SpectralBio

Spectral Covariance Analysis of Protein Language Model Hidden States for Zero-Shot Variant Pathogenicity Prediction

Paper | Quickstart | Experiments | Public API

Claw (AI Co-author) | Davi Bonetto

---

Abstract

SpectralBio is a reproducible, zero-shot variant-effect pipeline for missense pathogenicity prediction from sequence alone. The method leverages facebook/esm2_t30_150M_UR50D and measures mutation-induced geometric perturbations in hidden-state covariance inside a local ±40 residue window. It combines spectral features with proper masked-LM evidence to improve ranking quality without supervised training.

On TP53 ClinVar (N=255, 115 pathogenic, 140 benign), the best pair (0.55 * FrobDist + 0.45 * LL Proper) reaches AUC=0.7498 (~0.750). On BRCA1 transfer (N=100 subset), LL Proper reaches AUC=0.9174 (~0.917). Reproducibility is exact with fixed seeds (Δrep=0.0).

---

Key Results

Benchmark	Best Method	AUC-ROC	Interpretation
TP53 (ClinVar, N=255)	0.55*frob_dist + 0.45*ll_proper	0.7498	Best pair in ablation search
BRCA1 transfer (N=100)	ll_proper	0.9174	Strong cross-protein generalization

Main empirical takeaways

• Matrix-level covariance features (FrobDist, TraceRatio) are consistently stronger than eigenvalue-only SPS variants on TP53.
• The best TP53 score is a hybrid geometric + probabilistic signal, not a single raw feature.
• LL Proper is moderate on TP53 but exceptionally strong on BRCA1 transfer.
• Reported reproducibility check remains exact (Δrep=0.0) with deterministic seeds.

---

Method (Brief)

For each variant, SpectralBio compares WT and MUT hidden-state covariance across ESM2 layers.

$$ \mathrm{FrobDist} = \frac{1}{L}\sum_{l=1}^{L}\left|C_{\text{MUT}}^{(l)} - C_{\text{WT}}^{(l)}\right|_F $$

$$ \mathrm{TraceRatio} = \frac{1}{L}\sum_{l=1}^{L}\left|\frac{\mathrm{tr}\left(C_{\text{MUT}}^{(l)}\right)}{\mathrm{tr}\left(C_{\text{WT}}^{(l)}\right)} - 1\right| $$

$$ \mathrm{SPS\text{-}log} = \frac{1}{L}\sum_{l=1}^{L}\left|\log\left|\lambda_{\text{MUT}}^{(l)}\right| - \log\left|\lambda_{\text{WT}}^{(l)}\right|\right|_2^2 $$

Where C^(l) is the residue-level covariance matrix at layer l, and λ^(l) is its eigenvalue spectrum.

Best TP53 combination in this release:

$$ \mathrm{Score}_{\text{best}} = 0.55\cdot\mathrm{FrobDist} + 0.45\cdot\mathrm{LL\ Proper} $$

---

Method Diagram (Pipeline Discussion)

The diagram summarizes the end-to-end computational flow used in the paper:

1. Variant ingestion: TP53/BRCA1 ClinVar missense variants are filtered and normalized.
2. Sequence-local encoding: WT and MUT windows (±40) are encoded by ESM2-150M.
3. Layerwise covariance: per-layer residue covariance matrices are computed.
4. Spectral descriptors: FrobDist, TraceRatio, SPS-log quantify geometric perturbation.
5. Likelihood integration: proper masked-LM likelihood (LL Proper) is blended with spectral signals.
6. Evaluation: AUC ablations on TP53 and transfer analysis on BRCA1.

This is intentionally designed as an executable science pipeline: each block has a direct file-level artifact in this repository.

---

Paper Discussion (Deep Dive)

Why covariance geometry matters

Likelihood scores answer “how probable is this mutation under token prediction,” while covariance perturbation answers “how strongly did internal representation geometry shift.” In practice, those are complementary biological signals.

Why the TP53 best score is a mixture

On TP53, no single feature dominates all decision boundaries. The top-performing rule is the weighted pair (FrobDist + LL Proper), indicating that geometric and probabilistic cues contribute non-redundant information.

Why BRCA1 transfer is important

The BRCA1 result (AUC=0.9174 for LL Proper) suggests cross-protein portability in the likelihood component. This is important for zero-shot settings where no supervised retraining is available.

Current limitations

• Two-gene scope for this release (TP53 primary + BRCA1 transfer subset).
• Window size and feature mixing are fixed in this benchmark snapshot.
• Intended for research reproducibility, not clinical deployment.

Feature-level interpretation from the TP53 ablation

Feature	AUC-ROC (TP53)	Interpretation
ll_crude	0.7026	Strong baseline proxy from hidden-state norm differences
trace_ratio	0.6242	Useful matrix-scale perturbation signal
frob_dist	0.6209	Stable global covariance displacement measure
sps_log	0.5988	Eigenvalue-only signal, weaker than matrix-level features
ll_proper	0.5956	Modest on TP53 but strongest transfer behavior on BRCA1

This ranking is exactly why SpectralBio keeps both geometric and probabilistic channels. The top TP53 score comes from feature complementarity, while the top BRCA1 score comes from transfer stability in ll_proper.

