Activation Maximization

The core mechanism involves starting from random noise and iteratively updating the input to maximize a neuron's activation. Gradient ascent is performed on the input space, not the model weights. The model parameters are frozen while the input nucleotides are treated as learnable parameters. This process reveals the preferred stimulus of any neuron in a genomic network, effectively answering: 'What sequence pattern makes this filter fire most strongly?'

Unconstrained optimization often produces adversarial noise patterns that are unrecognizable. To synthesize interpretable genomic sequences, strong regularization is essential:

• L2 decay penalizes extreme nucleotide values
• Total variation penalizes high-frequency, non-biological transitions
• Jitter and rotation prevent overfitting to specific positional coordinates
• Frequency-domain constraints enforce realistic k-mer usage patterns These priors ensure the resulting sequence resembles a biologically plausible motif.

The output of this technique is a Maximally Activating Sequence (MAS)—a synthetic or naturally occurring genomic string that causes a specific filter or neuron to fire most strongly. In genomic models, these sequences often correspond to transcription factor binding motifs, splice sites, or conserved regulatory elements. By visualizing the MAS for each filter in a convolutional layer, researchers can build a complete dictionary of the sequence features the model has learned.

Beyond single-neuron maximization, the technique extends to:

• Activation Maximization by Direction: Optimizing inputs that align with a specific direction in activation space, revealing features that distinguish two classes of genomic elements
• Class Model Visualization: Maximizing the logit score for a specific output class (e.g., 'enhancer' vs. 'promoter') to generate a prototypical sequence for that class
• Diversity-driven optimization: Using a diversity term to prevent multiple runs from converging to the same motif, uncovering the full repertoire of features a neuron responds to.

Activation Maximization is a generative interpretability technique, distinct from attribution methods like Integrated Gradients or DeepLIFT. While attribution methods ask 'Which parts of this specific input were important?', Activation Maximization asks 'What is the ideal input for this neuron?' The two approaches are complementary: attribution explains individual predictions, while activation maximization characterizes the general feature detectors learned by the model. Together, they provide a complete picture of model behavior.

The biological validity of synthesized sequences is validated by:

• Motif database comparison: Aligning MAS against JASPAR, HOCOMOCO, or TRANSFAC databases to confirm correspondence with known transcription factor binding sites
• In-silico mutagenesis (ISM): Systematically mutating each position in the MAS to verify the predicted importance pattern matches experimental data
• Deep Mutational Scan (DMS) correlation: Comparing the model's sensitivity profile around the MAS with high-throughput experimental measurements of variant effects
• Cross-species conservation: Checking if the synthesized motif is conserved across orthologous genomic regions.

Activation maximization synthesizes sequences from scratch to reveal the preferred binding motifs of a neuron. Starting from random nucleotides, gradient ascent iteratively modifies the sequence to maximize a neuron's activation. The resulting maximally activating sequences often converge to consensus motifs that recapitulate known transcription factor binding sites, such as the TATA box or p53 response elements, without any prior biological knowledge. This technique has rediscovered canonical motifs and identified novel, uncharacterized patterns in deep learning models trained on ChIP-seq data.

In genomic convolutional neural networks, activation maximization visualizes what each first-layer convolutional filter detects. Early filters typically learn to recognize short, fundamental patterns:

• GC-rich regions associated with CpG islands
• Periodic dinucleotide patterns reflecting DNA structural properties
• Splice donor/acceptor sites with conserved GT/AG dinucleotides Deeper layer neurons combine these primitives to detect higher-order features like helix-turn-helix motifs or composite regulatory modules spanning 20-50 base pairs.

Activation maximization can probe individual neurons in recurrent or transformer layers of genomic language models to understand their specialized functions. A single neuron in a DNA language model might encode:

• Positional specificity relative to the transcription start site
• Strand orientation preferences for asymmetric motifs
• Cooperative binding logic where two motifs must appear with a specific spacing By synthesizing sequences that maximize that neuron, researchers decode the precise cis-regulatory grammar the model has internalized during self-supervised pre-training on reference genomes.

Instead of targeting a single neuron, activation maximization can be applied to the output logit for a specific genomic class. This generates the idealized input sequence the model associates most strongly with that phenotype:

• Enhancer vs. promoter classification: reveals distinguishing k-mer frequencies
• Disease-associated variant prediction: synthesizes sequences that maximize pathogenicity scores
• Cell-type-specific accessibility: generates sequences predicted to be open in HepG2 but not K562 cells The resulting synthetic sequences serve as prototypical examples of what the model considers the most definitive instance of each class.

Unconstrained activation maximization often produces adversarial-looking sequences with unrealistic nucleotide compositions. To ensure biological relevance, optimization is regularized with:

• Entropy penalties that encourage sharp, position-specific nucleotide preferences
• GC content constraints matching genomic background frequencies (typically ~40-42%)
• Total variation denoising to smooth adjacent nucleotide probabilities
• Sequence priors from a generative model of real genomic DNA These constraints force the synthesized sequence to remain within the manifold of realistic genomic sequences while still maximizing the target activation.

Motifs discovered through activation maximization are validated against experimental ground truth to confirm biological relevance:

• ChIP-seq peak summits: synthesized motifs should match the central enriched k-mers
• HT-SELEX data: high-throughput in vitro binding assays provide independent motif validation
• Deep Mutational Scanning (DMS) : the impact of single-nucleotide changes on model activation should correlate with measured binding affinity changes
• Cross-species conservation: maximally activating motifs should show elevated phyloP or phastCons scores Strong concordance between computationally synthesized motifs and experimental data builds regulatory confidence in the model's learned representations.