Skip to content
Other

SpongeScope v1

a fast AI model for predicting lncRNA–miRNA interactions

2 min readupdated

This compact model from sawcheet predicts whether a long non-coding RNA (lncRNA) interacts with a microRNA (miRNA). Under the hood, a lightweight 697,857-parameter architecture uses cross-attention (a mechanism that aligns context across two inputs) between candidate binding-site context and the miRNA sequence. Unlike traditional methods that often inflate accuracy by relying on simple sequence matches, the model was rigorously evaluated on hard negatives and unseen RNA clusters. It runs on a CPU in roughly 5 ms per pair once the lncRNA embedding is cached, making it practical for bioinformaticians shortlisting candidates ahead of wet-lab experiments.

Key features

  • Compact footprint: just 697,857 parameters and ~5 ms CPU inference per pair with a cached lncRNA representation.
  • Strict cold-start validation: test lncRNAs do not overlap with training clusters under MMseqs2 (80% sequence identity threshold), and 10% of miRNA families are completely held out from training.
  • Honest dataset: trained on 15,962 pairs from ENCORI CLIP alongside hard negatives (an ~8:1 ratio, half of which share a 7mer-site match without actual CLIP support).
  • Transparent limitations: the author explicitly documented negative experiments—adding tissue expression data and individual site labels did not improve pair-level predictions.

What you can do

  • Rank lncRNA–miRNA pairs to prioritize wet-lab validation (testing the top 10 candidates yields roughly 2.7 true interactions compared to 1 with random selection).
  • Use the built-in deterministic splits and bootstrap-interval evaluation module as a standardized benchmark for your own RNA-interaction models.
  • Run large-scale candidate interaction screening on a standard CPU without renting GPU compute.

Example

The model card does not include an inference snippet, but outlines the evaluation workflow: the model takes a long non-coding RNA (from GENCODE 50 transcripts) paired with a mature microRNA sequence (from miRBase), computes an interaction score via cross-attention over candidate binding sites, and outputs a ranked list of pairs. On the cold lncRNA split, it reaches an AUROC of 0.841 and precision@10 = 0.274.

Read moreDetails
Test metrics: cold lncRNA — AUROC 0.841 (AUPR 0.303, ECE 0.009); seed-controlled — AUROC 0.753 (AUPR 0.312); cold lncRNA + miRNA — AUROC 0.853 (AUPR 0.297).
Baseline comparison: gradient boosting on the same features scores 0.824 AUROC; cross-attention adds roughly 0.025 AUPR (though no statistically significant AUPR gain was observed across the 149 test lncRNAs).
Hardware: runs comfortably on a standard CPU (~5 ms per pair after caching the lncRNA embedding).
Limitations: intended solely for ranking candidates prior to experiments; it does not replace wet-lab validation and is not calibrated for clinical diagnostics.
Discuss Telegram 0 29

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 0.0 GB · ready files0.5 GB
  • BF16weights 0.0 GB0.5 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · F32 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # sawcheet/spongescope
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download sawcheet/spongescope --local-dir ~/models/spongescope
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to spongescope
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of spongescope (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

sawcheet/spongescope
Author
sawcheet
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

290

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments