Use AF-Traj Online

Commercially Available AF-Traj No-Code Web Server

AF-Traj: Subsampled AlphaFold2 for Protein Conformational Ensembles and Population Shifts

AF-Traj is the Tamarind Bio implementation of the subsampled AlphaFold2 approach described by Monteiro da Silva, Cui, Dalgarno, Lisi and Rubenstein (Brown University and collaborators) in Nature Communications. The method turns AlphaFold2 from a single-structure predictor into a high-throughput generator of protein conformational ensembles, and uses how often each state appears as a proxy for its relative population. The authors' accompanying data repository is released under GPL-3.0.

AlphaFold2 normally returns one ground-state structure per sequence, yet many proteins, including kinases, work by shifting between several conformations. Experimental methods such as NMR can resolve these states but are slow, and enhanced-sampling molecular dynamics (MD) simulations are expensive. The paper asks whether AlphaFold2 can recover conformational distributions without a physics engine.

The answer relies on a simple lever: randomly subsampling the input multiple sequence alignment (MSA) and running many predictions, so that the model reports a spread of structures rather than a single answer.

Item

Detail from the paper

Core idea

Subsample the MSA, run many AlphaFold2 predictions, and use state frequency as a proxy for population

Key parameters

max_seq and extra_seq; 256:512 performed best for Abl1

MSA inputs

JackHMMR searches against UniRef90, Small BFD and MGnify

Run settings

32 seeds, 5 models per seed, 3 recycles, dropout enabled; 480 predictions per test across 3 runs

Validation

Compared with NMR data and enhanced-sampling MD trajectories

Why Predicting Conformational Distributions Matters

Function often depends on which conformations a protein can adopt, and on how a mutation tips the balance between them. Drug-resistance mutations in kinases are a clear example. A fast computational filter that ranks mutations by their effect on conformational equilibrium helps researchers decide which variants deserve NMR or other follow-up work.

How AF-Traj Generates Conformational Ensembles

  1. Build an MSA: sequences are gathered with JackHMMR against UniRef90, Small BFD and MGnify. For Abl1 the MSA exceeded 600,000 sequences.

  2. Subsample the alignment: the max_seq and extra_seq settings control how much of the MSA AlphaFold2 sees, which changes the structures it returns.

  3. Predict many structures: each run used 32 random seeds with dropout during inference (10% in the Evoformer, 25% in the structure module), 5 models per seed and 3 recycles, giving 160 predictions per run.

  4. Analyze the ensemble: structures are clustered and the frequency of each state is compared between wild type and mutants.

Validation Results on Kinases and Cytokines

Abl1, Src and Anc-AS

For the Abl1 kinase core (residues 229 to 515), the ground state was predicted about nine times more often than the I2 state, consistent with NMR. Ground-state predictions were 97% for Src, 89% for Abl1 (235 to 497 construct) and 93% for Anc-AS, matching the experimental ordering.

Mutational series

Across the Abl1 allelic series, the direction and size of the population change were predicted correctly in more than 80% of cases. The method also handled GMCSF, built from an MSA of only 112 sequences, and reproduced the NMR trend that H15 and H83 mutations perturb the backbone more than H87.

System

Result

Abl1 core (229-515)

Ground state about 9x more frequent than I2

Src

97% ground state

Anc-AS

93% ground state

Abl1 (235-497)

89% ground state

Abl1 allelic series

Over 80% correct direction and change

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 AF-Traj on Tamarind Bio

  • Mutation triage: filter large allelic series to choose mutations for NMR or other follow-up studies.

  • Drug resistance: predict how resistance mutations shift conformational equilibria.

  • Alternative states: help identify alternative, possibly metastable, conformations for drug discovery.

  • Design and mechanism: support fold-switching protein design and the study of acquired drug resistance.

How to Use AF-Traj on Tamarind Bio

  1. Open the tool: log in to tamarind.bio and open AF-Traj.

  2. Provide the sequence: enter the target protein sequence the method takes as input.

  3. Set subsampling: choose max_seq and extra_seq values if exposed; the paper found 256:512 best for Abl1 but gives no universal recipe.

  4. Set the sampling budget: if exposed, pick the number of seeds and models; the paper used 32 seeds and 5 models.

  5. Run the job: submit and wait for the ensemble of predicted structures.

  6. Download outputs: retrieve the predicted structures for analysis.

  7. Interpret states: cluster the structures and compare state frequencies between variants.

  8. Plan follow-up: use the results to prioritize experiments or simulations.

Things to Keep in Mind

  • The method is qualitative: it shows whether a state's population rises or falls, and the authors note it is not shown to predict Boltzmann ratios directly.

  • The direction of the M290L effect was predicted wrongly, which led to double mutants being underestimated.

  • Very low-occupancy states may be missed; the authors suggest a detection threshold of roughly 2 to 6%.

  • Shallow MSAs and aggressive subsampling can yield partially unfolded predictions.

  • Settings that work for one protein may not suit another, and fast or small changes such as Fyn-SH3 folding were poorly captured.

Source: Monteiro da Silva G, Cui JY, Dalgarno DC, Lisi GP, Rubenstein BM and colleagues. High-throughput prediction of protein conformational distributions with subsampled AlphaFold2. Nature Communications 15, 2464 (2024). DOI: 10.1038/s41467-024-46715-9. Article: nature.com/articles/s41467-024-46715-9. Data and notebooks: github.com/GMdSilva/gms_natcomms_1705932980_data.

Supporting 10,000+ scientists around the world,

from leading biotechs, and global biopharma