Glossary

Basenji

A deep convolutional neural network that predicts cell-type-specific epigenetic and transcriptional profiles directly from 131-kilobase DNA sequences, capturing long-range regulatory interactions without explicit chromatin contact data.

Get in touch Learn more

Data scientist building training data pipeline on laptop, data preprocessing visible, technical workspace.

REGULATORY GENOMICS

What is Basenji?

Basenji is a deep convolutional neural network architecture designed to predict cell-type-specific epigenetic and transcriptional profiles directly from raw DNA sequences spanning 131 kilobases.

Basenji is a deep convolutional neural network that predicts functional genomic activity—including chromatin accessibility, transcription factor binding, and gene expression—directly from 131-kilobase DNA sequences. By processing extended genomic contexts, the model captures long-range regulatory interactions between distal enhancers and target promoters without requiring experimental chromatin conformation data as input.

The architecture employs dilated convolutions with exponentially increasing dilation rates to achieve a receptive field covering 131 kb, enabling it to model regulatory syntax across megabase-scale distances. Trained on thousands of epigenomic datasets from the ENCODE and Roadmap Epigenomics consortia, Basenji produces cell-type-specific predictions and can perform in silico saturation mutagenesis to prioritize functional non-coding variants associated with human disease.

ARCHITECTURAL INNOVATIONS

Key Features of Basenji

Basenji is a deep convolutional neural network that predicts cell-type-specific epigenetic and transcriptional profiles directly from 131-kilobase DNA sequences, capturing long-range regulatory interactions without explicit chromatin contact data.

Massive 131 kb Receptive Field

Basenji processes 131,072 base pairs of raw DNA sequence in a single forward pass, dramatically exceeding earlier models like DeepSEA (1 kb). This expansive window is achieved through dilated convolutions that exponentially increase the receptive field without a proportional increase in parameters. The architecture captures long-range enhancer-promoter interactions spanning tens of kilobases, enabling the model to learn regulatory grammar that depends on distal elements.

Multi-Task Prediction Across Assays

A single Basenji model simultaneously predicts thousands of genomic tracks spanning multiple functional classes:

DNase-seq chromatin accessibility profiles
ChIP-seq transcription factor binding
CAGE cap analysis gene expression
Histone modification marks (H3K4me3, H3K27ac, etc.)

Shared representations across tasks improve generalization, particularly for assays with limited training data.

Dilated Convolutional Architecture

Basenji employs a residual dilated convolutional network where dilation rates double across successive layers (1, 2, 4, 8, 16, 32, 64, 128, 256, 512). This design:

Exponentially expands the receptive field to 131 kb
Maintains parameter efficiency — no increase in kernel size
Preserves base-resolution prediction capability
Avoids the quadratic complexity of self-attention mechanisms

The architecture predates Enformer's transformer-based approach while achieving comparable long-range modeling.

Direct Sequence-to-Profile Prediction

Basenji maps raw DNA sequence directly to experimental coverage profiles without requiring:

Pre-computed chromatin contact maps (Hi-C)
Explicit enhancer-promoter annotations
Hand-engineered genomic features

The model learns regulatory syntax end-to-end from one-hot encoded nucleotide sequences, discovering motifs, motif combinations, and long-range interaction rules autonomously through supervised training on experimental data.

In Silico Saturation Mutagenesis

Basenji enables computational perturbation experiments by systematically introducing virtual mutations and measuring predicted changes in regulatory activity. This approach:

Identifies causal non-coding variants associated with disease
Quantifies the effect size of every possible single-nucleotide substitution
Prioritizes variants from genome-wide association studies (GWAS)
Reveals allele-specific regulatory effects without requiring heterozygous samples

The method transforms variant interpretation from statistical association to mechanistic prediction.

Cross-Species Regulatory Conservation

Basenji models trained on human data can be applied to orthologous sequences from other species to assess regulatory conservation. By comparing predicted activity profiles across species at aligned loci, researchers can:

Identify conserved regulatory elements under purifying selection
Detect lineage-specific regulatory innovations
Validate enhancer function through evolutionary constraint
Prioritize elements for experimental follow-up in model organisms

BASENJI MODEL CLARIFICATIONS

Frequently Asked Questions

Precise technical answers to common questions about the Basenji architecture, its training methodology, and its application to regulatory genomics prediction tasks.

