Learning and actioning general principles of cancer cell drug sensitivity

Francesco Raimondi, Pierluigi Di Chiaro, Mariangela Morelli, Chakit Arora, Luisa Bisceglia, Natalia De Oliveira Rosa, Alice Cortesi, Sara Franceschi, Francesca Lessi, Anna Luisa Di Stefano, Orazio Santo Santonocito, Francesco Pasqualetti, Paolo Aretini, Pasquale Miglionico, Giuseppe R. Diaferia, Fosca Giannotti, Píetro Lió, Miquel Duran‐Frigola, Chiara Maria Mazzanti, Gioacchino NatolibioRxiv 2024 · DOI

TL;DR

Researchers have developed CellHit, a machine learning framework that predicts how cancer cells will respond to drugs by combining cell-line transcriptomics with biological insights curated by Large Language Models (LLMs). The model doesn't just predict outcomes; it "learns" the underlying biology, successfully translating findings from lab-grown cell lines to real-world patient data and identifying effective treatments for lethal cancers like glioblastoma.

What Problem Does This Paper Address?

For decades, scientists have used high-throughput screening to test thousands of drugs on hundreds of cancer cell lines (CCLs). While these datasets are massive, two major hurdles remain:

  1. The Translation Gap: A drug that kills a cancer cell in a petri dish often fails in a human patient. Aligning the molecular profile of a lab cell line with a complex patient tumor is computationally difficult.
  2. The "Black Box" Problem: Many predictive models are accurate but opaque. They don't explain why a drug works, making it hard for clinicians to trust them or for researchers to discover new biological mechanisms.

The authors of this paper sought to create a model that is both highly accurate and biologically interpretable, using AI to bridge the gap between the lab and the clinic.

Methods

The authors developed CellHit, which primarily utilizes XGBoost (an efficient gradient-boosting algorithm) to predict drug sensitivity (measured as IC50 or log-fold change in cell viability).

The most novel aspect is the MOA-priming pipeline:

  • LLM Curation: The team used LLMs (GPT-4 and Mixtral Instruct) to read drug metadata and PubMed abstracts, generating detailed descriptions of a drug's Mechanism of Action (MOA).
  • Pathway Mapping: These descriptions were semantically matched to biological pathways in the Reactome database.
  • Feature Selection: Instead of feeding the model every single gene (which leads to overfitting), they "primed" the model by training it only on genes within these LLM-identified pathways.
  • Data Alignment: They used a tool called Celligner to harmonize the gene expression data of cell lines with bulk RNA-seq data from actual patients, ensuring the model could make "apples-to-apples" comparisons.

Datasets

The study leveraged several cornerstone datasets in cancer research:

  • GDSC (Genomics of Drug Sensitivity in Cancer): Used for training models on 286 drugs across 686 cell lines.
  • PRISM: A massive dataset used to extend the model to over 6,000 drugs, including non-oncological compounds.
  • CCLE (Cancer Cell Line Encyclopedia): Provided the baseline transcriptomic data for the cell lines.
  • TCGA (The Cancer Genome Atlas): Used for clinical inference on over 10,000 patient samples across 33 cancer types.
  • Primary Patient Samples: New experimental data from Pancreatic Ductal Adenocarcinoma (PDAC) and Glioblastoma (GBM) patients were used for final validation.

Key Insights and Ideas

  • Biology over Chemistry: Surprisingly, the model performed best when using "one-hot encoding" (simple identifiers) for drugs rather than complex chemical fingerprints. This suggests that the cell's internal state (transcriptomics) is a much stronger predictor of drug response than the drug's structure itself.
  • The Model "Learns" Targets: Through interpretability tools like SHAP values, the authors found that the genes the model deemed "most important" often matched the drug's actual biological target (e.g., the model identified BCL2 as the key feature for the drug Venetoclax).
  • LLMs as Scientific Curators: LLMs proved highly effective at extracting "dark data"—biological context buried in text—to improve model performance. MOA-primed models often outperformed "all-gene" models, proving that biological "priors" help AI focus on what matters.
  • Essentiality Matters: The model discovered that drug sensitivity is often driven by a "cross-talk" between the drug's specific target and the genes the cell needs to survive (core essential genes).

