pore-codec-rsq742c36a-519

PoreCodec-RSQ742C36A

PoreCodec-RSQ742C36A is a lightweight neural codec for nanopore electrical signals. It converts continuous ionic current measurements into discrete token sequences while preserving the original temporal resolution of the signal. The model combines a fully convolutional dilated encoder with a production-oriented Residual Finite Scalar Quantization (Residual FSQ) module to generate hierarchical discrete representations suitable for downstream foundation models.

Unlike conventional neural codecs that compress the temporal dimension, PoreCodec maintains a one-to-one correspondence between signal samples and latent tokens. Each input sample produces one discrete token, enabling accurate temporal alignment for biological sequence modeling, retrieval, and representation learning.

Instead of learned vector codebooks, the codec employs deterministic finite scalar quantization with exhaustive codebooks stored directly inside the model checkpoint, providing stable optimization, deterministic encoding, and constant-time decoding.

---

Highlights

• One token generated for every input signal sample
• Full-resolution temporal representation
• Five-layer dilated convolutional encoder and decoder
• Residual Finite Scalar Quantization (Residual FSQ)
• Exhaustive codebooks stored inside safetensors
• Constant-time O(1) decoding lookup
• Hierarchical discrete token generation
• HuggingFace AutoModel compatible
• End-to-end signal reconstruction
• Foundation-model ready discrete representation

---

Architecture

Raw Nanopore Signal
        │
        ▼
 Dilated CNN Encoder
        │
        ▼
512-D Latent Features
        │
        ▼
Linear Projection
        │
        ▼
Residual Finite Scalar Quantization
        │
        ├──────────────► Hierarchical Token IDs
        │
        ▼
Quantized Latent Features
        │
        ▼
Linear Projection
        │
        ▼
 Dilated CNN Decoder
        │
        ▼
Reconstructed Signal

The model consists of three major components.

---

Dilated CNN Encoder

The encoder extracts local electrical signal patterns using a five-layer one-dimensional dilated convolutional network.

Unlike conventional neural codecs, the encoder does not perform temporal downsampling. Progressive dilation enlarges the receptive field while preserving the original sequence length, allowing every input signal sample to correspond to one latent representation.

---

Residual Finite Scalar Quantization

The latent representation is discretized using a production-oriented Residual Finite Scalar Quantization (Residual FSQ) module.

Instead of learning vector embeddings as in conventional VQ-VAE methods, PoreCodec performs deterministic scalar quantization over a finite Cartesian lattice.

Each residual stage progressively quantizes the remaining reconstruction error, producing increasingly accurate hierarchical representations.

Compared with learned vector quantizers, Residual FSQ provides:

• Deterministic encoding
• Stable optimization
• No codebook collapse
• No dead entries
• Constant-time decoding
• Exact lookup reconstruction

All exhaustive codebooks are precomputed and stored directly inside the model checkpoint.

Quantization Configuration

---

Hierarchical Tokens

Each latent position produces hierarchical residual quantization indices.

Latent Vector

↓

Residual Quantizer 1

↓

Residual Quantizer 2

↓

Residual Indices

↓

Unified Token ID

Supported hierarchy levels:

Since the encoder preserves the original temporal resolution, the number of generated tokens is identical to the number of input signal samples.

---

Signal Preprocessing

The accompanying FeatureExtractor automatically preprocesses raw nanopore electrical signals before inference.

Pipeline:

1. Physical boundary correction
2. Spike removal
3. Median-MAD normalization
4. Optional median filtering
5. Smooth nonlinear clipping

Supported preprocessing strategies:

apple (default)

• Error repair
• Spike removal
• Median-MAD normalization
• Median filtering
• Smooth clipping

mongo

• Error repair
• Spike removal
• Median-MAD normalization
• Smooth clipping

---

Quick Start

import numpy as np
from transformers import AutoFeatureExtractor, AutoModel

model_name = "ShuaiAnwo/pore-codec-rsq742c36a-512"

feature_extractor = AutoFeatureExtractor.from_pretrained(
    model_name,
    trust_remote_code=True,
)

model = AutoModel.from_pretrained(
    model_name,
    trust_remote_code=True,
)

model.eval()

raw_signal = np.random.normal(
    loc=70,
    scale=8,
    size=1855,
).astype(np.float32)

signal = feature_extractor(
    raw_signal,
    return_tensors="pt",
)["signal"]

token_ids = model.encode_signal(
    signal,
    layer=2,
)

reconstructed = model.decode_token(
    token_ids,
    layer=2,
)

print(signal.shape)
print(token_ids.shape)
print(reconstructed.shape)

Expected output

torch.Size([1, 1, 1855])
torch.Size([1, 1855])
torch.Size([1, 1, 1855])

---

API Reference

encode_signal

Convert a normalized nanopore signal into unified discrete token IDs.

token_ids = model.encode_signal(
    signal,
    layer=2,
)

Returns

[B, N]

---

decode_token

Reconstruct normalized nanopore signals directly from unified token IDs.

signal = model.decode_token(
    token_ids,
    layer=2,
)

---

tokenize_indices

Convert hierarchical residual quantization indices into unified token IDs.

token_ids = model.tokenize_indices(
    level_indices,
    layer=2,
)

---

decode_indices

Reconstruct signals directly from hierarchical residual indices.

signal = model.decode_indices(
    level_indices,
    layer=2,
)

---

forward

Perform complete end-to-end encoding and decoding.

reconstruction, level_indices = model(signal)

Returns:

• Reconstructed signal
• Hierarchical residual indices

---

Temporal Alignment

Unlike conventional neural codecs that reduce temporal resolution, PoreCodec preserves the original signal length throughout the encoding pipeline.

For an input signal of length N, the model produces exactly N discrete tokens.

This one-to-one temporal correspondence simplifies downstream integration with biological foundation models, retrieval systems, sequence alignment algorithms, and multimodal learning pipelines.

---

Validation

The released checkpoint has been validated using the provided evaluation pipeline.

---

Model Configuration

---

Applications

PoreCodec is suitable for:

• Nanopore signal tokenization
• Biological signal representation learning
• Neural signal compression
• Nanopore foundation models
• Token-based biological language modeling
• Retrieval and indexing of nanopore reads
• Generative modeling over nanopore electrical signals
• Signal-to-token interfaces for multimodal biological AI

---

Advantages over Conventional VQ

---

Limitations

This model is designed for preprocessed nanopore electrical current signals using the accompanying FeatureExtractor.

The model produces discrete signal representations and signal reconstructions. It is not intended to perform direct DNA basecalling or nucleotide sequence prediction.

---

Citation

If you use PoreCodec in your research, please cite:

@misc{jiao2026porecodec,
  title={PoreCodec: Residual Finite Scalar Quantization for Full-Resolution Nanopore Signal Tokenization},
  author={Shuai Jiao},
  year={2026}
}

---

License

This model is released under the Apache-2.0 License.