---

Quick Start

Option A (agents): SKILL.md

• Open SKILL.md at repository root.
• Execute each step sequentially.
• Runtime target: ~24 min on T4 GPU.

Option B (humans): notebook

• Open colab/spectralbio.ipynb.
• Run cells in order.
• Outputs are written to colab/results/.

---

🤖 For AI Agents (Reproduction-First)

1) Where to start

• Primary entrypoint: SKILL.md
• Notebook mirror: colab/spectralbio.ipynb
• Paper text: paper/spectralbio.tex and paper/spectralbio_clawrxiv.md

2) Minimal execution contract

• Model: facebook/esm2_t30_150M_UR50D
• Seeds: 42 (torch, numpy, random)
• Core deps: torch, transformers, scipy, scikit-learn, numpy
• Expected key metrics: TP53 ~0.750, BRCA1 LL ~0.917, Δrep=0.0

3) Agent navigation map

If you need...	Open this path	What you get
Reproducible workflow	SKILL.md	Ordered execution steps and validation criteria
Raw TP53 variants	colab/results/tp53_variants.json	Curated benchmark variant list
Raw per-variant scores	colab/results/scores.json	11-feature table for each TP53 variant
Final metric snapshot	colab/results/summary.json	AUCs, counts, timing, reproducibility delta
Visual diagnostics	colab/results/figures.png	ROC + ablation + distributions
Interactive demo code	huggingface/app.py	Space logic and scoring API
Dataset metadata	huggingface/dataset_card.md	Schema, splits, usage notes
Submission flow	submit/submit.py	clawRxiv publish script

4) Output checklist after running

• colab/results/summary.json exists and contains TP53 + BRCA1 metrics.
• colab/results/scores.json exists and has TP53 rows with spectral/LL features.
• colab/results/figures.png exists and renders correctly.
• reproducibility_delta in summary remains 0.0.

5) Submission notes for agents

• submit/submit.py expects local submit/api_key.txt.
• Never commit API keys.
• Publish target is clawRxiv (http://18.118.210.52).

6) 60-second bootstrap commands (agents)

git clone https://github.com/DaviBonetto/SpectralBio
cd SpectralBio
# Primary reproducibility path
# Follow SKILL.md step-by-step and validate outputs in colab/results/

7) Fast troubleshooting map

Symptom	Likely cause	Fix
Equations or metrics not matching expected values	Seed drift or partial pipeline execution	Re-run with seeds fixed to 42 and execute full SKILL.md sequence
Missing figures.png or scores.json	Early stop before evaluation steps	Continue notebook/SKILL flow through final validation block
Submission script fails	Missing API key file	Create local submit/api_key.txt (never commit)
Slow runtime	CPU execution	Prefer T4/A100 runtime for expected wall-clock behavior

---

Project Structure

SpectralBio/
├── README.md
├── LICENSE
├── .gitignore
├── SKILL.md
├── Claw4S_conference.md
├── colab/
│   ├── spectralbio.ipynb
│   └── results/
│       ├── tp53_variants.json
│       ├── tp53_sequence.txt
│       ├── summary.json
│       ├── scores.json
│       └── figures.png
├── paper/
│   ├── spectralbio.tex
│   ├── spectralbio.pdf
│   ├── spectralbio_clawrxiv.md
│   ├── references.bib
│   └── assets/
├── submit/
│   └── submit.py
└── huggingface/
    ├── app.py
    ├── dataset_card.md
    ├── requirements.txt
    ├── README.md
    ├── assets/
    └── data/

---

Reproducibility Snapshot

• Model: facebook/esm2_t30_150M_UR50D
• Window radius: ±40
• Deterministic seeds: 42
• TP53: N=255, AUC_best_pair=0.7498
• BRCA1 transfer subset: N=100, AUC_ll_proper=0.9174
• Reproducibility delta: 0.0
• Scoring time: ~1447s (~24 minutes)

---

Links

• GitHub: DaviBonetto/SpectralBio
• Demo: Hugging Face Space
• Dataset: Hugging Face Dataset
• Full paper PDF: paper/spectralbio.pdf
• Claw4S 2026: claw4s.github.io
• clawRxiv: 18.118.210.52

---

Citation

@inproceedings{claw2026spectralbio,
  title={SpectralBio: Spectral Covariance Analysis of Protein Language Model Hidden States for Zero-Shot Variant Pathogenicity Prediction},
  author={Claw and Bonetto, Davi},
  booktitle={Claw4S Conference 2026, Stanford--Princeton},
  year={2026}
}

---

License

MIT License.

---

Acknowledgments

• Claw4S 2026 organizers
• Stanford University
• Princeton University
• Meta AI (ESM2)
• ClinVar (NCBI/FDA)