Key Results

  • High Accuracy: The joint model achieved a Pearson correlation of ρ=0.89 and a low Mean Square Error (MSE=1.55) on the GDSC dataset.
  • Clinical Recovery: When applied to the TCGA patient database, CellHit correctly identified the approved drug for a patient's specific cancer type in 90% of cases.
  • Repurposing Potential: The model identified potential new uses for existing drugs, such as using the Parkinson’s drug PARDOPRUNOX for certain cancers.
  • Experimental Validation: In the lab, the authors tested CellHit’s predictions on primary cultures from pancreatic and glioblastoma patients. The drugs the model predicted would be effective (like AZD5991 for specific GBM subtypes) were validated by significant reductions in cell viability.

Limitations and Future Work

The authors acknowledge a few key areas for improvement:

  • Specific Activity Profiles: The current regression-based approach struggles with drugs that have very high specificity (i.e., they only work on a tiny fraction of cell lines), as there isn't enough variance for the model to learn from.
  • Computational Bottlenecks: The alignment process (Celligner) currently requires retraining the model whenever new patient data is added. The authors suggest moving toward Variational Auto Encoders (VAEs) to create a more flexible, "on-the-fly" alignment system.
  • Web-App Deployment: The team plans to release CellHit as a web application, allowing researchers to upload patient RNA-seq data and receive immediate drug sensitivity predictions.

Summary updated

Abstract

Abstract High-throughput screening platforms for the profiling of drug sensitivity of hundreds of cancer cell lines (CCLs) have generated large datasets that hold the potential to unlock targeted, anti-tumor therapies. In this study, we leveraged these datasets to create predictive models of cancer cells drug sensitivity. To this aim we trained explainable machine learning algorithms by employing cell line transcriptomics to predict the growth inhibitory potential of drugs. We used large language models (LLMs) to expand descriptions of the mechanisms of action (MOA) for each drug starting from available annotations, which were matched to the semantically closest pathways from reference knowledge bases. By leveraging this AI-curated resource, and the interpretability of our model, we demonstrated that pathways enriched for genes crucial for prediction often matched known drug-MOAs and essential genes, suggesting that our models learned the molecular determinants of drug response. Furthermore, we demonstrated that by incorporating only LLM-curated genes associated with MOAs, we enhanced the predictive accuracy of our drug models. To enhance translatability to a clinical setting, we employed a pipeline to align bulk RNAseq from CCLs, used for training the models, to those from patient samples, used for inference. We proved the effectiveness of our approach on TCGA samples, where patients’ best scoring drugs matched those prescribed for their cancer type. We further showed its usefulness by predicting and experimentally validating effective drugs for the patients of two highly lethal solid tumors, i.e. pancreatic cancer and glioblastoma. In summary, our method facilitates the inference and interpretation of cancer cell line drug sensitivity and holds potential to effectively translate them into new cancer therapeutics. Highlights -Interpretable drug-response prediction models were trained on large scale pharmacogenomics data sets (i.e. GDSC and PRISM). -Large language models were used to enhance the curation of biological pathways associated to drugs MOA -Unbiased interpretation of the models demonstrated learning of drug MOAs and gene essentiality -Inference of TCGA cohort samples recovered mono- and combination cancer drug prescriptions and indicate potential repurposing candidates. -Drug candidates predicted from bulk RNAseq samples of pancreatic cancer and glioblastoma were experimentally validated.
BibTeX
@misc{raimondi2024learning,
  author = {Francesco Raimondi and Pierluigi Di Chiaro and Mariangela Morelli and Chakit Arora and Luisa Bisceglia and Natalia De Oliveira Rosa and Alice Cortesi and Sara Franceschi and Francesca Lessi and Anna Luisa Di Stefano and Orazio Santo Santonocito and Francesco Pasqualetti and Paolo Aretini and Pasquale Miglionico and Giuseppe R. Diaferia and Fosca Giannotti and Píetro Lió and Miquel Duran‐Frigola and Chiara Maria Mazzanti and Gioacchino Natoli and Francesco Raimondi and Francesco Raimondi},
  title = {{Learning and actioning general principles of cancer cell drug sensitivity}},
  year = {2024},
  doi = {10.1101/2024.03.28.586783},
  howpublished = {bioRxiv},
}
Figure