Use NOS Online
Commercially Available NOS No-Code Web Server
NOS: Diffusion Optimized Sampling for Guided Discrete Diffusion in Protein Sequence Design
NOS (diffusioN Optimized Sampling) is a guidance method for discrete diffusion models that lets researchers steer protein sequence generation toward desired properties. It was introduced by Nate Gruver, Samuel Stanton, Nathan Frey, Tim Rudner, Isidro Hotzel, Julien Lafrance-Vanasse, Arvind Rajpal, Kyunghyun Cho and Andrew Gordon Wilson, with authors from New York University, Prescient Design and Genentech, in an arXiv preprint from 2023. The authors also use NOS inside LaMBO-2, a Bayesian optimization method for multi-objective antibody design.
A common strategy in protein design is to pair a generative model with a discriminative model that scores properties. Structure-based guided diffusion followed by inverse folding is widely pursued, but it runs into data scarcity and the difficulty of inverse design. NOS works directly in sequence space, which avoids the structure-to-sequence step.
Fact | Detail |
|---|---|
Task | Guided sampling from discrete diffusion models for protein sequences |
Guidance mechanism | Gradients in the hidden states of the denoising network |
Variants | NOS-C (Gaussian corruption on embeddings) and NOS-D (mask-based discrete corruption) |
Application | LaMBO-2 for antibody optimization of expression yield and binding affinity |
In vitro outcome | 99% expression rate and 40% binding rate in the final round |
How NOS Works: Guidance in the Denoiser's Hidden States
Hidden-state guidance: NOS follows the gradients of a property predictor with respect to the denoising network's hidden states.
Langevin sampling with a KL penalty: a penalty keeps guided samples likely under the unguided model, which helps preserve natural-looking sequences.
Whole-sequence refinement: unlike PPLM, NOS iteratively refines tokens across the entire sequence at each step.
Two corruption schemes: NOS-C operates on continuous embeddings, while NOS-D uses mask-based discrete corruption.
LaMBO-2 for Antibody Optimization
Multiple objectives: LaMBO-2 generalizes LaMBO to several objectives, using ensemble uncertainty and hypervolume-improvement acquisition.
Edit budget: saliency maps pick which positions to edit under a limited budget of edits.
Developability constraints: constraints on cysteines and glycosylation motifs are enforced.
NOS Results: Guided Design and Antibody Wet-Lab Validation
In silico, NOS gave better trade-offs between likelihood and objective value than PPLM, DiGress and the structure-based baselines tested. In a comparison with genetic algorithms using 32 designs per method, LaMBO-2 remained feasible through 16 sampling steps, while the baselines AdaLead and PEX produced constraint violations or infeasible samples earlier.
In vitro measure | Result |
|---|---|
Campaign size | 20 seed antibodies, 4 targets, 3 rounds, 374 designs in total |
Final round | 56 antibodies tested |
Expression rate | 99% |
Binding rate | 40% |
Expressing designs above seed yield | 43% |
Binders with higher affinity than the seed | 21% |
Final designs had a median of 5 edits and included submicromolar binders for all 4 targets.
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 NOS on Tamarind Bio
Antibody optimization: improve expression yield and binding affinity from seed sequences with a small number of edits.
Property-guided sequence design: steer generation toward objectives such as beta-sheet content or solvent accessibility.
CDR infilling: the authors show that sequence-only diffusion can infill complementarity-determining regions.
Constrained design: keep designs close to a seed while respecting developability rules.
How to Use NOS on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and open NOS.
Provide the seed sequence: enter the protein or antibody sequence to optimize.
Select objectives: choose the properties to guide toward, if exposed.
Set constraints: set the edit budget and any sequence constraints, if exposed.
Tune guidance: adjust guidance strength and sampling parameters, if exposed. The paper describes a KL penalty strength, a step size and a temperature.
Run the job: submit the run and wait for candidate sequences.
Download outputs: retrieve the designed sequences and predicted scores.
Use downstream: filter designs and validate them experimentally, as the authors did in vitro.
Things to Keep in Mind
The authors describe their wet-lab validation as preliminary and exploratory.
The model was not pretrained, and the authors suggest a better pretraining corpus could help.
Length changes rely on sequence alignments, so other domains such as small molecules would need variable-length corruption.
LaMBO-2 with NOS-C yields high predicted objective values but lower naturalness.
The reliability of saliency maps is contested in the literature.
Source: Gruver N, Stanton S, Frey NC, Rudner TGJ, Hotzel I, Lafrance-Vanasse J, Rajpal A, Cho K, Wilson AG. arXiv:2305.20009, 2023 (https://arxiv.org/abs/2305.20009). Code: https://github.com/ngruver/NOS.