Basenji is a deep convolutional neural network architecture that predicts cell-type-specific epigenetic and transcriptional profiles directly from 131-kilobase (131,072 base pair) DNA sequences. The model processes raw genomic sequence through a series of dilated convolutional layers that exponentially expand the receptive field, enabling it to capture long-range regulatory interactions spanning over 100 kilobases without requiring explicit chromatin contact data such as Hi-C. Basenji operates as a multi-task regression model, simultaneously predicting thousands of functional genomics tracks—including CAGE expression, DNase-seq accessibility, and ChIP-seq binding profiles—across multiple human and mouse cell types. By training on 128-base-pair resolution bins, the model learns to recognize the sequence determinants of regulatory activity, including transcription factor binding motifs, their combinatorial grammar, and distal enhancer-promoter interactions. The architecture's key innovation is its ability to model distal regulation through dilated convolutions rather than attention mechanisms, making it computationally efficient while maintaining the capacity to capture interactions across the full 131-kilobase input window.

ARCHITECTURAL COMPARISON

Basenji vs. Enformer vs. DeepSEA

Comparative analysis of three foundational deep learning architectures for predicting regulatory genomic profiles directly from DNA sequence.

Feature	Basenji	Enformer	DeepSEA
Input Sequence Length	131 kb	200 kb	1 kb
Core Architecture	Dilated CNNs	Transformer + CNNs	Standard CNNs
Long-Range Interaction Modeling
Multi-Task Prediction
Base-Resolution Output
Explicit Chromatin Contact Data Required
Training Dataset	CAGE, ChIP-seq, DNase-seq	CAGE, ChIP-seq, DNase-seq	ChIP-seq, DNase-seq, TF binding
Receptive Field Expansion Method	Exponential dilation	Multi-head attention	Stacked convolutions

Enabling Efficiency, Speed & Accuracy

Intelligent Analysis, Decision & Execution

We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.

Talk to Us

Search across company data

Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.

Useful when people spend too long searching or get different answers from different systems.

Enterprise searchRAGPermissions

Automate internal workflows

Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.

Useful when repetitive work moves across multiple tools and teams.

AI agentsWorkflow automationGovernance

Add AI to products and internal tools

Build assistants, guided actions, or decision support into the software your team or customers already use.

Useful when AI needs to be part of the product, not a separate tool.

AI integrationDecision supportModel routing

BASENJI ECOSYSTEM

Related Terms

Key concepts, architectures, and evaluation methods that contextualize the Basenji framework within the broader landscape of genomic deep learning.

Enformer

A transformer-based successor architecture that extends the receptive field to 200 kilobases, explicitly modeling long-range interactions through multi-head attention. While Basenji uses dilated convolutions to capture 131kb dependencies, Enformer leverages self-attention mechanisms to directly weigh the influence of distal regulatory elements on gene expression prediction.

EXPLORE

Dilated Convolution

The core architectural innovation enabling Basenji's 131-kilobase receptive field. Unlike standard convolutions, dilated convolutions insert gaps between kernel elements, exponentially expanding the field of view without increasing parameter count:

Exponential dilation: Each layer doubles the spacing, e.g., rates of 1, 2, 4, 8
No pooling: Preserves base-resolution information unlike max-pooling architectures
Parameter efficiency: Captures long-range dependencies with far fewer weights than dense layers

Multi-Task Learning

Basenji jointly predicts thousands of epigenetic and transcriptional profiles across multiple cell types and assays from a single DNA input. This training paradigm forces the shared hidden layers to learn a universal regulatory grammar, improving generalization on low-data assays by leveraging common sequence features learned from high-data tracks.

Hold-Out Chromosome Validation

Basenji employs a rigorous cross-validation strategy where entire chromosomes (typically chromosomes 7 and 8) are reserved for testing. This prevents homology leakage—the risk that highly similar sequences appear in both training and test splits—which would artificially inflate performance metrics and mask true generalization capability.

In Silico Mutagenesis

A computational perturbation method used to interpret Basenji predictions. By systematically introducing virtual nucleotide substitutions and measuring the resulting change in predicted regulatory activity, researchers can identify causal regulatory variants and quantify the effect size of non-coding mutations on gene expression without wet-lab experiments.

CAGE Signal Prediction

Basenji predicts Cap Analysis of Gene Expression (CAGE) signals directly from DNA sequence, providing a quantitative readout of transcription start site activity. This enables the model to infer promoter usage and transcript abundance without requiring RNA-seq data, linking distal regulatory elements to their target gene outputs.

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.

Limited slotsGet a Free AI Consultation

How We Work

Custom AI workflows for your Business

One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.