Part 4: Partial-Connected Neural Networks

Model Architecture

In partial-connected neural networks, SNPs can be divided into groups by users, and each group connects to its own intermediate trait in the middle layer. This is useful when:

• Different genomic regions affect different intermediate traits
• You want to model pathway-specific effects
• You have prior biological knowledge about SNP-trait relationships

  Genotype Group 1 ─────► Omics/Latent 1 ─┐
  Genotype Group 2 ─────► Omics/Latent 2 ─┼──► Phenotype
  Genotype Group 3 ─────► Omics/Latent 3 ─┘

Example: Partial-Connected Network with Observed Omics

This example demonstrates:

• Number of genotype groups: 3 (simulated by splitting SNPs)
• Number of omics features: 3 (one per genotype group)
• Bayesian method: BayesC
• Activation function: sigmoid

# Step 1: Load packages
using NNMM
using NNMM.Datasets
using DataFrames
using CSV
using Statistics
using Random

Random.seed!(123)

# Step 2: Load simulated data
geno_path = Datasets.dataset("genotypes_1000snps.txt", dataset_name="simulated_omics_data")
pheno_path = Datasets.dataset("phenotypes_sim.txt", dataset_name="simulated_omics_data")
pheno_df = CSV.read(pheno_path, DataFrame)

# Step 3: Split genotypes into 3 groups (for demonstration)
# Read full genotype file and split into 3 parts
geno_full = CSV.read(geno_path, DataFrame)
n_markers = ncol(geno_full) - 1  # Exclude ID column
markers_per_group = div(n_markers, 3)

# Create 3 genotype group files
geno1_cols = vcat(:ID, names(geno_full)[2:markers_per_group+1])
geno2_cols = vcat(:ID, names(geno_full)[markers_per_group+2:2*markers_per_group+1])
geno3_cols = vcat(:ID, names(geno_full)[2*markers_per_group+2:end])

geno1_df = geno_full[:, geno1_cols]
geno2_df = geno_full[:, geno2_cols]
geno3_df = geno_full[:, geno3_cols]

genofile1 = "geno_group1.csv"
genofile2 = "geno_group2.csv"
genofile3 = "geno_group3.csv"

CSV.write(genofile1, geno1_df)
CSV.write(genofile2, geno2_df)
CSV.write(genofile3, geno3_df)

# Step 4: Create omics file (3 features, one per genotype group)
omics_df = pheno_df[:, [:ID, :omic1, :omic2, :omic3]]
omics_file = "omics_partial.csv"
CSV.write(omics_file, omics_df; missingstring="NA")

# Step 5: Create phenotype file
trait_df = pheno_df[:, [:ID, :trait1]]
trait_file = "trait_partial.csv"
CSV.write(trait_file, trait_df; missingstring="NA")

# Step 6: Define layers
# KEY: Provide multiple genotype files as a vector for partial connection
layers = [
    # Layer 1: Multiple genotype files = partial-connected network
    Layer(
        layer_name = "geno",
        data_path = [genofile1, genofile2, genofile3]  # Vector of 3 files
    ),
    # Layer 2: Observed omics (one per genotype group)
    Layer(
        layer_name = "omics",
        data_path = omics_file,
        missing_value = "NA"
    ),
    # Layer 3: Phenotypes
    Layer(
        layer_name = "phenotypes",
        data_path = trait_file,
        missing_value = "NA"
    )
]

# Step 7: Define equations
# The equation uses a single layer name ("geno"), but NNMM internally
# creates separate equations for each genotype group (geno1, geno2, geno3)
equations = [
    Equation(
        from_layer_name = "geno",
        to_layer_name = "omics",
        equation = "omics = intercept + geno",  # Internal: omic1=intercept+geno1; omic2=intercept+geno2; omic3=intercept+geno3
        traits = ["omic1", "omic2", "omic3"],  # Must match number of genotype files
        method = "BayesC",
        estimatePi = true
    ),
    Equation(
        from_layer_name = "omics",
        to_layer_name = "phenotypes",
        equation = "phenotypes = intercept + omics",
        traits = ["trait1"],
        method = "BayesC",
        activation_function = "sigmoid"
    )
]

# Step 8: Run analysis
out = runNNMM(layers, equations;
    chain_length = 5000,
    burnin = 1000,
    output_folder = "nnmm_partial_results"
)

