tessera-foundation 0.1.0

TESSERA: a foundation model for the cancer genome (joint SNV+CNA self-supervised pretraining).

Tumour Embeddings via Self-Supervised Encoding and Reconstruction of Alterations
A foundation model for the cancer genome.

---

TESSERA is a self-supervised foundation model jointly pretrained on somatic single-nucleotide variants (SNVs) and copy-number alterations (CNAs) from the TCGA Pan-Cancer Atlas. A single learned representation, produced once and reused without retraining, supports variant pathogenicity prediction, pan-cancer tumour-type classification, unsupervised molecular subtyping, prognostic stratification, and counterfactual treatment-effect estimation.

This repository contains the reference implementation, the pretrained-weights pointer, the inference utilities described in the accompanying paper, and the end-to-end analysis pipelines that reproduce every panel of Figures 1-6 and Supplementary Figures 1-12.

Quick start

The fastest way to use TESSERA is via the public inference API on Hugging Face; no local installation required. Upload SNV and/or CNA data, get back per-variant predictions and embeddings:

🔗 Inference API: huggingface.co/spaces/JW-Sidhom-Lab/tessera (coming soon)

From Python (pip install gradio_client):

import time
from gradio_client import Client, handle_file

client = Client("JW-Sidhom-Lab/tessera")        # the public Spaces URL also works

# Submit returns (status_html, job_id) immediately; inference runs async
_, job_id = client.predict(
    handle_file("snv.csv"),         # SNV CSV; or None
    handle_file("cna.csv"),         # CNA CSV; or None. At least one required.
    True,                           # apply TCGA quantile normalization to CNA
    "[e-mail]",              # email address for the download link
    "GRCh37",                       # genome assembly: "GRCh37" or "GRCh38"
    api_name="/submit",
)

# Poll for completion (the same URL also gets emailed when the job finishes)
while True:
    status = client.predict(job_id, api_name="/status")
    if status["status"] in ("done", "failed"):
        break
    time.sleep(10)

print(status["url"])    # 24h pre-signed S3 download URL with the result ZIP

The API serves the foundation-model outputs only (per-token embeddings + per-token reconstruction predictions, returned as .npy files inside the result ZIP). Downstream task heads (tumour-type classifier, treatment-effect score) are available on request under a Data Use Agreement.

CSV column conventions:

• SNV: Tumor_Sample_Barcode, Chromosome (no chr prefix), Start_Position, Reference_Allele, Tumor_Seq_Allele2, plus either vaf or both t_alt_count + t_ref_count. Single-base substitutions only.
• CNA: Tumor_Sample_Barcode, Chromosome, Start, End, Segment_Mean (log2 ratio); optional LOH column triggers the with-LoH model variant.

Local installation

For users who want to run inference offline, integrate TESSERA into a custom pipeline, or retrain on their own data:

# Clone
git clone https://github.com/JW-Sidhom-Lab/tessera.git
cd tessera

# Recommended: a virtual environment so deps don't clash with system Python
python3 -m venv .venv && source .venv/bin/activate

# Install all dependencies
pip install -r requirements.txt

# Download reference genome (default: GRCh37)
bash tessera/ref_genomes/download_ref_genomes.sh

requirements.txt covers the foundation-model package, all manuscript-reproduction scripts (pretraining, classifiers, prognostic / predictive-biomarker analyses), and the Gradio inference API. A trimmer subset for deploying only the inference API is at inference_api/requirements.txt.

Weights are hosted on Hugging Face Hub at huggingface.co/JW-Sidhom-Lab/tessera-foundation under CC-BY-NC-4.0. The shortest path from raw dataframes to feature tensors is the featurize one-liner, which downloads weights on first call (cached afterwards), lifts non-hg19 coordinates, builds the dataset, and runs both per-modality feature heads:

