It seems like you guys criticize viability screens for lacking "immune-mediated killing, stromal interactions, and metabolic context." However, doesn't your data and model still rely on cell lines in a dish? Even if you measure the transcriptome, a cell line in a plastic well still lacks the immune system and stroma right?
Yes i agree I think these are actually separate issues entirely. IMO viability screens are basically just a coarse uninformative readout. The immune context and stroma (coculture or in vivo or orthotopic etc etc) are questions that can only be dissected through a 1) single cell resolution method 2) model systems. I think here, there are several layers and the point about the viability screens is just that it does not enable the point #1 of single cell resolution. We'll never get to an understanding of ecosystem with viability, basically ; what we are not saying is that this current set of experiments/data already gets us to a full understanding of ecoysystem.
super exciting direction. But on a scale of -1.0 to +1.0, a correlation of -0.10 is arguably negligible. While the difference is statistically significant (due to the large sample size yielding a low p-value), the biological predictive power of such a weak correlation is questionable. It implies the drug is only reversing a tiny fraction of the disease signature, leaving 90%+ of the transcriptomic state unaccounted for or uncorrelated. What am I getting wrong about the above?
well the signatures and the model systems are in entirely different contexts. and, this correaltion value here is on the full set of diverse cancer cell lines - so its pretty diverse and were taking the median. Actually, there is a fair bit of variance between models which we parse internally to see if there are association with certain baseline expression level/organ/driver mutations etc (more corr in similar organs? or not? etc). but of course you can imagine blowing this up into that level of granularity generates a lot of hypotheses and then triaging/filtering/ranking of these.
Very exciting work. It is interesting to me that the average in Figure 1 is not centered around 0. Do you think that has any meaning or impact on the interpretation of the numerical scores?
I think it's primarily a function of the drugs tested - a lot in Tahoe100M are heavily anti-cancer so the distribution makes sense to have a left skew. Other ways to look at it are by digging into the signature that it is reversing. In this case it is "adjacent normal" to "primary tumor". If you imagine drugs that push in that direction - you may have drugs that agonize WNT signaling, or agonists of MYC/MAPK etc - but 1) inherently agonism is harder and 2) oftentimes in cancer these are driven by hyperactive genetic point or CNV mutations, which are pretty hard for drugs to mimic. So this actually tracks a lot with all the priors that we came into this analysis with. It's a thoughtful point though. And based on the signatures we use, we keep an eye on how these distributions are centered.
It seems like you guys criticize viability screens for lacking "immune-mediated killing, stromal interactions, and metabolic context." However, doesn't your data and model still rely on cell lines in a dish? Even if you measure the transcriptome, a cell line in a plastic well still lacks the immune system and stroma right?
Yes i agree I think these are actually separate issues entirely. IMO viability screens are basically just a coarse uninformative readout. The immune context and stroma (coculture or in vivo or orthotopic etc etc) are questions that can only be dissected through a 1) single cell resolution method 2) model systems. I think here, there are several layers and the point about the viability screens is just that it does not enable the point #1 of single cell resolution. We'll never get to an understanding of ecosystem with viability, basically ; what we are not saying is that this current set of experiments/data already gets us to a full understanding of ecoysystem.
how well does this do compared to CMap which has long used differential expression to match drug signatures to disease signatures?
well actually, claude is working on that :)
working on that
super exciting direction. But on a scale of -1.0 to +1.0, a correlation of -0.10 is arguably negligible. While the difference is statistically significant (due to the large sample size yielding a low p-value), the biological predictive power of such a weak correlation is questionable. It implies the drug is only reversing a tiny fraction of the disease signature, leaving 90%+ of the transcriptomic state unaccounted for or uncorrelated. What am I getting wrong about the above?
well the signatures and the model systems are in entirely different contexts. and, this correaltion value here is on the full set of diverse cancer cell lines - so its pretty diverse and were taking the median. Actually, there is a fair bit of variance between models which we parse internally to see if there are association with certain baseline expression level/organ/driver mutations etc (more corr in similar organs? or not? etc). but of course you can imagine blowing this up into that level of granularity generates a lot of hypotheses and then triaging/filtering/ranking of these.
Very exciting work. It is interesting to me that the average in Figure 1 is not centered around 0. Do you think that has any meaning or impact on the interpretation of the numerical scores?
I think it's primarily a function of the drugs tested - a lot in Tahoe100M are heavily anti-cancer so the distribution makes sense to have a left skew. Other ways to look at it are by digging into the signature that it is reversing. In this case it is "adjacent normal" to "primary tumor". If you imagine drugs that push in that direction - you may have drugs that agonize WNT signaling, or agonists of MYC/MAPK etc - but 1) inherently agonism is harder and 2) oftentimes in cancer these are driven by hyperactive genetic point or CNV mutations, which are pretty hard for drugs to mimic. So this actually tracks a lot with all the priors that we came into this analysis with. It's a thoughtful point though. And based on the signatures we use, we keep an eye on how these distributions are centered.