DiffBio

End-to-End Differentiable Bioinformatics Pipelines

Built on Datarax, Artifex, Opifex, and Calibrax | Powered by JAX & Flax NNX

---

⚠️ Early Development - API Unstable

DiffBio is currently in early development and undergoing rapid iteration. Please be aware of the following implications:

Area	Status	Impact
API	🔄 Unstable	Breaking changes are expected. Public interfaces may change without deprecation warnings. Pin to specific commits if stability is required.
Tests	🔄 In Flux	Test suite is being expanded. Some tests may fail or be skipped. Coverage metrics are improving but not yet full.
Documentation	🔄 Evolving	Docs may not reflect current implementation. Code examples might be outdated. Refer to source code and tests for accurate usage.

We recommend waiting for a stable release (v1.0) before using DiffBio in production. For research and experimentation, proceed with the understanding that APIs will evolve.

---

Overview

DiffBio is a framework for building end-to-end differentiable bioinformatics pipelines. By replacing discrete operations with differentiable relaxations, DiffBio enables gradient-based optimization through entire analysis workflows.

DiffBio is the biology-specific differentiable operator layer of a wider JAX/NNX scientific ML ecosystem. It uses:

• Datarax for operator and dataflow contracts
• Artifex for reusable model-building and transformer components
• Opifex for scientific ML and advanced optimization primitives
• Calibrax for metrics, benchmarking, comparison, and regression control

Traditional bioinformatics pipelines use discrete operations (hard thresholds, argmax decisions) that block gradient flow. DiffBio addresses this by:

• Soft quality filtering using sigmoid-based weights instead of hard cutoffs
• Differentiable pileup with soft position assignments via temperature-controlled softmax
• Soft alignment scoring replacing discrete Smith-Waterman with continuous relaxations
• End-to-end training of complete pipelines using gradient descent

This enables learning optimal pipeline parameters directly from data, rather than manual tuning.

Features

• 40+ Differentiable Operators covering alignment, variant calling, single-cell analysis, epigenomics, RNA-seq, preprocessing, normalization, multi-omics, drug discovery, and protein/RNA structure
• 6 End-to-End Pipelines for variant calling, enhanced variant calling, single-cell analysis, differential expression, perturbation, and preprocessing
• GPU-Accelerated computation via JAX's XLA compilation
• Composable Architecture built on the Datarax, Artifex, Opifex, and Calibrax stack
• Training Utilities with gradient clipping, custom loss functions, and synthetic data generation

For complete operator and pipeline listings, see the Operators Overview and Pipelines Overview in the documentation.

Installation

# Clone the repository
git clone https://github.com/avitai/DiffBio.git
cd DiffBio
# Install with uv
uv sync

Quick Start

Using Individual Operators

importjaximportjax.numpyasjnpfromflaximportnnxfromdiffbio.operatorsimportDifferentiableQualityFilter, QualityFilterConfig# Quality filtering with learnable threshold (default initial_threshold=20.0)quality_filter=DifferentiableQualityFilter(
QualityFilterConfig(initial_threshold=20.0),
rngs=nnx.Rngs(0),
)
# Apply to a one-hot encoded sequence with per-position quality scoresquality_scores=jnp.array([35.0, 15.0, 28.0, 10.0])
sequence=jax.nn.one_hot(jnp.array([0, 1, 2, 3]), 4) # (length, alphabet=4)data= {"sequence": sequence, "quality_scores": quality_scores}
filtered_data, _, _=quality_filter.apply(data, {}, None)
# filtered_data["sequence"] — sequence with low-quality positions softly suppressed# filtered_data["quality_scores"] — pass-through quality values

Using the Variant Calling Pipeline

fromdiffbio.pipelinesimport (
VariantCallingPipeline,
VariantCallingPipelineConfig,
create_variant_calling_pipeline,
)
# Create pipeline with default configurationpipeline=create_variant_calling_pipeline(
reference_length=100,
num_classes=3, # ref, SNP, indelhidden_dim=32,
seed=42,
)
# Process readsbatch_data= {
"reads": reads, # (num_reads, read_length, 4)"positions": positions, # (num_reads,)"quality": quality, # (num_reads, read_length)
}
result, _, _=pipeline.apply(batch_data, {}, None)
# result["logits"] contains per-position variant predictions# result["probabilities"] contains class probabilities

Training a Pipeline

fromdiffbio.utilsimport (
Trainer,
TrainingConfig,
cross_entropy_loss,
create_synthetic_training_data,
data_iterator,
)
# Generate synthetic training datainputs, targets=create_synthetic_training_data(
num_samples=100,
num_reads=10,
read_length=50,
reference_length=100,
variant_rate=0.1,
)
# Configure trainingconfig=TrainingConfig(
learning_rate=1e-3,
num_epochs=50,
log_every=10,
grad_clip_norm=1.0,
)
# Create trainertrainer=Trainer(pipeline, config)
# Define loss functiondefloss_fn(predictions, targets):
returncross_entropy_loss(
predictions["logits"],
targets["labels"],
num_classes=3,
)
# Traintrainer.train(
data_iterator_fn=lambda: data_iterator(inputs, targets),
loss_fn=loss_fn,
)
# Access trained pipelinetrained_pipeline=trainer.pipeline

Architecture

DiffBio sits on a layered ecosystem rather than standing alone:

Layer	Library	Role In DiffBio
Execution contracts	Datarax	Operator, data-source, and pipeline contracts
Modeling substrate	Artifex	Reusable transformer and generative-model components
Scientific ML substrate	Opifex	Scientific optimization, operator learning, and advanced training methods
Evaluation substrate	Calibrax	Metrics, benchmarking, comparison, profiling, and regression checks

DiffBio itself sits on top of these as the biology-specific layer: differentiable biological operators and end-to-end pipeline compositions (alignment, variant calling, single-cell analysis, drug discovery, structural biology, multi-omics).

