LLMsynthlet: Zero-shot LLM prediction of synthetic lethal interactions

---

Paper: Zero-shot biological reasoning with open-weights large language models reproduces CRISPR screen based prediction of synthetic lethal interactions
Authors: Aurel Prosz, Zsofia Sztupinszki, Miklos Diossy, Bogumil Zimon, Istvan Csabai, Zoltan Szallasi

---

Why this project exists

Synthetic lethality (SL) is a powerful way to uncover cancer-specific vulnerabilities, but experimental discovery (e.g., genome-wide CRISPR screens) is expensive and hard to scale, especially for multi-gene interactions. Most computational SL predictors are data-driven and often lack human-readable mechanistic explanations.

This project tests a different idea: open-weights LLMs can use their internal biological knowledge to reason about gene–gene (and gene–drug) interactions in a zero-shot way, producing both a score and a short rationale.

---

What we show (main results)

• Multiple open-weights LLMs predict Olaparib/PARP1 synthetic lethal hits from genome-wide CRISPR knockout screens better than random.
• Performance generally increases with model size and saturates around ~32B parameters in our comparisons.
• Adding curated pathway/GO annotations and adding cell-line genetic background does not consistently improve performance (and sometimes hurts), suggesting models already carry much of the needed prior knowledge.
• Selected Qwen2.5-32B-Instruct as the best cost/quality tradeoff (0.715 AUROC) and used it to run an in silico screen of 398,277 clinically relevant gene pairs, releasing:
  • full de novo predictions
  • an LLM-assisted filtered subset with novelty/feasibility assessment

---

Repository contents

You can find the data files here: https://drive.google.com/drive/folders/1tS3s9dauWEEm_GElWve36PhaCMFdVn0K?usp=sharing Uploading the analysis scripts are work in progress.

sorted_hypotheses.jsonl

De novo LLM predictions for 398,277 clinically relevant gene pairs (Qwen2.5-32B-Instruct), including a synthetic lethality score and model reasoning.

scored_hypotheses.csv

A filtered subset produced by an LLM-assisted review pipeline that estimates:

• Novelty (0 = known, 10 = likely novel)
• Feasibility (0 = implausible / impractical, 10 = feasible)

batch_infer_no_vllm.py

Inference script used in the study. After downloading the model weights, you can use this script to make predictions.

---

Data notes

• Scores are intended primarily for ranking / prioritization (extreme tails), not binary classification.
• Predictions are hypotheses and require experimental validation.

---

Citation

If you use these outputs, please cite the paper:

Prosz A, Sztupinszki Z, Diossy M, Zimon B, Csabai I, Szallasi Z. Zero-shot biological reasoning with open-weights large language models reproduces CRISPR screen based prediction of synthetic lethal interactions.