Maximally Activating Sequences

The foundational feature visualization technique that synthesizes an input sequence to maximally activate a target neuron. Starting from random nucleotides, it iteratively optimizes the input via gradient ascent in the input space. For genomic models, this often reveals the preferred sequence motif a convolutional filter or attention head has learned to detect. Regularization terms, such as L2 decay or total variation, are critical to prevent the generation of unrealistic, high-frequency nucleotide patterns that do not correspond to biological sequences.

A systematic perturbation-based method for validating MAS. ISM computationally mutates every nucleotide in a given sequence and measures the change in the model's prediction or neuron activation. This produces a mutation sensitivity map that identifies which base pairs are critical for function. When applied to a maximally activating sequence, ISM confirms whether the high-activation pattern is robust to single-nucleotide variants, distinguishing a genuine biological motif from a brittle adversarial artifact.

Transcription Factor Motif Discovery from Importance Scores is a method that clusters high-contribution genomic subsequences identified by attribution maps into recurring, biologically meaningful patterns. It bridges the gap between raw per-nucleotide importance scores and the sequence logos biologists expect. By applying TF-MoDISco to a large set of sequences, one can recover the distinct motifs that drive activation in different model channels, providing a direct comparison to known transcription factor binding sites from databases like JASPAR.

A graphical representation of a conserved nucleotide motif where the height of each letter is proportional to its information content (measured in bits). Sequence logos are the standard visual language for presenting the results of a MAS analysis. The x-axis represents sequence position, and stacked letters indicate the relative frequency of A, C, G, and T. A tall, single letter indicates a highly conserved position, while a short stack of multiple letters indicates degeneracy. Tools like Logomaker generate publication-ready logos directly from position weight matrices.

A suite of diagnostic tests designed to verify that a feature visualization or attribution method is genuinely sensitive to the model's learned parameters. The core test is model parameter randomization: the weights of the trained model are progressively randomized from the top layer down. A valid method must produce different MAS or attribution maps for the randomized model compared to the trained one. If the synthesized sequence remains unchanged, the method is an edge detector independent of the model's training, and its results are not trustworthy.

The quantitative difference in a model's prediction score between a reference allele and an alternate allele. In the context of MAS, delta scores are used to measure the functional impact of introducing a single-nucleotide variant into a maximally activating sequence. A large negative delta score indicates the variant disrupts the motif and silences the neuron. This metric is the foundation of variant effect prediction, linking the abstract concept of a synthetic MAS to the clinically relevant task of scoring pathogenic mutations.