT-based transformer without MSA will outperform specialized architectures. As a result, it achieves higher accuracy than AlphaFold 3 in antibody-antigen binding prediction and also outperforms competing models such as Chai-1 and Boltz-1. In single-sequence mode, it provides approximately a 10-fold increase in inference speed compared to MSA-based methods, making it suitable for large-scale screening. From the perspective of a biotechnology researcher, the most noteworthy aspect is its ability to design therapeutic binders that have been validated in the laboratory. The Biohub team designed binders with nanomolar affinity for five cancer and immune targets (EGFR, PDGFRβ, PD-L1, CTLA-4, and CD45) with hit rates of 36-88% for compact minibinders and 15-29% for antibody-derived formats, and the PD-L1 binder has even demonstrated the ability to restore T-cell signaling in vivo. This reduces the target binder design cycle from months or years to just a few days. Furthermore, through ESM Atlas, it is possible to explore 6.8 billion sequences and 1.1 billion predicted structures, and to map the function of unannotated proteins using approximately 16,000 interpretable features based on a sparse autoencoder (SAE). The model and code are fully open-source under the MIT license, allowing both academic and industrial researchers to run it directly on their local GPUs or access it through the Biohub platform API.

When should I use ESMC / ESMFold2?

Protein Biology World Model — A platform for understanding and designing proteins at an evolutionary scale.

What is a biomedical use case for ESMC / ESMFold2?

Design of Antibody Therapeutic Candidate Binders: Utilize ESMFold2's de novo binder design protocol to generate minibinders against immune checkpoint targets such as PD-L1 and CTLA-4. Achieve a hit rate of 36-88% and nanomolar affinity based on compact minibinder criteria. After design, visualize the 3D binding structure with PyMOL or ChimeraX to determine the priority for experimental validation. This significantly reduces the time to reach therapeutic candidates by shortening the design cycle from several months to 3-5 days.

📄 Official Docs🐙 GitHub

📝 Update Notes

  1. vv3.4.19/15/2026

    이번 v3.4.1 업데이트에서는 FoldCP 지원이 추가되어 단백질 구조 예측의 정밀도를 더욱 높일 수 있게 되었습니다. 특히 Hugging Face 스타일의 체크포인트를 불러올 수 있게 되어, 기존 HF 생태계의 다양한 모델을 연구 워크플로우에 훨씬 손쉽게 통합하여 활용할 수 있습니다. 여러 버그 수정도 함께 이루어져 더욱 안정적인 단백질 구조 분석 환경을 제공합니다.

  2. vv3.2.38/25/2026

    이번 v3.2.3 업데이트에서는 분자 복합체(MolecularComplex)에 체인 ID가 추가되어, 여러 체인이 얽힌 복잡한 단백질 구조를 더욱 명확하게 식별할 수 있어요. mmCIF 변환 시 biotite 라이브러리를 활용하도록 개선되어 구조 데이터 처리의 안정성도 한층 높아졌습니다. 또한, MSA(다중 서열 정렬) 과정의 시퀀스 필터링 로직이 정교해져 예외적인 서열 데이터도 더욱 정확하게 처리할 수 있게 되었어요. 이번 업데이트를 통해 단백질 복합체 분석과 서열 기반 모델링의 신뢰도를 더욱 높일 수 있습니다.

  3. vv3.4.08/24/2026

    이번 업데이트에서는 최신 단백질 언어 모델인 ESMC와 ESMFold2를 완벽하게 지원하여 더욱 정교한 단백질 구조 예측과 분석이 가능해졌어요. 특히 Hugging Face 포트와의 호환성 레이어가 추가되어, 기존에 활용하던 다양한 사전 학습 모델들을 훨씬 간편하게 연구 워크플로우에 통합할 수 있습니다. 또한 기존 모델과의 하위 호환성을 유지하고 있어, 기존에 구축해둔 연구 파이프라인을 중단 없이 안정적으로 이어갈 수 있다는 점이 큰 장점이에요.

  4. vv3.2.2.post26/17/2026

    이번 업데이트에서는 단백질뿐만 아니라 DNA, RNA, 리간드를 포함한 복합체 구조를 예측할 수 있는 'All-Atom' 기능이 새롭게 도입되었습니다. 이를 통해 단백질과 핵산, 혹은 화합물 간의 상호작용을 더욱 정교하게 분석할 수 있어 신약 개발 및 구조 생물학 연구의 범위를 크게 넓힐 수 있습니다. 또한, MSA(다중 서열 정렬) 처리 효율을 높이는 기능과 SDK 안정성 개선이 함께 이루어져 더욱 빠르고 신뢰도 높은 구조 예측 파이프라인을 구축할 수 있습니다.

🧪 Related Code of Life

No related Code of Life posts yet.