Source-linked AI summary
Drug cell line interaction prediction
Pengfei Liu
TL;DR
Phenotypic screening is important for anti-cancer drug discovery, but prior prediction work had not applied CNNs to this problem. The paper introduces tCNNS, a twin CNN using drug SMILES and cancer cell-line inputs, and reports better coefficient-of-determination and Pearson-correlation results than previous work.
Problem
Phenotypic screening supports anti-cancer drug discovery, while prior prediction research had not applied CNNs to cancer-cell-line drug-response prediction.
Method
tCNNS uses twin CNN networks: one for drugs represented in SMILES format and another for cancer cell lines.
Results
tCNNS achieved much better coefficient-of-determination and Pearson-correlation results than previous works on the newer GDSC dataset.
Takeaways & Limitations
tCNNS provides predictions for missing GDSC values and insights into phenotypic screening, while using SMILES inputs and fewer cancer-cell-line features.
Takeaways & Limitations
tCNNS is less accurate at the two ends of the experimental IC50 range and could be further optimized there.
Abstract
from arXiv · showhide
Understanding the phenotypic drug response on cancer cell lines plays a vital rule in anti-cancer drug discovery and re-purposing. The Genomics of Drug Sensitivity in Cancer (GDSC) database provides open data for researchers in phenotypic screening to test their models and methods. Previously, most research in these areas starts from the fingerprints or features of drugs, instead of their structures. In this paper, we introduce a model for phenotypic screening, which is called twin Convolutional Neural Network for drugs in SMILES format (tCNNS). tCNNS is comprised of CNN input channels for drugs in SMILES format and cancer cell lines respectively. Our model achieves $0.84$ for the coefficient of determinant($R^2$) and $0.92$ for Pearson correlation($R_p$), which are significantly better than previous works\cite{ammad2014integrative,haider2015copula,menden2013machine}. Besides these statistical metrics, tCNNS also provides some insights into phenotypic screening.
1 Introduction
Phenotypic screening remains important in drug discovery, but predicting drug responses on cancer cell lines has mainly relied on conventional machine learning and engineered drug features. The paper introduces tCNNS to apply CNNs directly to drug SMILES and cell-line data.
- Phenotypic screening has produced many approved first-in-class drugs, despite uncertainty in identifying and validating therapeutic targets.
- GDSC and CCLE are major public datasets for studying pharmacologic drug functions and cancer cell-line responses.
- Earlier GDSC work used neural networks, kernelized Bayesian matrix factorization, and multivariate random forests to predict drug responses.
- CNNs support pattern detection for classification and regression and have shown strong results in pharmaceutical research.
- tCNNS introduces twin CNN branches for drugs in SMILES format and cancer cell lines, addressing a problem not previously studied with deep CNNs.
2 Relate Work
Prior work increasingly applies deep learning directly to molecular structures, while drug-response prediction on cancer cell lines has used less advanced architectures and features. This paper positions tCNNS as a CNN-based alternative for drug–cell-line interaction prediction.
- Deep-learning studies have used molecular structures directly because molecular descriptors are a decisive step in many QSAR and classification models.
- SMILES have been converted into learned vectors with autoencoders for drug features, fingerprints, and similarity-based drug-discovery tasks.
- CNNs have also been applied to drug images and molecular 3D structures, but these representations involve accuracy or coverage trade-offs.
- An RNN study used SMILES-derived graphs to predict drug solubility but did not model interactions with cell lines or proteins.
- The benchmark study analyzed IC50 values with a neural network, whereas this paper uses a CNN-based model to predict drug–cell-line interactions.
3 Methods
The method preprocesses GDSC drug, cell-line, and IC50 data, converts canonical SMILES into fixed-size one-hot matrices, and predicts IC50 using twin CNN branches followed by a fully connected regressor. Performance is assessed with R2, Pearson correlation, and RMSE.
- Data and preprocessing: GDSC describes cancer cell lines through genetic features and drug effects through IC50 measurements.
- Data and preprocessing: The preprocessing retains 223 drugs and 948 cell lines after removing duplicates, inconsistent entries, incomplete cell-line features, and unusable IC50 records.
- Data and preprocessing: The dataset contains 211,404 possible drug–cell-line pairs, with 81.4% of IC50 values provided and 18.6% missing.
- Data and preprocessing: Logarithmic IC50 values are exponentiated and normalized to the (0, 1) interval for training and comparison.
- Drug representation: Canonical SMILES are padded to length 188 and represented as 72 × 188 one-hot matrices, with symbols as rows and positions as columns.
- Network architecture: Twin 1D CNN branches separately distill drug and cell-line features, while a fully connected network regresses IC50 from their outputs.
- Implementation and evaluation: The model is implemented in TensorFlow, and evaluation uses R2, Pearson correlation coefficient, and RMSE.
4 Results and Discussion
tCNNS improves drug–cell line IC50 prediction over prior work, while analyses show strong performance on limited, accurate data but weaker generalization to unseen drugs and boundary values. Predictions also recover plausible drug–cell line relationships, although the model appears to rely mainly on statistical SMILES patterns rather than biological structure.
- 0.84 R2, 0.92 Rp, and 0.027 RMSE improve on prior values of 0.72, 0.85, and 0.83, respectively.
- tCNNS is most accurate for middle-range IC50 values and less accurate at the two extremes.The model underpredicts very large IC50 values and does not produce sufficiently small outputs for very small experimental values.
- PaDEL-based results are horizontally stratified, indicating weaker representation power than the tCNNS feature representation.
- Bortezomib has the strongest missing-value predictions, with all 40 lowest predicted IC50 pairs involving this drug.This agrees with its strong performance on existing pairs and reported evidence that it sensitizes cell lines to other anticancer drugs.
- A predicted Bortezomib–NCI-H2342 IC50 of 1.19 ∗10^-4µg is the third-smallest prediction and is supported by lung-cancer-related evidence.
- At 1% of existing IC50 values, tCNNS reaches nearly 0.5 R2 and around 0.7 Rp.
- Data below the maximum screening concentration perform nearly as well as the full dataset despite representing only 37% of the data.
- Low-quality extrapolated data contribute little to accuracy, although they can improve scalability and slightly reduce variation or increase Rp.
5 Conclusion
The paper presents tCNNS for phenotypic screening between cancer cell lines and anti-cancer drugs, evaluated on a newer, larger GDSC dataset. The model improves coefficient of determination and Pearson correlation, supports evidence-based predictions for missing GDSC values, and can converge with limited training data and fewer cell-line features.
- tCNNS predicts interactions between cancer cell lines and anti-cancer drugs for phenotypic screening.
- The model is tested on a newer GDSC version containing more data than previous studies.
- tCNNS achieves better coefficient of determination and Pearson correlation than previous works.
- The model produces predictions for missing GDSC values with trustful evidence and can converge using little training data and fewer cell-line features.