# Step 9: Check accuracy
ebv = out["EBV_NonLinear"]
ebv.ID = string.(ebv.ID)
pheno_df.ID = string.(pheno_df.ID)
results = innerjoin(ebv, pheno_df[:, [:ID, :genetic_total]], on=:ID)
accuracy = cor(Float64.(results.EBV), results.genetic_total)
println("Prediction accuracy: ", round(accuracy, digits=4))

# Cleanup
rm(genofile1, force=true)
rm(genofile2, force=true)
rm(genofile3, force=true)
rm(omics_file, force=true)
rm(trait_file, force=true)

Example: Partial-Connected Network with Latent Traits

When you don't have observed omics data but want to use a partial-connected architecture:

# Step 1: Load packages
using NNMM
using NNMM.Datasets
using DataFrames
using CSV
using Statistics
using Random

Random.seed!(123)

# Step 2: Load simulated data and split genotypes (same as above)
geno_path = Datasets.dataset("genotypes_1000snps.txt", dataset_name="simulated_omics_data")
pheno_path = Datasets.dataset("phenotypes_sim.txt", dataset_name="simulated_omics_data")
pheno_df = CSV.read(pheno_path, DataFrame)

geno_full = CSV.read(geno_path, DataFrame)
n_markers = ncol(geno_full) - 1
markers_per_group = div(n_markers, 3)

geno1_cols = vcat(:ID, names(geno_full)[2:markers_per_group+1])
geno2_cols = vcat(:ID, names(geno_full)[markers_per_group+2:2*markers_per_group+1])
geno3_cols = vcat(:ID, names(geno_full)[2*markers_per_group+2:end])

CSV.write("geno_group1.csv", geno_full[:, geno1_cols])
CSV.write("geno_group2.csv", geno_full[:, geno2_cols])
CSV.write("geno_group3.csv", geno_full[:, geno3_cols])

# Step 3: Create latent trait file (all missing)
n_individuals = nrow(pheno_df)
latent_df = DataFrame(
    ID = pheno_df.ID,
    latent1 = fill(missing, n_individuals),
    latent2 = fill(missing, n_individuals),
    latent3 = fill(missing, n_individuals)
)
CSV.write("latent_partial.csv", latent_df; missingstring="NA")

# Step 4: Create phenotype file
CSV.write("trait_partial.csv", pheno_df[:, [:ID, :trait1]]; missingstring="NA")

# Step 5: Define layers
layers = [
    Layer(layer_name="geno", data_path=["geno_group1.csv", "geno_group2.csv", "geno_group3.csv"]),
    Layer(layer_name="latent", data_path="latent_partial.csv", missing_value="NA"),
    Layer(layer_name="phenotypes", data_path="trait_partial.csv", missing_value="NA")
]

# Step 6: Define equations
equations = [
    Equation(from_layer_name="geno", to_layer_name="latent",
             equation="latent = intercept + geno",
             traits=["latent1", "latent2", "latent3"], method="BayesC"),
    Equation(from_layer_name="latent", to_layer_name="phenotypes",
             equation="phenotypes = intercept + latent",
             traits=["trait1"], activation_function="tanh")
]

# Step 7: Run analysis
out = runNNMM(layers, equations; chain_length=5000, burnin=1000, output_folder="nnmm_partial_latent")

# Step 8: Check accuracy
ebv = out["EBV_NonLinear"]
ebv.ID = string.(ebv.ID)
pheno_df.ID = string.(pheno_df.ID)
results = innerjoin(ebv, pheno_df[:, [:ID, :genetic_total]], on=:ID)
println("Accuracy: ", round(cor(Float64.(results.EBV), results.genetic_total), digits=4))

# Cleanup
for f in ["geno_group1.csv", "geno_group2.csv", "geno_group3.csv", "latent_partial.csv", "trait_partial.csv"]
    rm(f, force=true)
end

Key Points

1. Number of files = Number of omics: The number of genotype files must exactly match the number of omics/latent features.

2. Automatic naming: When you provide 3 genotype files and layer_name="geno", NNMM internally names them geno1, geno2, geno3.

3. Independent models: Each genotype group gets its own independent Bayesian model for predicting its corresponding omics feature.

4. Mixed with residuals: You can add extra latent nodes for residual effects by including additional (all-missing) columns in the omics file.

Comparison: Fully-Connected vs Partial-Connected

Output Files

Output files are similar to fully-connected networks, but with group-specific marker effect files:

• MCMC_samples_marker_effects_geno1_omic1.txt
• MCMC_samples_marker_effects_geno2_omic2.txt
• MCMC_samples_marker_effects_geno3_omic3.txt