PoseBench: Protein-Ligand Structure Prediction Benchmarking

The ESM Metagenomic Atlas is a repository containing over 600 million metagenomic protein structures predicted using the ESMFold model (esm.pretrained.esmfold_v0()).

It is divided into two main offerings:

1. High quality MGnify30 structures (~1TB): A subset built from a 30% sequence identity clustering of MGnify90. The best structure from each cluster is selected based on the score pTM * pLDDT. Structures are filtered to have pTM > 0.7 and pLDDT > 0.7.
2. Full database (~15TB): The complete set of PDB structures, stratified by pTM and pLDDT scores to allow for selective downloading based on quality requirements.

RoseTTAFold All-Atom (RFAA) is a biomolecular structure prediction neural network designed to predict a wide range of biomolecular assemblies. This includes:

• Proteins
• Nucleic acids
• Small molecules
• Covalent modifications
• Metals

RFAA provides error estimates alongside predictions, which users can use to assess the reliability of the generated structures.

The SE(3)-Transformer is a Graph Neural Network (GNN) designed for processing 3D points and graphs using a variant of self-attention. It is equivariant under continuous 3D roto-translations (the SE(3) group), meaning that if the input (graphs or point sets) undergoes a rigid transformation (rotation or translation), the model's outputs either remain invariant or transform accordingly.

This implementation is optimized for high throughput and low memory consumption compared to the official implementation, and it supports:

• Multi-GPU training and inference (DDP).
• Mixed precision training (AMP).
• Edge features of any degree (not just scalar degree 0).
• Automatic downloading of the QM9 dataset from DGL.
• Computation of spherical harmonics and Clebsch–Gordan coefficients via the e3nn library.

After running a prediction, results are organized in an output directory (out_dir/).

• lightning_logs/: Logs from training or evaluation.
• predictions/: Contains subdirectories for each input file. Each input subdirectory contains:
  • [input_file]_model_[N].cif: Predicted structure in CIF format (includes per-token pLDDT scores).
  • confidence_[input_file]_model_0.json: Aggregated confidence scores.
  • affinity_[input_file].json: Affinity scores.
  • pae_[input_file]_model_0.npz: Predicted PAE scores for every pair of tokens.
  • pde_[input_file]_model_0.npz: Predicted PDE scores for every pair of tokens.
  • plddt_[input_file]_model_0.npz: Predicted pLDDT scores for every token.
• processed/: Contains the processed input files used during inference.

Boltz-2 provides two distinct types of affinity predictions in its output. They should be used for different stages of molecular design:

1. affinity_probability_binary: A value ranging from 0 to 1 representing the predicted probability that the ligand is a binder. Use this for hit-discovery (detecting binders from decoys).
2. affinity_pred_value: A value reported as log(IC50) (derived from IC50 measured in μM). Use this for ligand optimization (hit-to-lead and lead-optimization) to measure how specific affinity changes with molecular modifications.

DiffDock outputs a confidence score ($c$) for the top pose. This score measures the quality of the predicted 3D structure, not the binding affinity.

General Guidelines for Score $c$:

• $c > 0$: High confidence
• $-1.5 < c < 0$: Moderate confidence
• $c < -1.5$: Low confidence

Important Caveats:

• These intervals assume the complex is similar to the training set (e.g., a drug-like molecule bound to a medium-sized protein in a conformation similar to the bound one).
• If you are working with large ligands, large protein complexes, or unbound protein conformations, you should shift these intervals down (expect lower scores).
• The score is not a direct measure of binding affinity. For affinity prediction, it is recommended to combine DiffDock predictions with tools like GNINA or MM/GBSA, ideally after relaxing the predicted structure with a force field.

The model follows a graph-to-scalar architecture (though it can be adapted to graph-to-graph):

1. Stacked Layers: Consists of equivariant graph self-attention and equivariant normalization layers.
2. Tensor Field Network (TFN): A TFN convolution is applied to obtain invariant features.
3. Pooling: Graph pooling (mean or max over nodes) is applied to these features.
4. MLP: A final Multi-Layer Perceptron (MLP) processes the pooled features to produce scalar predictions.

To obtain a graph-to-graph network, the pooling step can be removed. To output different feature types (e.g., 3D vectors), the final TFN can be modified.