Skip to content

supported_models.md

Latest commit

 

History

History
57 lines (45 loc) · 3.28 KB

supported_models.md

File metadata and controls

57 lines (45 loc) · 3.28 KB

Protenix Supported Models

Protenix provides various pre-trained models to suit different computational resources, inference speeds, and prediction accuracy requirements. This document details the characteristics and configurations of these models.

Naming Convention

Model names follow the format: protenix_{model_size}_{features}_{version}

  • model_size: Size of the model, including base, mini (lightweight), and tiny (minimal).
  • features: Functional characteristics such as default, constraint (distance constraints), esm (includes ESM embeddings), etc. Multiple features are separated by -.
  • version: Version number, e.g., v0.5.0, v1.0.0.

Supported Models Summary

Model Name ESM MSA Constraint RNA MSA Template Params Training Data Cutoff
protenix_base_default_v1.0.0 ❌ ✅ ❌ ✅ ✅ 368.48 M 2021-09-30
protenix_base_20250630_v1.0.0 * ❌ ✅ ❌ ✅ ✅ 368.48 M 2025-06-30
protenix_base_default_v0.5.0 ❌ ✅ ❌ ❌ ❌ 368.09 M 2021-09-30
protenix_base_constraint_v0.5.0 ❌ ✅ ✅ ❌ ❌ 368.30 M 2021-09-30
protenix_mini_esm_v0.5.0 ✅ ✅ ❌ ❌ ❌ 135.22 M 2021-09-30
protenix_mini_ism_v0.5.0 ✅ ✅ ❌ ❌ ❌ 135.22 M 2021-09-30
protenix_mini_default_v0.5.0 ❌ ✅ ❌ ❌ ❌ 134.06 M 2021-09-30
protenix_tiny_default_v0.5.0 ❌ ✅ ❌ ❌ ❌ 109.50 M 2021-09-30

*Note: For practical application scenarios, protenix_base_20250630_v1.0.0 is trained based on the 2025-06-30 wwPDB cutoff and is also released to the community. For fair benchmarks of model improvements across different versions, please use protenix_base_default_v1.0.0 (trained on the 2021-09-30 cutoff).

Model Detailed Descriptions

1. Base Models

  • Characteristics: Full-parameter models with the highest prediction accuracy.
  • Key Configurations:
    • N_cycle: 10 (Number of recycle iterations).
    • sample_diffusion.N_step: 200 (Diffusion steps for higher quality sampling).
  • Use Case: Scientific research requiring maximum precision.

2. Mini & Tiny Models

  • Characteristics: Significant reduction in parameters and faster inference speed.
  • Key Configurations:
    • N_cycle: 4
    • sample_diffusion.N_step: 5
  • Difference: Mini has more layers in Pairformer and Transformer modules compared to Tiny.
  • Use Case: High-throughput screening or scenarios with limited computational resources.

3. Constraint Model (protenix_base_constraint_v0.5.0)

  • Characteristics: Allows incorporating additional experimental constraints during inference (e.g., Pocket, Contact).
  • Included Features:
    • pocket_embedder: Handles binding pocket information.
    • contact_embedder: Handles contact point information.
  • Use Case: Predictions with available structural priors.

4. ESM & ISM Models

  • Characteristics: Integrates the single-sequence protein language model (ESM2-3B), performing better when MSAs are unavailable.
  • Difference: ESM uses standard ESM2 embeddings, while ISM uses specific ISM embeddings.
  • Note: For efficiency, these models do not use MSA by default.