Can a neural network predict which cancer drugs work on which cells?
Combining a graph network for drug structure with gene features predicts drug sensitivity well on familiar cell lines, but error more than doubles on cell lines never seen in training.
Source
SWnet: a deep learning model for drug response prediction from cancer genomic signatures and compound chemical structures
Study at a glance
- Design
- Computational / modelling — Deep model combining a GNN over drug molecular graphs with a 1D-CNN gene branch, trained on cancer cell-line screens with ablations
- N
- N=177128 · Drug–cell-line instances in the processed GDSC dataset (plus a smaller CCLE dataset of 10853 instances)
- Population
- Cancer cell lines screened with anticancer drugs (GDSC and CCLE)
- Outcome
- Mean squared error and R-squared for predicted log IC50
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors built SWnet, which encodes each drug's molecular graph with a graph neural network and each cell line's gene expression and mutations with a convolutional branch, then merges them to predict drug sensitivity (log IC50). They trained and tested it on large cancer cell-line screening datasets using a random 90/10 split. They ran ablations on how mutations are combined, multi-task gene weight layers, self-attention over drug similarity, and graph radius, and finally tested a split where whole cell lines were held out.
What they found
The best model, using multiple gene weight layers plus self-attention, reached an MSE of 0.9384 and R-squared of 0.8583 on the random split and beat earlier published methods. Simply adding mutation data to expression made results slightly worse, while a learned gene weight layer combined them better. When whole cell lines were held out, MSE rose to 2.2793, showing much weaker generalisation to new cell lines.
The limits
What it doesn't show
The headline accuracy comes from a random split where the same cell lines and drugs appear in both training and test data, which overstates performance on truly new cases. Results are on in-vitro cell lines, not patients, so clinical usefulness is untested. Comparisons with prior methods rely on reported numbers rather than a single controlled re-run, and the authors note the model is hard to train on small datasets. The biological interpretation of gene weights is suggestive, based on protein interaction networks, not experimental validation.
Key terms
- IC50
- The drug concentration needed to inhibit a biological process (here cell growth) by half; lower means more sensitive.
- Graph neural network
- A network that learns representations of graphs, here molecules with atoms as nodes and bonds as edges.
- Self-attention
- A mechanism that weights inputs by their similarity to one another; here it lets similar drugs share information.
- Multi-task learning
- Training one model on several related prediction tasks at once so they share representations.
- Data split leakage
- When test examples share entities with training data, making performance look better than it would on truly new cases.
Flashcards
0 of 10 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What part of SWnet encodes drug structure?
Common questions
Why did performance drop so much when cell lines were held out?
In the random split, the model had already seen each cell line with other drugs; with new cell lines it had to generalise across tissues, and error more than doubled.
Why not just add mutation data to expression data?
Simple addition changed the expression data's distribution and slightly hurt accuracy; a learned weight layer combined them better.
Does this mean the model can pick drugs for patients?
No. It was only tested on lab cell lines, and even there it struggled with unseen cell lines.
More on Graph neural networks