Compare · 10 models
Protein language models you can run online
Protein language models extract biological representations, score mutations, and design protein sequences or structures from primary amino acid chains and structural tokens.
What protein language models models do
Protein language models learn the statistical and functional patterns of proteins from large corpora of sequences and structural tokens. Most models in this category employ masked language modelling objectives, where hidden residues are predicted from surrounding sequence context to produce high-dimensional vector representations. Others employ autoregressive sequence generation or diffusion processes to generate new sequences and coordinates directly.
These numerical representations serve as inputs for downstream supervised prediction tasks, such as variant-effect scoring and functional annotation, or as generative engines for computational protein design. While standard architectures operate strictly on linear amino acid strings, specialised variants incorporate discrete structural alphabets or joint coordinate diffusion to bridge sequence and three-dimensional tertiary structure.
Compared
- ESM2-3BMeta FAIRA 3-billion-parameter protein language model generating 1,280-dimensional sequence embeddings.Commercial OKReady
- ESM2-650MMeta FAIRGenerates 1,280-dimensional embeddings and masked residue predictions from protein sequences.Commercial OKReady
- ProGen2Salesforce ResearchGenerates protein and antibody sequences conditionally using autoregressive language modelling.Commercial OKReady
- ESMC 300MbiohubA 300M-parameter protein language model for sequence embedding, masked language modelling, and classification.Check termsOn demand
- ESMC 600MbiohubA 600M-parameter protein language model providing sequence embeddings and masked language modelling.Check termsOn demand
- ESMC 6BbiohubA 6-billion-parameter protein language model producing sequence representations and masked residue predictions.Check termsOn demand
- SaProtSu et al., Westlake UniversityA structure-aware protein language model predicting variant effects and generating embeddings using 3Di tokens.Commercial OKOn demand
- DPLM-2Nanjing University / ByteDanceGenerates protein sequences and structures simultaneously using a unified diffusion process.Commercial OKNot yet
- ESMC 300M (Dec 2024)biohubA 300-million-parameter protein language model for generating protein embeddings in biological workflows.See its pageNot yet
- ESMC 600M (Dec 2024)biohubA 600-million-parameter protein language model checkpoint for protein engineering and variant-effect analysis.See its pageNot yet
How to choose
For embedding extraction and standard sequence evaluation, the ESM2 and ESMC families provide scalable masked language models. Between Meta FAIR checkpoints, ESM2-650M requires only 8 GB of GPU memory, making it practical for high-throughput batch inference where computational resources are constrained. ESM2-3B demands 16 GB of GPU memory and delivers higher general accuracy, producing 1,280-dimensional representations that also serve as inputs for ESMFold.
The ESMC series from Biohub offers architectures configured with SwiGLU activations, rotary positional embeddings, and Pre-LN normalisation without linear biases, all subject to a 2,048-token context window. ESMC 300M fits within 16 GB of GPU memory on consumer hardware. For larger capacities, ESMC 600M and the 80-layer ESMC 6B support bfloat16 FlashAttention-2 and parameter-efficient fine-tuning via LoRA, though extracting attention maps requires falling back to slower unoptimised attention routines. The ESMC 300M (Dec 2024) and ESMC 600M (Dec 2024) checkpoints exist specifically to support legacy pipelines that rely on the original esm package format.
When experimental or predicted 3D coordinates are available, SaProt provides structure-aware representations by combining amino acid residues with Foldseek 3Di tokens. SaProt outperforms ESM2 across most functional prediction benchmarks when structural inputs are provided, and it accounts for structural uncertainty by masking AlphaFold 2 coordinates where pLDDT scores fall below 70. However, SaProt requires 16 GB of GPU memory and cannot generate frozen embeddings from sequence alone without task-specific fine-tuning.
For generative design, ProGen2 and DPLM-2 address different requirements and both demand 24 GB of GPU memory. DPLM-2 applies a diffusion process to generate sequence and tertiary structure concurrently, outputting full all-atom PDB coordinates and co-designed sequences constrained by specified lengths and functional motifs.
Questions
Can I run these models without a GPU or installing anything?
7 of the 10 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
Which models can generate both amino acid sequences and 3D atomic structures?
DPLM-2 unifies sequence and structure design within a single diffusion process, outputting both the amino acid sequence and complete all-atom PDB coordinates. ProGen2, the ESM2 models, and the ESMC models generate or embed sequences only, while SaProt requires structural inputs to produce structure-aware sequence tokens.
Can SaProt be used if only primary amino acid sequences are available?
SaProt requires paired sequence and Foldseek 3Di structural tokens to generate its default frozen embeddings. If only primary sequences are supplied without structural coordinates, the model requires task-specific fine-tuning to function.
What hardware is required to deploy ESM2-650M compared to ESM2-3B?
ESM2-650M requires 8 GB of GPU memory, allowing large-scale batch processing on standard hardware. ESM2-3B requires 16 GB of GPU memory, which increases computing requirements and inference latency during sequence screening.
What is the primary difference between the standard ESMC checkpoints and the December 2024 releases?
The ESMC 300M (Dec 2024) and ESMC 600M (Dec 2024) checkpoints are maintained for backwards compatibility with the original GitHub esm package format, whereas the standard ESMC checkpoints are intended for current Hugging Face workflows and support modern acceleration like FlashAttention-2.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.