Use Profluent E1 Online

Commercially Available Profluent E1 No-Code Web Server

Profluent E1: Retrieval-Augmented Protein Language Model for Fitness Prediction and Contact Maps

Profluent-E1 (E1) is a family of retrieval-augmented protein encoder language models that can condition on homologous sequences at inference time. It was developed by Sarthak Jain, Joel Beazer, Jeffrey A. Ruffolo, Aadyot Bhatnagar and Ali Madani at Profluent Bio, and described in a bioRxiv preprint posted in November 2025. The 150M, 300M and 600M parameter models are released on Hugging Face, with code under Apache 2.0 and weights under a permissive license that carries attribution requirements.

Protein language models trained on single sequences capture general constraints but often miss the family-specific signal that evolutionary context provides. Methods that use multiple sequence alignments supply that signal but require alignment-based pipelines.

E1 addresses this by letting the model read homologs directly as context, without requiring them to be aligned to each other. The authors report that this captures general and family-specific constraints without fine-tuning.

Property

Detail

Architecture

Bidirectional transformer encoder trained with masked language modeling

Model sizes

150M, 300M and 600M parameters

Training scale

4 trillion tokens per model, from the Profluent Protein Atlas and UniRef

Context

Up to 512 sequences per multi-sequence instance

Modes

Single-sequence and retrieval-augmented

How E1 Uses Homologous Sequences for Retrieval-Augmented Protein Modeling

  • Block-causal multi-sequence attention: residues can attend to residues in preceding homologs.

  • Alternating attention: global block-causal attention appears every three layers, and the other layers use intra-sequence attention.

  • Sequence-ID embeddings and RoPE: position IDs restart per sequence in intra-sequence layers.

  • Context curriculum: instance length grew from 8,192 to 32,768 tokens, and sequence count from 2 to 512 during training.

Training Data

Homolog sets were built from the Profluent Protein Atlas and UniRef (version 2411), clustered at 50% and 90% identity with DIAMOND searches. The 600M model took 25 days on 64 H100 GPUs.

E1 Benchmark Results: Zero-Shot Fitness and Unsupervised Contact Prediction

Benchmark

Reported finding

ProteinGym v1.3 (217 substitution assays), single-sequence

E1 outperforms ESM-2 and ESMC at comparable sizes

ProteinGym, retrieval-augmented

Substantial gains over single-sequence mode; reportedly ahead of MSA Pairformer and PoET among public models

CAMEO and CASP15 contacts

E1 beats the ESM family at all scales in single-sequence mode; homologs add consistent gains

Fitness was scored with the masked-marginal method using Spearman correlation and NDCG@10, and contacts were derived with the categorical Jacobian. Retrieval gains were consistent for low- and medium-depth assays, and larger models performed better on average.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

  • A user-friendly graphical interface for setting up and launching experiments

  • A robust API for integration into existing research pipelines

  • An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with Profluent E1 on Tamarind Bio

  • Variant effect scoring: rank mutations by zero-shot fitness using masked-marginal scoring, including in silico site saturation mutagenesis.

  • Protein embeddings: extract per-residue representations as a drop-in replacement for ESM-family embeddings.

  • Contact-map inference: obtain unsupervised contact predictions for a sequence.

  • Family-aware modeling: add homologs of your protein to sharpen predictions for low- and medium-depth families.

How to Use Profluent E1 on Tamarind Bio

  1. Open the tool: log in to tamarind.bio and select Profluent E1.

  2. Enter the query sequence: provide the protein as an amino-acid string.

  3. Optionally add homologs: for retrieval-augmented mode, supply homologous sequences as context, if exposed.

  4. Choose a model size: 150M, 300M or 600M, if exposed.

  5. Run the job: submit it and wait for completion.

  6. Download outputs: retrieve the logits, embeddings or scores returned.

  7. Interpret results: compare mutation scores or inspect embeddings for your protein.

  8. Use downstream: prioritize variants or feed embeddings to predictors.

Things to Keep in Mind

  • It is unclear how much E1 relies on knowledge in its weights versus information in the supplied homologs.

  • Scaling was tested only up to 600M parameters.

  • E1 is sequence-only; no structural information was used in pretraining.

  • Whether prompting with specific sequences can steer design toward particular functions remains open.

Source: Jain, Beazer, Ruffolo, Bhatnagar and Madani, Profluent-E1: Retrieval-Augmented Protein Encoder Models, bioRxiv preprint, 2025, DOI 10.1101/2025.11.12.688125. Code: https://github.com/Profluent-AI/E1.

Supporting 10,000+ scientists around the world,

from leading biotechs, and global biopharma