🧬 biology

Large-scale AI-Ready Data for Anti-Cancer Drug Response Modeling

This paper presents a significantly expanded, large-scale AI-ready dataset for anti-cancer drug response modeling that integrates millions of measurements and over 50,000 new compounds, demonstrating that models trained on this resource achieve superior generalization to unseen drugs compared to the original IMPROVE benchmark.

Original authors: Vincent Lavelle, Yitan Zhu, Kaitlyn Marlor, Thomas Brettin, Rick Stevens

Published 2026-08-13

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Vincent Lavelle, Yitan Zhu, Kaitlyn Marlor, Thomas Brettin, Rick Stevens

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve the ultimate mystery: why does a specific medicine work like a superhero for one person's cancer but fail completely for another's? This is the world of pharmacogenomics, a field where scientists try to match the right drug to the right patient by looking at two big clues: the chemical makeup of the drug and the genetic "fingerprint" of the cancer cells. For a long time, these detectives had a hard time because their clue books were too small and messy. They had data from a few experiments, but the chemicals looked different in every book, and the cell types were described in confusing ways. It was like trying to learn a new language using dictionaries written in different alphabets. Without a massive, clean library of information, the computer programs (AI) designed to predict which drugs would win couldn't learn the deep patterns needed to find new cures. They were smart, but they were starving for data.

This paper is about building that massive, super-organized library. The researchers took a standard, well-organized dataset called IMPROVE and gave it a giant upgrade. They didn't just add a few pages; they merged in millions of new measurements from a huge collection of drug experiments called PharmacoDB. They cleaned up the chemical names so every drug had a single, perfect ID, standardized how they measured if a drug killed a cancer cell, and added over 53,000 different chemical compounds to the mix. They also included a wider variety of cancer types and genetic data, turning a small notebook into a giant encyclopedia.

When they tested their AI models on this new, expanded library, they found something exciting. The models got significantly better at predicting how the drug would work against new, unseen chemicals (a scenario they call "drug-blind") and when facing completely new combinations of drugs and cancer cells (the "disjoint" setting). It's as if the AI, after reading millions more pages, finally learned the general rules of chemistry and biology well enough to guess the outcome of a drug it had never seen before. However, the paper notes a small twist: the models didn't get much better at predicting outcomes for drugs and cancers they had already seen in the old, smaller dataset. In fact, their performance on those familiar cases dipped slightly. The authors suggest this might be because the new data is dominated by one specific type of experiment (from the NCI60 study) that uses a slightly different testing method than the others. So, while the AI became a master at guessing the unknown, it got a little less confident with the familiar. Ultimately, this work suggests that feeding AI more diverse data is the key to helping it discover new cancer treatments, even if it makes the model a bit more specialized in handling the unknown.

Technical Summary: Large-Scale AI-Ready Data for Anti-Cancer Drug Response Modeling

Problem Statement

Drug response prediction (DRP) models are critical for accelerating anticancer drug discovery, yet their predictive performance is frequently limited by the scale and diversity of available training data. Existing pharmacogenomic datasets often suffer from insufficient coverage of chemical and cancer spaces, inconsistent data formats, and heterogeneous preprocessing pipelines. Furthermore, the lack of standardized benchmarking practices hinders reliable model comparison. While the IMPROVE project established a unified framework for DRP evaluation, its underlying benchmark dataset did not fully capture the scale and diversity of modern pharmacogenomic data, limiting the generalizability of models trained upon it.

Methodology

Data Integration and Curation

The authors constructed an expanded, AI-ready pharmacogenomic dataset by integrating data from multiple sources, primarily PharmacoDB, alongside smaller studies involving patient-derived organoids (PDOs). The integration encompassed:

Sources: Cancer Cell Line Encyclopedia (CCLE), CTRPv2, GDSC (v1 and v2), gCSI, NCI-60, PRISM, FIMM, and various PDO studies.
Scale: The final resource contains approximately 5.46 million drug–response experiments, covering 1,362 cancer cell lines and 53,949 compounds (an increase of over 50,000 compounds compared to the original benchmark).
Multi-omics Coverage: The dataset includes six molecular profiling modalities: gene expression, copy number variation (CNV), somatic mutations, DNA methylation, protein expression (RPPA), and miRNA expression.

