TuningExample_config.yaml

# DATA DEFINITIONS ------------------------------------------------------------
# -----------------------------------------------------------------------------
DATA_TYPE: # TODO This variable specifies the details for each data type.
# if you don't use a data type, just comment it out, or delete it
  RNA:
    FILE_RAW: "data_mrna_seq_v2_rsem_formatted.parquet" # TODO your file name here. Standard is parquet format. DELIM-based files can be provided and should have file extension: "csv","tsv" or "txt".
    TYPE: "NUMERIC" # TODO (your data type here) MIXED, IMG, NUMERIC and ANNOTATION are possible
    SCALING: "Standard" # TODO (your scaling method here) NoScaler, MinMax, Standard, Robust, MaxAbs, are possilbe. Be sure to make this consistent with your loss function i.e. BCE and MinMax
    FILTERING: "Var" # SHOULDDO NoFilt, NonZeroVar, Var, MAD, Corr, Var+Corr are possible, Var is recommended because of faster computation
  METH:
    FILE_RAW: "data_methylation_per_gene_formatted.parquet"
    TYPE: "NUMERIC"
    SCALING: "Standard"
    FILTERING: "Var"
  ANNO:
    FILE_RAW: "data_clinical_formatted.parquet"
    TYPE: "ANNOTATION" ## ANNOTATION files will be ignored as data modality in AE
  # CLIN:
  #   FILE_RAW: "data_clinical_formatted.parquet" # TODO your file name here
  #   TYPE: "MIXED" # TODO (your data type here) MIXED, IMAGE, NUMERIC are possible, for clinical usually MIXED
  #   REL_CLIN_C: # TODO colnames of categorial features
  #     - "SUBTYPE"
  #     - "ONCOTREE_CODE"
  #     - "SEX"
  #     - "AJCC_PATHOLOGIC_TUMOR_STAGE"
  #     - "ETHNICITY"
  #     - "PATH_M_STAGE"
  #     - "PATH_N_STAGE"
  #     - "PATH_T_STAGE"
  #     - "GRADE"
  #   REL_CLIN_N: # TODO colnames of numerical features
  #     - "AGE"
  #     - "ANEUPLOIDY_SCORE"
  #     - "MSI_SCORE_MANTIS"
  #     - "TMB_NONSYNONYMOUS"
  #   SCALING: "Standard" # TODO MinMax, Standard, Robust, MaxAbs, are possilbe. Be sure to make this consistent with your loss function i.e. BCE and MinMax
  #   FILTERING: "NoFilt" # SHOULDDO NoFilt, Var, Corr, Var+Corr are possible, Var is recommended because of faster computation

K_FILTER: 1000 # SHOULDDO Number of features after filtering per data modality

## Model and Training --------------------------------------------------------
# ----------------------------------------------------------------------------
TRAIN_TYPE: "tune" # SHOULDDO train or tune are possible
MODEL_TYPE: "varix" #SHOULDO which autoencoder to use: vanillix, varix, stackix, ontix or x-modalix
BETA: 0.1 # OPTIONAL float that adds a weight to the KL divergence or MMD loss of variational autoencoder

LATENT_DIM_FIXED: 2 # SHOULDDO number of neurons in latent dimenion (all integer values are possible)
BATCH_SIZE: 512 # SHOULDDO (all integer values are possible, depending on your hardware and data size)
EPOCHS: 300 # SHOULDDO (all integer values are possible)
LR_FIXED: 0.0005 # SHOULDDO Learning Rate for training (all float values are possible)
VAE_LOSS: "KL" # SHOULDDO  loss function for VAE to fit standard normal distribution (KL or MMD)
RECONSTR_LOSS: "MSE" # SHOULDDO (BCE or MSE)

PREDICT_SPLIT: "all" # SHOULDDO data split the trainied model uses to generate latent space (all, train, test, valid are possible) default is all

# # Ontix specific
# NON_ONT_LAYER: 1 # OPTIONAL Specify additional layer for ontix if should LATENT_DIM_FIXED < Ontology Dimension
# FILE_ONT_LVL1: "full_ont_lvl1_reactome.txt" # OPTIONAL Mandotory for ontix; LVL1 specifies relationship of features and first ontology level in sparse decoder
# FILE_ONT_LVL2: "full_ont_lvl2_reactome.txt" # OPTIONAL for ontix; LVL2 specifies relationship of first ontology level and a second level (dim(lvl1) > dim(lvl2)) in sparse decoder


