Baseline models - Synthesize Bio

Overview

Baseline models generate synthetic gene expression data from metadata alone. You describe the biological conditions—tissue type, disease state, perturbations, cell type, etc.—and the model generates realistic expression profiles matching those conditions. This is the most common use case: generating synthetic data for conditions where real data may be scarce or unavailable.

Available Models

gem-1-bulk: Bulk RNA-seq baseline model
gem-1-sc: Single-cell RNA-seq baseline model

library(rsynthbio)

Creating a Query

The structure of the query required by the API is specific to each model. Use get_example_query() to get a correctly structured example for your chosen model.

# Get the example query structure for a specific model
example_query <- get_example_query(model_id = "gem-1-bulk")$example_query

# Inspect the query structure
str(example_query)

The query consists of:

sampling_strategy: The prediction mode that controls how expression data is generated:
- “sample generation”: Generates realistic-looking synthetic data with measurement error (bulk only)
- “mean estimation”: Provides stable mean estimates of expression levels (bulk and single-cell)
inputs: A list of biological conditions to generate data for

Each input contains metadata (describing the biological sample) and num_samples (how many samples to generate).

Making a Prediction

Once your query is ready, send it to the API to generate gene expression data:

# Create a query for the bulk model
query <- get_example_query(model_id = "gem-1-bulk")$example_query

# Submit and get results
result <- predict_query(query, model_id = "gem-1-bulk")

The result is a list containing two data frames: metadata and expression.

Single-Cell Example

# Create a query for the single-cell model
sc_query <- get_example_query(model_id = "gem-1-sc")$example_query

# Submit and get results
sc_result <- predict_query(sc_query, model_id = "gem-1-sc")

Single-cell models only support "mean estimation" mode.

Query Parameters

In addition to metadata, queries support several optional parameters that control the generation process.

sampling_strategy (character, required)

Controls the type of prediction the model generates. This parameter is required in all queries. Available modes:

“sample generation”: The model generates realistic-looking synthetic data that captures measurement error. This mode is useful when you want data that mimics real experimental measurements. (Bulk only)
“mean estimation”: The model creates a distribution capturing biological heterogeneity consistent with the supplied metadata, then returns the mean of that distribution. This mode is useful when you want a stable estimate of expected expression levels. (Bulk and single-cell)

# Bulk query with sample generation
bulk_query <- get_example_query(model_id = "gem-1-bulk")$example_query
bulk_query$sampling_strategy <- "sample generation"

# Bulk query with mean estimation
bulk_query_mean <- get_example_query(model_id = "gem-1-bulk")$example_query
bulk_query_mean$sampling_strategy <- "mean estimation"

# Single-cell query (must use mean estimation)
sc_query <- get_example_query(model_id = "gem-1-sc")$example_query
sc_query$sampling_strategy <- "mean estimation" # Required for single-cell

total_count (integer, optional)

Library size used when converting predicted log CPM back to raw counts. Higher values scale counts up proportionally.

Default: 10,000,000 for bulk; 10,000 for single-cell

# Create a query and add custom total_count
query <- get_example_query(model_id = "gem-1-bulk")$example_query
query$total_count <- 5000000

deterministic_latents (logical, optional)

If TRUE, the model uses the mean of each latent distribution (p(z|metadata)) instead of sampling. This removes randomness from latent sampling and produces deterministic outputs for the same inputs.

Default: FALSE (sampling is enabled)

# Create a query and enable deterministic latents
query <- get_example_query(model_id = "gem-1-bulk")$example_query
query$deterministic_latents <- TRUE

seed (integer, optional)

Random seed for reproducibility when using stochastic sampling.

