Localizing predictive concepts encoded in neural representations via hierarchical testing
Organizations: Max Delbrück Center for Molecular Medicine, Berlin, Germany · Humboldt-Universität zu Berlin, Germany · Department of Artificial Intelligence, Fraunhofer Heinrich Hertz Institute, Berlin, Germany · BIFOLD – Berlin Institute for the Foundations of Learning and Data, Germany ·
Abstract
Understanding which regulatory sequence motifs a genomic deep learning model has learned, with statistical error guarantees, is essential for trusting its use in downstream tasks such as variant effect prediction. Existing approaches to identifying which units in a trained network encode human- understandable concepts fall into two groups. Exploratory methods rank and visualize units but offer no statistical guarantee that the selected units are associated with the prediction target across the population of inputs. Testing-based methods test only single units or one pre-defined set of them, and do not control the error across the many sets of units. We introduce CUTTIE (Cluster-tree Unit Top-down Testing of Independence with Error control), which clusters the units of each layer into a tree and tests the clusters top-down with conditional independence tests, localizing information about the target while controlling the family-wise error rate. Thereby, we identify clusters that carry information about the target beyond the rest of the layer. For the units in the identified clusters, conditional attribution maps then reveal what motifs they encode. In a controlled simulation, the procedure identifies the clusters whose concepts affect the target. We apply the procedure to a convolutional model predicting cell-type-specific chromatin accessibility from DNA sequence, where the concepts encoded in the identified clusters map onto known regulatory motifs.
Figures & tables
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
AttnGen: Attention-Guided Saliency Learning for Interpretable Genomic Sequence Classification
Causal dictionary learning reveals and validates transcription-factor binding features in genomic language models
concept'' inside a model is real rather than an artifact of sequence composition. We introduce a framework that combines sparse dictionary learning with causal intervention to extract, validate, and causally test interpretable features in genomic foundation models. Training top-$k$ sparse autoencoders on the hidden activations of two architecturally distinct models, Nucleotide Transformer ($6$-mer tokenization) and DNABERT-2 (byte-pair encoding), we recover thousands of monosemantic features that map to transcription-factor (TF) sequence motifs. We show that the naive validation of such features against position weight matrices is severely confounded by GC composition and repetitive elements, producing hundreds of spurious TF features'', and we develop a composition-matched, binding-resolved protocol that removes these confounds. Critically, we move beyond correlation: by ablating individual dictionary directions during the model's forward pass and measuring the induced shift in the model's own predictive distribution, we establish that specific features are causally used to represent cell-type-specific TF binding, not merely motif presence. Across three transcription factors (CTCF, GATA1, REST) and both architectures, causally validated binding features emerge reproducibly (-- of tested features per condition), while two classes of negative control, scrambled binding labels and randomly selected features, yield no detectable signal. The framework is purely computational, uses only public data, and provides a reusable standard for interpretability claims in genomic deep learning.