Use ProFam Online
Commercially Available ProFam No-Code Web Server
ProFam: Open-Source Protein Family Language Model for Fitness Prediction and Design
ProFam is an open-source suite of data, weights and code for protein family language modelling. Its model, ProFam-1, is a 251-million-parameter autoregressive Transformer that supports zero-shot fitness prediction and homology-guided sequence generation. It was introduced in a December 2025 bioRxiv preprint by Jude Wells, Alex Hawkins Hooker, Micha Livne, Weining Lin, David Miller, Christian Dallago, Nicola Bordin, Brooks Paige, Burkhard Rost, Christine Orengo and Michael Heinzinger (University College London, NVIDIA, Duke University, TUM and Helmholtz Munich). Code, weights and training data are all publicly released.
Protein language models trained on single sequences learn general patterns, but masked models are ill-suited to scoring insertions, deletions and distant variants. Autoregressive models handle these naturally, yet unconditional generation samples from all of sequence space rather than a family of interest.
Family language models fix this by conditioning on homologous sequences supplied in the prompt, steering scoring and generation with evolutionary context and no fine-tuning.
Attribute | Detail |
|---|---|
Model | ProFam-1, Llama-style Transformer, 16 layers, 251M parameters |
Training data | ProFam Atlas: almost 40 million families, about 481M sequences |
Context | Up to 8192 tokens of concatenated, unaligned family sequences |
Tasks | Variant scoring and family-conditioned generation |
ProteinGym Spearman | 0.47 substitutions, 0.53 indels |
How ProFam-1 Scores Variants and Generates Sequences from Protein Families
Family prompt: homologous sequences are concatenated with separator tokens; sequence order is randomized in training so the model is order-invariant.
Scoring: a variant is scored by its average per-token log-probability under the family prompt, including the end-of-sequence token.
Generation: new sequences are sampled conditioned on the prompt, without fixed length or alignment to a seed.
Prompt ensembling: large families are split into multiple prompts, and scores or next-token probabilities are averaged across them.
The ProFam Atlas
Training families come from the AlphaFold Database, TED domain families and the OpenFold OpenProteinSet, plus single UniRef90 sequences.
Results: Benchmarks for Fitness Prediction, Family Statistics and Synthetic MSAs
Evaluation | Result |
|---|---|
ProteinGym, single prompt (average) | Spearman 0.42 |
ProteinGym, 120-prompt ensemble | Spearman 0.47 (substitutions) |
Sequences in acceptable length range, multi-sequence prompt | 95.1% vs 83.8% for PoET |
Covariance correlation, single-sequence prompt | PCC 0.64 vs 0.49 for PoET |
ColabFold on hard CASP15/16 targets, lDDT | 0.7 with synthetic MSAs vs 0.5 single-sequence, 0.9 natural MSAs |
Fitness performance peaked at mid-range average log-likelihood (about -1.3), and ensembling many prompts made prompt engineering largely unnecessary. Generated sequences showed structural similarity to natural family members comparable to PoET and outperformed a random-mutation baseline.
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 ProFam on Tamarind Bio
Zero-shot variant ranking: score substitutions and indels against a family context to prioritize variants.
Family-guided design: generate sequences that follow the evolutionary patterns of a chosen family.
Synthetic MSAs: augment sparse evolutionary signal, which improved ColabFold predictions for hard targets.
Open foundation: use the released weights and data for further method development.
How to Use ProFam on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and select ProFam.
Provide homologs: supply related sequences as the family prompt, as the interface defines.
Choose the task: scoring of candidate variants, or generation of new sequences, if exposed.
Provide candidates or settings: for scoring, the variants to evaluate; for generation, options such as number of sequences, if exposed.
Use ensembling: if exposed, multiple prompts improve scoring and generation diversity.
Run and download: launch the job and retrieve scores or sequences.
Interpret: rank variants by log-likelihood and compare designs against natural family members.
Continue downstream: fold top sequences and select candidates for experiments.
Things to Keep in Mind
The authors have no experimental validation of generated sequences yet.
Matching PoET zero-shot scoring needed 120 forward passes versus 15 for PoET, so ensembles cost more compute.
Pushing average log-likelihood too high can reduce fitness correlation.
Synthetic MSAs improved lDDT but did not match natural MSAs.
Source: Wells J., Hawkins Hooker A., Livne M., et al., "ProFam: Open-Source Protein Family Language Modelling for Fitness Prediction and Design", bioRxiv preprint (2025), DOI 10.64898/2025.12.19.695431. Code: github.com/alex-hh/profam; weights: huggingface.co/judewells/ProFam-1.