Glossary

TF-MoDISco

TF-MoDISco is a method that clusters high-contribution genomic subsequences identified by attribution maps into recurring, biologically meaningful motif patterns.

Get in touch Learn more

ML engineer managing model training cluster on laptop, GPU utilization visible, technical deep learning setup.

MOTIF DISCOVERY FROM IMPORTANCE SCORES

What is TF-MoDISco?

A computational method that transforms scattered nucleotide-level attribution maps into consolidated, high-resolution sequence motifs representing the recurring patterns a deep learning model has learned to recognize.

TF-MoDISco (Transcription Factor Motif Discovery from Importance Scores) is a multi-stage algorithm that clusters high-contribution genomic subsequences identified by feature attribution maps into biologically meaningful motif patterns. It operates on per-nucleotide importance scores—typically generated by methods like DeepLIFT or Integrated Gradients—and groups seqlets (short, high-scoring sequence segments) using affinity propagation based on their sequence similarity and spatial alignment.

The output is a set of position weight matrices visualized as sequence logos, where each cluster represents a distinct binding preference learned by the model. By collapsing thousands of individual high-attribution regions into a small number of consolidated motifs, TF-MoDISco bridges the gap between opaque neural network predictions and human-interpretable transcription factor binding site biology, enabling rigorous validation of genomic deep learning models.

MOTIF DISCOVERY

Key Features of TF-MoDISco

TF-MoDISco distills massive, noisy attribution maps into a small set of high-quality, recurring sequence motifs. It bridges the gap between per-nucleotide importance scores and biologically meaningful patterns recognized by transcription factors.

Contribution-Weighted Clustering

TF-MoDISco clusters high-scoring genomic subsequences based on their contribution scores from attribution maps. It uses DeepLIFT or Integrated Gradients scores to weight sequences, ensuring that the discovered patterns directly reflect the model's learned decision logic rather than just statistical over-representation.

Seqlet Identification

The algorithm first identifies seqlets—short, high-contribution genomic subsequences—from the attribution map. These seqlets serve as the raw material for motif discovery:

Extracted from regions with high attribution scores
Typically span 10-30 base pairs
Represent the model's focal points within a longer input sequence

Multi-Stage Refinement Pipeline

TF-MoDISco employs a multi-stage clustering and refinement process to produce non-redundant motifs:

Affinity Propagation groups similar seqlets
Trimming removes low-contribution flanks
Merging combines highly similar clusters
Rediscarding eliminates statistically insignificant patterns This yields a concise set of distinct, high-confidence motifs.

Synthetic Motif Generation

Each discovered cluster is aggregated into a Position Weight Matrix (PWM) and visualized as a sequence logo. TF-MoDISco generates a synthetic representation of the motif by averaging the aligned seqlets, providing a human-interpretable summary of the pattern the model has learned to recognize.

Hypothesis-Driven Discovery

TF-MoDISco can be run in a hypothesis-driven mode by restricting seqlet extraction to regions associated with a specific prediction class (e.g., high vs. low expression). This allows researchers to isolate motifs that are causally linked to a particular biological outcome, moving beyond correlation to mechanistic insight.

Continuous Score Integration

Unlike binary peak-calling methods, TF-MoDISco integrates continuous attribution scores throughout the clustering process. The contribution weight of each nucleotide influences seqlet alignment and motif construction, preserving the nuanced importance gradients that binary thresholding would discard.

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

TF-MODISCO EXPLAINED

Frequently Asked Questions

Clear, technical answers to the most common questions about TF-MoDISco, the algorithm that transforms genomic attribution maps into biologically meaningful sequence motifs.

TF-MoDISco (Transcription Factor Motif Discovery from Importance Scores) is a post-hoc interpretability algorithm that clusters high-contribution genomic subsequences identified by feature attribution maps into recurring, biologically meaningful motif patterns. It operates in three stages: first, it extracts seqlets—short genomic subsequences with high attribution scores—from per-nucleotide importance maps generated by methods like DeepLIFT or Integrated Gradients. Second, it embeds these seqlets using their attribution-weighted sequence content and clusters them using affinity propagation or density-based spatial clustering. Third, it aligns the seqlets within each cluster and generates a sequence logo representing a consolidated motif. The output is a set of non-redundant, human-interpretable motifs that directly correspond to the decision logic learned by the deep genomic model, bridging the gap between opaque neural network predictions and established biological knowledge of transcription factor binding sites.

INTERPRETABILITY ECOSYSTEM

Related Terms

TF-MoDISco operates within a broader landscape of feature attribution and motif discovery methods. These related techniques provide the foundational importance scores that TF-MoDISco clusters, or serve as alternative approaches for decoding genomic model logic.

DeepLIFT

A backpropagation-based attribution algorithm that computes importance scores by comparing neuron activations to a reference state. DeepLIFT uses the rescale rule and revealcancel rule to handle non-linearities, efficiently assigning contribution scores to individual nucleotides. It is a common upstream input for TF-MoDISco's clustering pipeline because it satisfies the summation-to-delta property, meaning the sum of all feature contributions equals the difference in output from the reference.

Summation-to-Delta

Key Property

Integrated Gradients

An axiomatic attribution method that computes the path integral of gradients from a baseline input to the actual input. It is the only method that satisfies the completeness axiom (attributions sum to the output difference) while also satisfying sensitivity and implementation invariance. In genomics, a zero-vector or shuffled sequence often serves as the baseline. The resulting nucleotide-level importance scores can be fed directly into TF-MoDISco for motif discovery.

Axiomatic

Theoretical Foundation

In-Silico Mutagenesis (ISM)

A systematic perturbation technique that computationally mutates every nucleotide in a sequence to quantify its impact on model predictions. By measuring the delta score between reference and alternate alleles at each position, ISM produces a high-resolution importance map. These maps are often used as input to TF-MoDISco, particularly when the goal is to identify motifs associated with variant effect prediction.

Per-Nucleotide

Resolution

SHAP

A unified framework based on Shapley values from cooperative game theory that assigns each genomic feature an importance score for a particular prediction. SHAP satisfies three desirable properties: local accuracy, missingness, and consistency. While exact Shapley values are computationally expensive for genomic models, approximations like DeepSHAP and KernelSHAP provide practical alternatives whose outputs can be clustered by TF-MoDISco.

Game Theory

Mathematical Basis

Sequence Logos

A graphical representation of a conserved nucleotide motif where the height of each letter is proportional to its information content. Sequence logos are the standard visualization for motifs discovered by TF-MoDISco. Each position in the logo shows the relative frequency of A, C, G, and T, with the total height indicating conservation. They enable direct comparison with known transcription factor binding motifs from databases like JASPAR and TRANSFAC.

Information Content

Measured in Bits

Attribution Sanity Checks

A suite of tests designed to verify that an attribution method is sensitive to the learned parameters of the genomic model. Key checks include model parameter randomization (attributions should change when model weights are scrambled) and data randomization (attributions should differ when trained on shuffled labels). These sanity checks are critical before using any attribution method as input to TF-MoDISco, ensuring the discovered motifs reflect genuine model logic.

Pre-Clustering

Validation Stage

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.