README.md

---
tags:
- rna
- sirna
- bioinformatics
- drug-discovery
- pytorch
- onnx
- webgpu
- research
library_name: pytorch
---

# KDMSi: Base3 and Base0 siRNA design models

KDMSi is a staged computational workflow for prioritizing unmodified siRNA target sites and ranking chemically modified siRNA duplexes. This repository contains the released KDMSi-Base3 and KDMSi-Base0 checkpoints, browser-compatible ONNX exports, the RNA-FM backbone used to construct residue representations, and the preprocessing assets required by the KDMSi web application.

- Interactive Space: [leighv318/KDMSi-webapp](https://huggingface.co/spaces/leighv318/KDMSi-webapp)
- Source code: [leighv318/KDMSi](https://github.com/leighv318/KDMSi)

## Model components

### KDMSi-Base3

`Base3_cross_access_v2_src_no_se` ranks unmodified siRNA candidates. It combines guide and local target-mRNA RNA-FM representations, recurrent shared-weight Transformer encoders, guide-to-mRNA cross-attention, handcrafted sequence descriptors, thermodynamic/accessibility features, and source information.

The standard inference release is a ten-fold ensemble:

- PyTorch: `base3/pytorch/0-inter/` through `base3/pytorch/9-inter/`
- ONNX: `base3/onnx/0-inter.onnx` through `base3/onnx/9-inter.onnx`

The `base3/pytorch/T/` and `base3/pytorch/R/` directories contain the Takayuki and Reynolds source-held-out development checkpoints. They are evaluation artifacts and are not part of the standard ten-fold inference ensemble.

### KDMSi-Base0

`Base0_mrna_full_no_se` ranks chemically modified antisense/sense duplexes using position-resolved nucleotide and chemical representations, local mRNA context, concentration, cell type, and biophysical descriptors.

The standard inference release is a ten-fold ensemble:

- PyTorch: `base0/pytorch/0-inter/` through `base0/pytorch/9-inter/`
- ONNX: `base0/onnx/0-inter.onnx` through `base0/onnx/9-inter.onnx`

Gene-named directories under `base0/pytorch/` contain leave-one-gene-out development checkpoints. They are provided for reproducibility and are not used by the standard web ensemble.

## Shared assets

- `shared/rnafm/RNA-FM_pretrained.pth` — RNA-FM PyTorch weights used to produce 640-dimensional residue representations.
- `shared/rnafm/rnafm_t12.onnx` — browser-compatible RNA-FM export used by the KDMSi Space.
- `shared/preprocessing/base3_pssm.json` — fold-specific Base3 PSSM data.
- `shared/preprocessing/base0_mod_features.json` — Base0 chemical representation lookup data.
- `shared/preprocessing/base0_fold_context.json` — fold-specific Base0 concentration and categorical context.
- `shared/preprocessing/mod_distributions.json` — empirical position-specific modification distributions used by the candidate sampler.
- `shared/runtime/vrna.wasm` — ViennaRNA WebAssembly runtime used to calculate browser-side structure and thermodynamic features.
- `manifest.json` — machine-readable release layout.
- `checksums.sha256` — SHA-256 integrity hashes for every released file.

RNA-FM was developed by the [ml4bio/RNA-FM](https://github.com/ml4bio/RNA-FM) authors. Its license is reproduced at `shared/rnafm/RNA-FM_LICENSE.txt`. Please cite the original RNA-FM publication when using these representations.

## Download

Download the complete release with `huggingface_hub`:

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="leighv318/KDMSi",
    repo_type="model",
)
print(local_dir)
```

To download only the browser inference assets:

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="leighv318/KDMSi",
    allow_patterns=[
        "base3/onnx/*",
        "base0/onnx/*",
        "shared/rnafm/rnafm_t12.onnx",
        "shared/preprocessing/*",
        "shared/runtime/vrna.wasm",
        "manifest.json",
    ],
)
```

The checkpoints use custom KDMSi model classes and feature construction; they are not directly loadable with `transformers.AutoModel`. Use the KDMSi source repository or the browser application for complete preprocessing and inference.

## Intended use

The models are intended for research-use prioritization of siRNA candidates before experimental validation. Base3 is used for unmodified target-site ranking; Base0 is used for ranking chemically modified duplexes under the represented molecular and assay context.

## Limitations

- Predicted scores do not replace experimental validation.
- Base0 supports only the chemical-modification vocabulary represented by its released encoder and training data.
- The standard Base0 score estimates inhibition under modeled sequence, chemistry, transcript, concentration, and cell-type conditions. It does not predict delivery, pharmacokinetics, toxicity, or clinical efficacy.
- Optional off-target annotation is implemented in the KDMSi workflow rather than these checkpoint files. It is sequence based and does not model modification-dependent changes in off-target activity.
- Base3 source-held-out results using adaptive batch normalization should be interpreted as transductive source-adaptation estimates rather than strict frozen zero-shot estimates.
- Sample-level cross-validation is a development/interpolation benchmark and should not be conflated with leave-one-gene-out or group-disjoint external evaluation.

## Release and licensing note

The KDMSi project license has not yet been finalized. Public availability of the files should not be interpreted as permission for clinical use. The bundled RNA-FM license applies to the RNA-FM component and is included separately. Contact the KDMSi authors regarding reuse outside research evaluation.

## Citation

The KDMSi manuscript citation will be added after publication. Until then, please cite the repository and the original RNA-FM work when using these files.