One-Hot Encoding

Each nucleotide is mapped to a unique binary vector of length four: A = [1,0,0,0], C = [0,1,0,0], G = [0,0,1,0], and T = [0,0,0,1]. This creates an orthogonal basis where all bases are equidistant in vector space.

• No ordinal relationships: The encoding prevents the model from falsely assuming 'A' is closer to 'C' than 'G', a critical failure mode of integer encoding.
• Sparse input tensor: A 1,000-base-pair sequence becomes a [4, 1000] matrix, where exactly one bit is 'hot' per column.
• Ambiguity handling: IUPAC ambiguity codes (e.g., 'N' for any base) are represented as [0.25, 0.25, 0.25, 0.25], distributing probability mass uniformly.

DNA is double-stranded and binding events are strand-agnostic. One-hot encoding enables a precise mathematical operation to enforce this symmetry.

• Reverse complement operation: The input tensor is flipped along the sequence axis, and the channel order is permuted: A↔T and C↔G.
• Data augmentation: Training samples are randomly reverse-complemented to teach the model strand invariance without doubling the dataset size.
• Prediction averaging: At inference, predictions from the forward and reverse-complement sequences are averaged to produce a strand-symmetric output, a standard practice in tools like DeepBind and BPNet.

The one-hot structure allows the first-layer convolutional filters to be directly interpreted as Position Weight Matrices (PWMs).

• Filter visualization: A learned [4, filter_width] weight matrix can be plotted as a sequence logo, revealing the binding motif a neuron has learned to detect.
• Activation maps: Scanning the filter across the one-hot input produces an activation track that corresponds to a motif scanning score, functionally equivalent to a PWM scan.
• TF-MoDISco: This interpretability tool clusters high-attribution subsequences from one-hot inputs to extract consolidated, non-redundant motifs learned by the network.

The representation is highly sparse (75% zeros), which has significant implications for storage and compute.

• Memory inefficiency: Storing a full 32-bit float tensor for a 3-billion-base-pair human genome requires approximately 48 GB of memory, which is prohibitive.
• Embedding layer optimization: In practice, a trainable embedding layer maps integer-encoded bases (A=0, C=1, G=2, T=3) to dense vectors, learning a continuous representation rather than using fixed one-hot vectors.
• Mixed precision: When one-hot is used directly, storing the tensor in float16 or bfloat16 halves the memory footprint without impacting model accuracy.

One-hot encoding enables a systematic computational perturbation assay to identify causal regulatory variants.

• Saturation mutagenesis: Every position in the input sequence is systematically mutated to all three alternative bases, and the change in the model's prediction is recorded.
• Attribution scores: The difference between the reference prediction and the mutated prediction quantifies the functional impact of a single nucleotide variant.
• Integrated Gradients: This interpretability method computes feature attributions by interpolating between a zero baseline and the one-hot input, accumulating gradients along the path to identify bases critical for binding.

One-hot encoded sequence serves as the primary input, but it is often concatenated with additional channels representing epigenomic data.

• Multi-channel input: Beyond the 4 nucleotide channels, additional tracks for chromatin accessibility, histone marks, or evolutionary conservation scores are stacked to create a [C, L] tensor.
• Unified representation: Models like DeepSEA and Enformer consume this multi-channel input to simultaneously predict hundreds of epigenetic assays from raw sequence.
• Distance encoding: For models predicting 3D genome folding, one-hot sequence is combined with a genomic distance channel to provide the linear separation between loci.

The sequence obtained by reversing the order of a DNA strand and swapping each base with its Watson-Crick pairing partner (A↔T, C↔G). This is a critical data augmentation technique for enforcing strand symmetry in genomic models. Because DNA is double-stranded and sequencing reads can originate from either strand, a model must recognize that a binding site is identical to its reverse complement. During training, one-hot encoded sequences are randomly reverse-complemented to double the dataset size and ensure the model learns strand-agnostic representations.

The broader category of vectorization strategies that convert raw nucleotide sequences into dense, learnable feature vectors. While one-hot encoding produces a fixed, sparse, binary matrix, modern embedding layers learn a continuous, lower-dimensional representation directly from data. These embeddings are initialized randomly and updated via backpropagation, allowing the model to discover semantic relationships between k-mers—similar to word embeddings in natural language processing. This is the first layer in architectures like DNABERT and Enformer.

A statistical matrix representing the log-odds probability of each nucleotide occurring at each position within a collection of aligned binding sites. A PWM is the classical, interpretable alternative to a neural network's learned filters. It is used to scan genomes for motif matches by calculating a score for every overlapping k-mer. In deep learning, the first convolutional layer of a model like DeepBind often learns filters that strongly resemble PWMs, demonstrating that neural networks can autonomously rediscover known binding motifs from one-hot encoded input.

A cross-validation strategy for genomic deep learning where entire chromosomes are reserved for testing. This prevents information leakage caused by sequence homology between training and validation splits. If sequences from the same gene family are split randomly, a model might memorize a motif in training and appear to generalize in validation, inflating performance metrics. By holding out chromosomes 8 and 9, for example, the model is tested on truly unseen genomic contexts, providing a rigorous evaluation of its ability to generalize from one-hot encoded sequences.

A convolutional operation where the filter kernel is applied over an input with defined gaps or holes, exponentially expanding the receptive field without increasing parameter count. In genomic models processing one-hot encoded sequences, standard convolutions capture only local motifs. Dilated convolutions allow the network to integrate information across hundreds or thousands of base pairs—critical for modeling enhancer-promoter interactions. Architectures like BPNet use dilated convolutions to predict base-resolution binding profiles while accounting for long-range regulatory grammar.

A computational normalization procedure that models and removes the systematic dependence of sequencing read coverage on local guanine-cytosine content. This bias arises from PCR amplification artifacts during library preparation, where fragments with extreme GC content amplify less efficiently. If uncorrected, a model trained on one-hot encoded sequences may learn to associate GC-rich regions with low signal, conflating biochemical reality with technical artifact. Correction ensures the neural network learns true sequence-binding relationships rather than library preparation biases.