Standardization Workflow

To ensure the data was "AI-ready," a rigorous, unified preprocessing pipeline was applied:

Drug Response Modeling: Raw viability measurements were processed using a unified dose-response fitting scheme. A four-parameter logistic (4PL) model was fitted to normalized viability data. To handle non-canonical responses (e.g., increasing viability with dose), a flexible fitting regime was employed, selecting the best fit based on the coefficient of determination ( $R^2$ ). Area Under the Curve (AUC) was computed as the primary summary metric.
Chemical Featurization: Compound representations were standardized using canonical SMILES strings. Molecules were cleaned (removing salts/counterions), and tautomers were collapsed. Two feature modalities were generated:
- Physicochemical descriptors: Computed via the Mordred library (excluding 3D descriptors).
- Molecular fingerprints: Extended-connectivity fingerprints (ECFP4) with a radius of 2 and 512-bit length.
Multi-omics Harmonization: Gene identifiers were mapped to a unified scheme (NCBI/Ensembl/Entrez) to ensure compatibility. Cell lines were mapped to Cellosaurus accession IDs. Gene expression data was normalized (log2(TPM+1)) and quantile-normalized across sources to reduce distributional differences while preserving rank information.

Model Training and Evaluation

Two distinct DRP architectures were evaluated to assess the impact of the expanded dataset:

UNO: A deep neural network using predefined physicochemical descriptors for drugs and gene expression for cell lines.
GraphDRP: A graph-based deep learning model that learns drug representations directly from canonical SMILES via graph convolutional networks.

Evaluation Protocol:
Models were trained on both the original IMPROVE benchmark and the expanded dataset. Performance was assessed using 10-fold cross-validation with a shared test framework to ensure direct comparability. The evaluation included four specific splitting strategies to test generalization:

Mixed: Standard random split (drug and cell line seen in training).
Cancer-blind: Cell lines unseen during training.
Drug-blind: Compounds unseen during training.
Disjoint: Both cell lines and compounds unseen during training (the most stringent test).

Key Results

The expanded dataset demonstrated significant improvements in generalization capabilities, particularly for unseen compounds:

Drug-Blind and Disjoint Performance: Models trained on the expanded dataset showed consistent improvements in predicting responses to unseen compounds.
- UNO: $R^2$ improved from 0.03 to 0.22 in the drug-blind setting and from -0.10 to 0.20 in the disjoint setting.
- GraphDRP: $R^2$ improved from -0.11 to 0.12 in the drug-blind setting and from -0.26 to 0.08 in the disjoint setting.
Mixed and Cancer-Blind Performance: Performance in these settings showed a slight decrease or remained comparable. The authors attribute this to the composition of the expanded dataset, which is heavily dominated by NCI-60 data. Since the test sets were derived from GDSC, CCLE, and CTRPv2 (using different assay protocols than NCI-60), the shift in training distribution may have introduced a bias that reduced performance on these specific "in-distribution" test sets.
Stability: The standard deviation of performance metrics generally decreased or remained stable in the drug-blind and disjoint settings, suggesting that the improvements were robust across cross-validation folds rather than driven by favorable data splits.

Significance and Claims

The paper positions the expanded dataset as a community resource designed to provide a richer foundation for developing DRP models. The authors claim that:

Enhanced Generalization: The primary contribution is the demonstration that increasing the scale and chemical diversity of training data leads to better generalization to previously unseen compounds, a critical requirement for discovering novel anticancer drugs.
Standardization: The work reinforces the importance of standardized preprocessing and unified interfaces for reproducible AI research in pharmacogenomics.
Foundation for Future Work: The dataset is well-suited for exploring pretraining strategies and foundation-model approaches. While the current analysis focused on transcriptomic and chemical features, the inclusion of diverse multi-omics modalities offers opportunities for future integration to further improve predictive performance.

The authors remain modest regarding the trade-offs observed, acknowledging that the current dataset introduces computational burdens and potential distributional biases due to the heavy weighting of NCI-60 data. However, they conclude that the resource successfully advances the ability of the community to train and evaluate models in challenging generalization scenarios.