# # X-modalix specific
# TRANSLATE: "KLIMBIM_to_ANDERERKLIMBIM" # TODO, TRAIN_TYPE=translate, when <FROMDATA>_to_<TODATA>
# GAMMA: 5 # OPTIONAL weighs the latent space classifier loss for th translation case
# DELTA_PAIR: 5 # OPTIONAL weigths for pairing loss term (xmodalix)
# DELTA_CLASS: 5 # OPTIONAL weigths for class/supervision loss term (xmodalix)
# PRETRAIN_TARGET_MODALITY: null #SHOULDDO For `x-modalix` VAE's can be pretrained before latent spaces will be aligned. Options are pretrain_image or gamma_anneal or null to switch off pretraining.
# ANNEAL_PRETRAINING: False #OPTIONAL (True or False) indicates if beta annealing should be used in pretraining target modality or while using gamma annealing in pretraining
# PRETRAIN_EPOCHS: 100 #OPTIONAL (all integer values are possible) number of epochs of gamma_anneal or pretrain_image (additional to EPOCHS) (set 0 if not wanted)
# CLASS_PARAM: "CANCER_TYPE_ACRONYM" #OPTIONAL Annotation parameter (str) to which the latent space is aligned for the MODEL_TYPE "x-modalix" and is a column as defined in the Annotation DATA_TYPE. Set to null if not wanted.

# OPTUNA TUNING VARS-----------------------------------------------------------
# -----------------------------------------------------------------------------
LAYERS_LOWER_LIMIT: 2 # OPTIONAL  minumum number of layers for hyperparamter tuning(all integer values are possible)
LAYERS_UPPER_LIMIT: 5 # OPTIONAL maximum number of layers for hyperparamter tuning(all integer values are possible)
LR_LOWER_LIMIT: 0.0001 # OPTIONAL minimum learning rate for hyperparamter tuning(all float values are possible)
LR_UPPER_LIMIT: 0.05 # OPTIONAL maximum learning rate for hyperparamter tuning(all float values are possible)
DROPOUT_LOWER_LIMIT: 0.00 # OPTIONAL minimum dropout rate for hyperparamter tuning(all float values are possible)
DROPOUT_UPPER_LIMIT: 0.50 # OPTIONAL maximum dropout rate for hyperparamter tuning(all float values are possible)
OPTUNA_TRIALS: 10 # OPTIONAL number of trials for hyperparamter tuning(all integer values are possible)


# EVALUATION and VISUALIZATION ------------------------------------------------
# -----------------------------------------------------------------------------
DIM_RED_METH: "UMAP" # OPTIPNAL PCA or UMAP or TSNE
CLUSTER_ALG: "KMeans" # OPTIONAL KMeans or HDBScan
CLUSTER_N: 32 # OPTIONAL Number of clusters for KMeans
MIN_CLUSTER_N: 20 # OPTIONAL Minimal number of samples per cluster as crucial parameter for HDBSCAN
CLINIC_PARAM: # TODO needs to be a column name of the ANNOTATION data
  - "CANCER_TYPE_ACRONYM" # Feature to be visualized on latent space plot and used for learning alling of downstream machine learning task
  - "TMB_NONSYNONYMOUS"
ML_TYPE: "Auto-detect" # TODO For the given CLINIC_PARAM either regression or classification ML tasks will be performed. If auto-detect, non-string parameters will be assumed to be regression tasks.
# ML_TYPE: # TODO manual specification if given CLINIC_PARAM are either regression or classification (recommended). Must be similar entries as CLINIC_PARAM
#   CANCER_TYPE: "classification" # OPTIONAL classification or regression
#   TMB_NONSYNONYMOUS: "regression" # OPTIONAL classification or regression
ML_ALG: # OPTIONAL which machine learning algorithms should be used for evaluation, Linear, RF, SVM are possible
  - Linear
  - RF
ML_SPLIT: "use-split" # OPTIONAL "use-split" or "CV-on-all-data"
ML_TASKS: # OPTIONAL which dimension reduction methods should be used for the ml task, Latent, UMAP, PCA, RandomFeature are possible
  - Latent
  - PCA

RUN_ID: TuningExample