27 ms·

That's fascinating. I started a company that cryopreserves adult stem cells. Would you mind sharing what kinds of datasets you are mining? Here or email (in my

That's fascinating. I started a company that cryopreserves adult stem cells. Would you mind sharing what kinds of datasets you are mining? Here or email (in my bio).

xaa 9y ago
Sure. The most mature part of this is analysis of expression data from GEO. We had to take the raw expression data and textual metadata, extract attributes from the semi-structured metadata into something statistically useful, and devise methods to do automated meta-analysis between experiments and platforms on the expression data. We are talking about >1M samples here by 20K-50K transcripts. Now we are basically extending the same idea to new data types and sources, like methylation and sequencing-based data. When we ask "what should we do next", we look for basically what archives are biggest, and if they conceivably have something useful to say about aging, we put it on the queue. The big-picture goal is to take this approach and apply it to all these different levels of organization (DNA, epigenetics, RNA, and ultimately hopefully protein) and put it all together in a cogent picture, using way more data than anyone has done before. We ask questions like "how does the aging process differ between tissues?". We also want to go "wide" by comparing species, as the whole "intervention X works in mice but not humans" is an important thing to address. Then on the smaller scale, the system is also of use and interest to individual investigators who want to know "what's happening with my gene X with age in tissue Y?". They like to include figures from us for example if they have some findings from mice and want to give evidence that the pattern may also hold in humans. Finally, we built the platform to be a pretty generic meta-analysis tool, so we have occasionally used it to answer questions completely unrelated to aging. Once you do all the background legwork, asking many different questions is largely a matter of changing the meta-analytic model.

That's fascinating. I started a company that cryopreserves adult stem cells. Would you mind sharing what kinds of datasets you are mining? Here or email (in my bio).

xaa 9y ago

Sure. The most mature part of this is analysis of expression data from GEO. We had to take the raw expression data and textual metadata, extract attributes from the semi-structured metadata into something statistically useful, and devise methods to do automated meta-analysis between experiments and platforms on the expression data. We are talking about >1M samples here by 20K-50K transcripts. Now we are basically extending the same idea to new data types and sources, like methylation and sequencing-based data. When we ask "what should we do next", we look for basically what archives are biggest, and if they conceivably have something useful to say about aging, we put it on the queue. The big-picture goal is to take this approach and apply it to all these different levels of organization (DNA, epigenetics, RNA, and ultimately hopefully protein) and put it all together in a cogent picture, using way more data than anyone has done before. We ask questions like "how does the aging process differ between tissues?". We also want to go "wide" by comparing species, as the whole "intervention X works in mice but not humans" is an important thing to address. Then on the smaller scale, the system is also of use and interest to individual investigators who want to know "what's happening with my gene X with age in tissue Y?". They like to include figures from us for example if they have some findings from mice and want to give evidence that the pattern may also hold in humans. Finally, we built the platform to be a pretty generic meta-analysis tool, so we have occasionally used it to answer questions completely unrelated to aging. Once you do all the background legwork, asking many different questions is largely a matter of changing the meta-analytic model.