tumour growth modelling
Baseline tumour growth rate may affect treatment efficacy. However, tumours grow at very different rates. For example, the untreated patient-derived colorectal cancer xenograft tumours in mice grow by 0.005-0.172 mm/day as shown in the Novartis mouse clinical trial in 2014. This is over 30-fold difference! Predicting tumour growth rate using gene expression data is hard. Firstly, which tumour growth rate law should you choose? Secondly, gene expression data are inherently noisy. This makes constructing gene signature risky.
Explain patient-to-patient variability by Una Milovanovic
✔ What are the sources of noise in high throughput data?
✔ Which strategy might be useful for tackling the noise?
✔ Can you generate reproducible predictive markers across clinical studies?
Tumour growth rate laws
We have tested 5 tumour growth rate laws using 224 untreated patient-derived xenografts grown in mice, which represent 6 histology including
gastric cancer (n = 44)
colorectal cancer (n = 42)
breast cancer (n = 39)
non-small cell lung cancer (n = 29)
pancreatic ductal adenocarcinoma (n = 37)
cutaneous melanoma (n = 33)
To fit this dataset, exponential-linear, logistic and Gompertz models collapse back down to a simpler rule: the exponential growth model. In addition, the cubic model (assuming tumour diameter changes linearly with time, i.e. tumour volume changes cubically with time) fits the data equally well.
RWR-LASSO regression
The single biggest challenge for gene signature is they are built on noisy data. Unfortunately, most signatures cannot be applied to independent datasets. In fact, most signatures for breast cancer metastasis is no better than random guess!
We are developing a network-based regression method which combines Random Walk with Restart and LASSO regression. It is a gradient-based method. The R implementation is fast. We tested it on the Novartis mouse clinical trial data. It is more accurate than the LASSO regression. It also highlights pathways in cancer that control tumour growth.
This offers an unprecedented pathway-based view of the data: instead of attributing tumour growth solely to the activity of individual genes (mRNA expression level in our case), we view it as the outcome of a system organised at two different levels:
1) Each canonical Hallmarks of Cancer pathway plays a role in mediating certain biological functions affecting the growth and death of tumour cells
2) Each gene on a pathway contributes to the apparent tumour growth
This algorithm helps quantify the relative importance of each pathway, and the contribution of each gene. This systems approach embeds genes in the context of pathways.