import tessera

result = tessera.featurize(
    snv_df=snv_df,                      # columns: Tumor_Sample_Barcode, Chromosome, Start_Position,
                                        #          Reference_Allele, Tumor_Seq_Allele2, vaf
    cna_df=cna_df,                      # columns: Tumor_Sample_Barcode, Chromosome, Start, End, Segment_Mean
    variant="joint_snv_cna_noloh",      # or "joint_snv_cna" for the with-LoH variant
    from_assembly="GRCh38",             # "GRCh37" / "hg19" is a no-op; otherwise UCSC liftover runs
)

result.snv_features      # (n_variants, 1169)  per-variant embeddings, row-aligned with result.snv_table
result.cna_features      # (n_segments, 688)   per-segment embeddings, row-aligned with result.cna_table
result.liftover_stats    # {"snv": {"n_in", "n_out", "n_dropped"}, "cna": {...}}

For finer-grained control there are still building blocks:

from tessera import load_pretrained, lift_snv, lift_cna

model = load_pretrained("joint_snv_cna_noloh")    # download + instantiate, ~3 s cold
snv_df, _ = lift_snv(snv_df, from_assembly="GRCh38")    # identity if from_assembly=="GRCh37"
cna_df, _ = lift_cna(cna_df, from_assembly="GRCh38")
result = model.featurize(snv_df=snv_df, cna_df=cna_df)  # repeat without re-downloading

UCSC chain files are downloaded on first use and cached at ~/.cache/pyliftover/; offline environments can point the loader at a bundled chain file via the chain_file= argument or the TESSERA_LIFTOVER_CHAIN environment variable.

Reproducing the manuscript

Every published panel is backed by a script in this repository. The pipeline runs in three stages:

1. Data preparation (data/): per-cohort download instructions, source-table provenance, and the create_training_data*.py / build_<cohort>_metadata.py builders that turn raw releases into the analysis-ready CSVs.
2. Foundation-model pretraining (scripts/tcga_pancan_*/): trains the SNV models, the CNA models, and the joint SNV+CNA InfoNCE-aligned foundation model on the TCGA Pan-Cancer Atlas.
3. Downstream analyses (scripts/): variant-pathogenicity (Fig. 1 h-o), cross-platform validation (Fig. 1 f-g, Fig. 2 d), tumour-type classification (Fig. 3, Fig. 4 b-e), prognostic UMAP + joint Cox (Fig. 5), doubly-robust counterfactual treatment-effect (Fig. 6 a-m), and DepMap cell-line transfer (Fig. 6 n).

scripts/README.md and data/README.md hold the full per-directory tables mapping each script and cohort to its manuscript figure.

Repository layout

tessera/
├── tessera/                        # foundation-model package
│   ├── base.py                     # BaseModel: shared data + training infrastructure
│   ├── input_keys.py               # input-key helpers
│   ├── model.py                    # TESSERA: foundation-model class
│   ├── data/
│   │   └── preprocessing.py        # SNV/CNA tokenization, FASTA lookup, sample bagging
│   ├── layers/                     # custom Keras layers (attention, masking, MIL, ...)
│   ├── training/                   # training utilities (callbacks, losses, schedules)
│   └── ref_genomes/                # reference-genome download script + indices
├── data/                           # per-cohort data preparation pipelines (data/README.md)
├── scripts/                        # analysis pipelines backing the manuscript figures (scripts/README.md)
└── README.md

Citing TESSERA

If you use TESSERA in your work, please cite:

citation pending publication

A BibTeX entry will be added on acceptance.

License

This repository is distributed under the PolyForm Noncommercial License 1.0.0 (see LICENSE). Use is permitted for academic research, education, public-research-organization use, and personal experimentation; commercial use is not permitted without a separate license. Pretrained foundation-model weights are released on the Hugging Face Hub under CC-BY-NC-4.0 (non-commercial, attribution required). Pretrained weights for downstream clinical task heads (CRC and PDAC treatment-effect models) remain available on request under a Data Use Agreement. Patents covering clinical applications of TESSERA are assigned to NewYork-Presbyterian; commercial licensing inquiries should be directed to NYP's technology transfer office.

Lab

TESSERA is developed in the JW Sidhom Lab at Weill Cornell Medicine.

For questions, collaborations, or commercial-licensing enquiries, contact the corresponding author.