Convolutional networks
Can a CNN predict how cancer cells respond to drugs?
Open access · cc by · source: Europe PMC
A two-branch convolutional network predicted drug response well for drugs it had seen before, but almost failed on entirely new drugs.
Study at a glance
- Design
- Computational / modelling — Regression of normalised IC50 from one-hot SMILES and 735 genetic features; random 80/10/10 pair splits repeated 50 times, plus drug-blind, cell-line-blind and tissue-blind splits.
- N
- N=172114 · 172,114 measured drug-cell line IC50 values among 223 drugs and 948 cell lines from GDSC.
- Population
- Anti-cancer drugs and cancer cell lines in the Genomics of Drug Sensitivity in Cancer database
- Outcome
- Coefficient of determination, Pearson correlation and RMSE between predicted and measured IC50
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
With random pair splits the model reached a mean R squared of 0.826 and Pearson correlation of 0.909 over 50 repeats, above an older neural network baseline. When whole drugs were held out, performance collapsed (mean Pearson correlation around 0.2), while held-out cell lines were handled much better. Shuffling or rearranging the SMILES barely changed accuracy, showing the network was exploiting statistical patterns rather than chemical meaning; it also stayed accurate with far fewer cell-line features.
Methodology
The authors built tCNNS, which runs one 1D convolutional network over a drug's SMILES string (encoded as a one-hot matrix) and another over a cell line's mutation and copy-number features, then combines them in fully connected layers to predict the drug's IC50. They used 172,114 measured drug-cell line pairs covering 223 drugs and 948 cell lines. Beyond random splits, they tested on held-out drugs, held-out cell lines and held-out tissues, and scrambled the SMILES input to see what the network actually learned.
Limitations
The comparison with the earlier baseline is indirect because the database version changed, which the authors say may be overly optimistic. Random pair splits let the same drugs and cell lines appear in training and testing, so the headline scores mainly reflect filling in a table, not predicting new drugs. Suggested new drug uses were supported only by literature searches, not lab tests, and the scrambling result means the model offers little chemical insight.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
Random splits can measure table-filling, not prediction for new drugs.
A CNN predicting cancer drug response reached Pearson 0.909 on random drug-cell-line pair splits but about 0.2 when whole drugs were held out, and shuffling SMILES barely changed accuracy.
Evidence for the claim as stated.
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
Evidence for the claim as stated.
Open questions
Tensions this paper is part of
From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
Related papers in this topic
Same topic cluster — not a recommendation engine.