configs/training.yml

# ============================================================================
# DiffDock CGModel Training Configuration
#
# IMPORTANT: Update all filesystem paths below to match your environment.
#   - cache_path, pdbbind_dir, moad_dir: dataset directories
#   - split_train, split_val: text files listing complex names (one per line)
#
# Supported scope (current migration):
#   dataset: pdbbind | moad | generalisation
# Unsupported (will fail fast):
#   dataset: pdbsidechain | distillation
#   all_atoms: true  (AAModel not migrated)
#   triple_training: true
# ============================================================================

runtime:
  run_name: diffdock_cg_example
  log_dir: examples/biosciences/diffdock/outputs/train
  device: auto
  seed: 0
  cudnn_benchmark: false
  wandb: false
  project: diffdock

data:
  dataset: pdbbind  # supported: pdbbind | moad | generalisation
  cache_path: ./cache
  pdbbind_dir: ${ONESCIENCE_DATASETS_DIR}/diffdock/datasets/PDBBind_processed
  moad_dir: ${ONESCIENCE_DATASETS_DIR}/diffdock/datasets/BindingMOAD_2020_processed
  split_train: ${ONESCIENCE_DATASETS_DIR}/diffdock/datasets/splits/timesplit_no_lig_overlap_train
  split_val: ${ONESCIENCE_DATASETS_DIR}/diffdock/datasets/splits/timesplit_no_lig_overlap_val
  protein_file: protein_processed
  limit_complexes: null
  num_conformers: 1
  num_workers: 1
  num_dataloader_workers: 0
  batch_size: 4
  pin_memory: false
  dataloader_drop_last: false
  remove_hs: true
  receptor_radius: 30
  c_alpha_max_neighbors: 10
  atom_radius: 5
  atom_max_neighbors: 8
  chain_cutoff: null
  max_lig_size: null
  matching_popsize: 20
  matching_maxiter: 20
  matching_tries: 1
  not_knn_only_graph: false
  include_miscellaneous_atoms: false
  all_atoms: false  # must stay false in the current migration; only the CGModel main path is supported
  triple_training: false  # unsupported in the current migration; will fail fast if set true
  combined_training: false
  double_val: false
  train_multiplicity: 1
  val_multiplicity: 1
  max_receptor_size: null
  remove_promiscuous_targets: null
  min_ligand_size: 0
  unroll_clusters: false
  enforce_timesplit: false
  crop_beyond: 20
  moad_esm_embeddings_path: null
  pdbbind_esm_embeddings_path: null
  moad_esm_embeddings_sequences_path: null
  esm_embeddings_model: null

diffusion:
  no_torsion: false
  tr_sigma_min: 0.1
  tr_sigma_max: 30.0
  rot_sigma_min: 0.1
  rot_sigma_max: 1.65
  tor_sigma_min: 0.0314
  tor_sigma_max: 3.14
  sampling_alpha: 1.0
  sampling_beta: 1.0
  tr_weight: 0.33
  rot_weight: 0.33
  tor_weight: 0.33
  backbone_loss_weight: 0.0
  sidechain_loss_weight: 0.0

model:
  # old_score_model is intentionally unsupported in the current migration; training only covers the CGModel main path.
  num_conv_layers: 2
  max_radius: 5.0
  scale_by_sigma: true
  norm_by_sigma: false
  ns: 16
  nv: 4
  distance_embed_dim: 32
  cross_distance_embed_dim: 32
  no_batch_norm: true
  use_second_order_repr: false
  cross_max_distance: 80
  dynamic_max_cross: false
  dropout: 0.0
  smooth_edges: false
  odd_parity: false
  embedding_type: sinusoidal
  sigma_embed_dim: 32
  embedding_scale: 1000
  no_aminoacid_identities: false
  sh_lmax: 2
  no_differentiate_convolutions: false
  tp_weights_layers: 2
  num_prot_emb_layers: 0
  reduce_pseudoscalars: false
  embed_also_ligand: true
  depthwise_convolution: false
  use_old_atom_encoder: false

optimization:
  n_epochs: 10
  lr: 0.001
  w_decay: 0.0
  scheduler: null
  scheduler_patience: 20
  lr_start_factor: 0.001
  warmup_dur: 4
  use_ema: true
  ema_rate: 0.999
  restart_dir: null
  restart_ckpt: last_model
  restart_lr: null
  pretrain_dir: null
  pretrain_ckpt: null
  save_model_freq: null

validation:
  test_sigma_intervals: false
  inference_samples: 1
  val_inference_freq: null
  train_inference_freq: null
  inference_steps: 20
  num_inference_complexes: 20
  inference_earlystop_metric: valinf_min_rmsds_lt2
  inference_secondary_metric: null
  inference_earlystop_goal: max

# ----------------------------------------------------------------------------
# NOTE: Confidence model training is a separate step after score-model training.
# Run: python -m onescience.confidence.diffdock.confidence_train
#            --original_model_dir <this_run_dir>
#            --data_dir <pdbbind_dir> --split_train <...> --split_val <...>
# There is no example-level confidence_train script; the module lives at
# src/onescience/confidence/diffdock/confidence_train.py
# ----------------------------------------------------------------------------