# Create a query with a specific seed
query <- get_example_query(model_id = "gem-1-bulk")$example_query
query$seed <- 42

Combining Parameters

You can combine multiple parameters in a single query:

# Create a query and add multiple parameters
query <- get_example_query(model_id = "gem-1-bulk")$example_query
query$total_count <- 8000000
query$deterministic_latents <- TRUE
query$sampling_strategy <- "mean estimation"

results <- predict_query(query, model_id = "gem-1-bulk")

Valid Metadata Keys

The input metadata is a list of lists. Here is the full list of valid metadata keys:

Biological

age_years
cell_line_ontology_id
cell_type_ontology_id
developmental_stage
disease_ontology_id
ethnicity
genotype
race
sample_type (“cell line”, “organoid”, “other”, “primary cells”, “primary tissue”, “xenograft”)
sex (“male”, “female”)
tissue_ontology_id

Perturbational

perturbation_dose (number and unit separated by a space, e.g., “10 um”)
perturbation_ontology_id
perturbation_time (number and unit separated by a space, e.g., “24 hours”)
perturbation_type (“coculture”, “compound”, “control”, “crispr”, “genetic”, “infection”, “other”, “overexpression”, “peptide or biologic”, “shrna”, “sirna”)

Technical

study (Bioproject ID)
library_selection (e.g., “cDNA”, “polyA”, “Oligo-dT” - see https://ena-docs.readthedocs.io/en/latest/submit/reads/webin-cli.html#permitted-values-for-library-selection)
library_layout (“PAIRED”, “SINGLE”)
platform (“illumina”)

Valid Metadata Values

The following are the valid values or expected formats for selected metadata keys:

Metadata Field	Requirement / Example
`cell_line_ontology_id`	Requires a Cellosaurus ID.
`cell_type_ontology_id`	Requires a CL ID.
`disease_ontology_id`	Requires a MONDO ID.
`perturbation_ontology_id`	Must be a valid Ensembl gene ID (e.g., `ENSG00000156127`), ChEBI ID (e.g., `CHEBI:16681`), ChEMBL ID (e.g., `CHEMBL1234567`), or NCBI Taxonomy ID (e.g., `9606`).
`tissue_ontology_id`	Requires a UBERON ID.

We highly recommend using the EMBL-EBI Ontology Lookup Service to find valid IDs for your metadata. Models have a limited acceptable range of metadata input values. If you provide a value that is not in the acceptable range, the API will return an error.

Modifying Query Inputs

You can customize the query inputs to fit your specific research needs:

# Get a base query
query <- get_example_query(model_id = "gem-1-bulk")$example_query

# Adjust number of samples for the first input
query$inputs[[1]]$num_samples <- 10

# Add a new condition
query$inputs[[3]] <- list(
  metadata = list(
    sex = "male",
    sample_type = "primary tissue",
    tissue_ontology_id = "UBERON:0002371"
  ),
  num_samples = 5
)

Working with Results

# Access metadata and expression matrices
metadata <- result$metadata
expression <- result$expression

# Check dimensions
dim(expression)

# View metadata sample
head(metadata)

You may want to process the data in chunks or save it for later use:

# Save results to RDS file
saveRDS(result, "synthesize_results.rds")

# Load previously saved results
result <- readRDS("synthesize_results.rds")

# Export as CSV
write.csv(result$expression, "expression_matrix.csv")
write.csv(result$metadata, "sample_metadata.csv")

​Overview

​Available Models

​Creating a Query

​Making a Prediction

​Single-Cell Example

​Query Parameters

​sampling_strategy (character, required)

​total_count (integer, optional)

​deterministic_latents (logical, optional)

​seed (integer, optional)

​Combining Parameters

​Valid Metadata Keys

​Biological

​Perturbational

​Technical

​Valid Metadata Values

​Modifying Query Inputs

​Working with Results

Overview

Available Models

Creating a Query

Making a Prediction

Single-Cell Example

Query Parameters

sampling_strategy (character, required)

total_count (integer, optional)

deterministic_latents (logical, optional)

seed (integer, optional)

Combining Parameters

Valid Metadata Keys

Biological

Perturbational

Technical

Valid Metadata Values

Modifying Query Inputs

Working with Results