Each DiffBio operator inherits from Datarax's OperatorModule and implements:

apply(data, state, metadata) -> (output_data, output_state, output_metadata)

This enables:

• Composition: Chain operators into pipelines
• Batch processing: Automatic vectorization via apply_batch()
• Gradient flow: End-to-end differentiability through the pipeline

Operator Composition

apply() runs an operator on a single element (no batch dimension). Operators are chained by threading the (data, state, metadata) triple returned by apply() into the next operator:

data, state, metadata=quality_filter.apply(element_data, {}, None)
data, state, metadata=pileup.apply(data, state, metadata)
data, state, metadata=classifier.apply(data, state, metadata)
# `data` is a dict of JAX arrays — read out the per-position predictionspredictions=data["logits"]

For batched data wrapped in a Datarax Batch, call the operator directly (or use apply_batch()); both delegate to the same code path:

fromdataraximportBatchbatch=Batch.from_parts(...) # construct from a list of elementsbatch=quality_filter(batch) # equivalent to quality_filter.apply_batch(batch)batch=pileup(batch)
batch=classifier(batch)

Testing

# Run all tests
uv run pytest -vv
# Run with coverage
uv run pytest -vv --cov=src/ --cov-report=term-missing
# Run specific test modules
uv run pytest tests/operators/ -vv
uv run pytest tests/pipelines/ -vv
uv run pytest tests/integration/ -vv

Project Structure

DiffBio/
├── src/diffbio/
│ ├── core/ # Base operators, graph utils, soft ops, neural components
│ ├── operators/ # 40+ differentiable operators
│ │ ├── alignment/ # Smith-Waterman, profile HMM, soft MSA
│ │ ├── assembly/ # GNN assembly, metagenomic binning
│ │ ├── crispr/ # Guide RNA scoring
│ │ ├── drug_discovery/ # Fingerprints, ADMET, AttentiveFP, MACCS keys
│ │ ├── epigenomics/ # Peak calling, chromatin state, contextual epigenomics
│ │ ├── foundation_models/ # Geneformer/scGPT adapters, transformer encoders
│ │ ├── mapping/ # Neural read mapping
│ │ ├── metabolomics/ # Spectral similarity
│ │ ├── molecular_dynamics/ # Force fields, MD integrators
│ │ ├── multiomics/ # Hi-C, spatial deconvolution, multi-omics VAE
│ │ ├── normalization/ # VAE normalizer, UMAP, PHATE, embeddings
│ │ ├── population/ # Ancestry estimation
│ │ ├── preprocessing/ # Adapter removal, duplicate weighting, error correction
│ │ ├── protein/ # Secondary structure
│ │ ├── rna_structure/ # RNA folding
│ │ ├── rnaseq/ # Splicing PSI, motif discovery
│ │ ├── singlecell/ # Clustering, trajectory, velocity, GRN, batch correction, ...
│ │ ├── statistical/ # HMM, NB GLM, EM quantification
│ │ └── variant/ # Pileup, classifiers, CNV segmentation
│ ├── pipelines/ # 6 end-to-end pipelines
│ ├── losses/ # Alignment, biological-regularization, single-cell, statistical, metric
│ ├── sources/ # Data loaders (FASTA, BAM, AnnData, MoleculeNet, indexed views)
│ ├── splitters/ # Random, stratified, scaffold, Tanimoto, sequence-identity
│ ├── samplers/ # Perturbation samplers
│ ├── sequences/ # DNA / RNA encoding utilities
│ ├── evaluation/ # Evaluation runner and graders
│ └── utils/ # Training utilities, dependency-runtime checks
├── tests/ # Unit, integration, and benchmark tests
├── benchmarks/ # Domain benchmarks with training + baselines
├── examples/ # Runnable example scripts paired with notebooks
└── docs/ # MkDocs documentation

Requirements

• Python 3.11+
• JAX 0.6.1+
• Flax 0.12+
• Optax 0.1.4+
• jaxtyping 0.2.20+
• Datarax, Artifex, Opifex, and Calibrax (installed automatically from PyPI)

License

MIT License. See LICENSE for details.

Acknowledgments

DiffBio builds on ideas from:

• Datarax: Composable data processing framework
• Artifex: Generative-model and transformer substrate
• Opifex: Scientific ML and advanced optimization substrate
• Calibrax: Benchmarking, comparison, and regression substrate
• Flax NNX: Neural network library for JAX