AFsample3 Is Now on Vecura: Diverse Conformational Sampling Without the Infrastructure

What is AFsample3?

AFsample3 is an enhanced sampling framework built upon the AlphaFold3 (AF3) inference pipeline that generates diverse protein conformational ensembles by randomly masking columns in the input Multiple Sequence Alignment (MSA) at run time. This MSA masking strategy disrupts co-evolutionary signals, making the system more susceptible to exploring alternative structural states. It requires no additional network training on top of the AF3 architecture and adds virtually no computational overhead beyond standard AF3 inference.

It helps users generate and select multiple distinct protein conformations, substantially improving alternate-state prediction accuracy and ensemble diversity compared to standard AlphaFold3. It is especially useful for studying proteins with multiple experimentally known conformational states — such as those that undergo large structural rearrangements depending on binding partners, cellular context, or signaling state — where capturing conformational heterogeneity is critical for understanding protein mechanism and function.

What can users do with AFsample3 on Vecura?

With REINVENT 4 on Vecura, users can:

• Submit a protein sequence and MSA to run multiple inference iterations with randomized MSA masking, generating a broad pool of diverse structural models.

• Recover high-quality models for both preferred and alternate conformational states in a single run, without separate pipelines.

• Use the DiSco protocol to rank and cluster models by structural distance, identifying distinct states — including intermediates — without needing a reference structure.

• Evaluate ensemble diversity using fill-ratio metrics and diversity plots to quantify how broadly the generated models span the conformational landscape.

What the output means

The output is a ranked ensemble of 3D protein structural models (coordinates), accompanied by TM-scores measuring similarity to known reference states, confidence scores, and diversity metrics (fill-ratio). When using DiSco state selection, outputs include cluster representatives ranked by structural diversity, identifying the most distinct conformations within the ensemble — potentially including intermediate or novel states not previously characterized experimentally.

Why this matters

Proteins are not rigid objects — they shift between multiple conformations to carry out their functions, respond to ligands, or switch between active and inactive states. Standard structure prediction tools like baseline AlphaFold tend to collapse toward a single dominant conformation, leaving the full conformational landscape unexplored. This is a significant blind spot for drug discovery, since many therapeutic targets (e.g., GPCRs, kinases, ion channels) must be modeled in multiple states to identify cryptic binding sites or understand allosteric mechanisms. AFsample3 addresses this by improving alternate-state accuracy for 28% of benchmarked targets while degrading performance for only 3%, and increasing the number of high-quality alternate-state predictions (TM-score > 0.8) by over 50% compared to standard AlphaFold3. Ensemble diversity is also markedly enhanced, enabling modeling of potential intermediate and additional states beyond just two end points. This makes it one of the most capable openly available tools for exploring protein conformational landscapes computationally.

• Developed by: Wallner Lab, Linköping University (Sweden)

• Source: Official GitHub Repository

• Reference: AFsample3: Generating and selecting multiple conformational states with AlphaFold3, bioRxiv (2026). DOI: 10.64898/2026.01.16.699904