Mun-Lee/txgemma-2b-pgx-lora-210k

TxGemma PGx LoRA Adapter — 210k Checkpoint

This repository provides the manuscript-specific LoRA adapter used for the pharmacogenomic (PGx) representation component of the study “Leveraging Pharmacogenomic Signals to Augment Genomic–Treatment Representations for Prognostic Prediction of 1-Year Mortality in Lung Cancer.”

The released adapter corresponds to the selected 210,000-step checkpoint used for PGx representation extraction. This repository distributes the adapter only; it does not redistribute Google TxGemma base-model weights.

Model artifact

• Base model: google/txgemma-2b-predict
• Adapter framework: LoRA / PEFT
• Selected checkpoint: 210,000 steps
• LoRA rank (r): 8
• LoRA alpha: 8
• LoRA dropout: 0.0
• Bias: none
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Task type: causal language modeling

The repository root contains:

• adapter_config.json
• adapter_model.safetensors
• LICENSE
• NOTICE
• MODIFICATIONS.md
• CHECKSUMS.sha256
• examples/synthetic_clinical_examples.jsonl
• examples/synthetic_pgx_contexts.jsonl

Base-model access and adapter loading

The adapter requires separate access to google/txgemma-2b-predict. Access to the base model is governed by Google's Health AI Developer Foundations (HAI-DEF) Terms of Use and may require acceptance of those terms on Hugging Face.

A minimal PEFT loading pattern is:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "google/txgemma-2b-predict"
adapter_path = "."  # repository root after cloning/downloading

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto",
    torch_dtype="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_path)
model.eval()

The code above demonstrates adapter loading only. It is not intended to reproduce the protected clinical cohort or the complete downstream survival-prediction analysis.

Synthetic clinical examples

The examples/ directory contains a small set of fully synthetic, fictional clinical examples created solely to illustrate the expected structure of study inputs.

• examples/synthetic_clinical_examples.jsonl contains five fictional patient-level examples.
• examples/synthetic_pgx_contexts.jsonl contains two illustrative PGx inference-context examples derived from those fictional records.

These examples:

• were created from scratch;
• do not contain real patient records;
• are not perturbations or transformations of real patient records;
• are not intended to reproduce the study cohort distribution; and
• are not intended to reproduce the reported model-performance estimates.

No outcome labels are included in the synthetic clinical examples. The study target was 1-year all-cause mortality from the predefined landmark.

PGx inference-context examples

The illustrative PGx context records demonstrate, at a structural level, how a medication-specific molecular representation and altered-gene context can be represented together for encoder input. They are provided for transparency only and are not asserted to reproduce an exact production prompt string or a real patient record.

These examples are distinct from the GDSC fine-tuning prompt format described in the Supplementary Methods.

Representation lineage

At a high level, the study pipeline can be summarized as:

Protected clinical / genomic source data
        ↓
Integrated patient-level clinical information
        ↓
 ┌──────────────────┬──────────────────────┬────────────────────────┐
 │ ENF representation│ PGx inference context│ CTE / E2E / GKC       │
 │                  │                      │ baseline representations│
 └──────────────────┴──────────────────────┴────────────────────────┘
                              ↓
                    Model-ready representations
                              ↓
                      EHR + PGx late fusion
                              ↓
                           XGBoost
                              ↓
                 1-year mortality prediction

The patient-level PGx representation described in the manuscript is obtained by aggregating context-level encoder representations. Protected patient-level clinical data and patient-level PGx feature vectors are not included in this public repository.

Scope and reproducibility boundary

This repository is intended to support model accessibility, methodological specification, and privacy-preserving transparency for the study.

Included:

• the manuscript-specific LoRA adapter;
• the exact PEFT adapter configuration distributed with that adapter;
• synthetic patient-level input examples; and
• illustrative PGx inference-context examples.

Not included:

• protected patient-level EHR or genomic data;
• real patient identifiers or outcome labels;
• pre-calculated real-patient PGx feature vectors;
• downstream patient-level prediction files; or
• downstream classifier weights fitted to the protected clinical cohort.

Accordingly, the public materials support inspection and reuse of the PGx adapter and the input structure without releasing protected clinical data. They do not by themselves reproduce the manuscript's patient-level performance estimates.

Modification notice

This repository contains a manuscript-specific LoRA model derivative of google/txgemma-2b-predict. The adapter weights were produced by study-specific LoRA fine-tuning, and the selected released artifact corresponds to the 210,000-step checkpoint. No Google base-model weights are redistributed here. See MODIFICATIONS.md for a concise artifact-level notice.

License and use restrictions

This adapter is a model derivative of TxGemma and is distributed subject to the Health AI Developer Foundations Terms of Use:

https://developers.google.com/health-ai-developer-foundations/terms

The HAI-DEF Prohibited Use Policy incorporated by those terms is available at:

https://developers.google.com/health-ai-developer-foundations/prohibited-use-policy

A local copy of the HAI-DEF Terms of Use is included in LICENSE. Users are responsible for complying with those terms, including applicable use and redistribution restrictions. Access to and use of the separate TxGemma base model are also subject to the base model's applicable terms. The required HAI-DEF distribution notice is included in NOTICE.

Google does not endorse this repository or the associated study. TxGemma, Gemma, and Google are referenced only to identify the upstream model and provider.

Associated manuscript

Leveraging Pharmacogenomic Signals to Augment Genomic–Treatment Representations for Prognostic Prediction of 1-Year Mortality in Lung Cancer

PLOS Digital Health — major revision materials.