Webinar 2pm today (Friday 6 June 2025): Genetics Centre of Excellence (Edinburgh Ponting lab): update on recent research

Would it be worth reaching out to PrecisionLife and expressing some of our questions and concerns? Maybe invite them to come on here and discuss their work with us?

Or might that be something better left until LOCOME publishes?
 
Agree with everything you say in your post, wanted to ask if you had heard these trials were already happening? I swear Sayoni Das said they would start after they had the locome results in a presentation i saw once.
Maybe it's that, I think I've seen a presentation suggesting the trials would be underway by now. I thought I'd also read that they had sorted the genetic test they need to identify sub groups protesting.

But I haven't seen any confirmation that those trials are in progress – that was an assumption. If no one else has heard it, but it's probably not the case!
 
Obviously, they are in a difficult situation as they're being funded by investors, who probably don't want to hear about the niceties of experiments and bumps along the road.

Actually, having spent quite a lot of time advising investors on biotech, sensible investors do the due diligence and want to know the details - even if that means getting them from a third party.

And this webinar was clearly designed for the benefit of patients, who are entitled to have a presentation that they can understand.
 
Action for ME:
Recording available!
This webinar was really well done. Ran smooth, clear concise presentations with actual data on a variety of projects with good information provided. And what a great way for Action to ME to share how the money that funded a Phd was put to use. I wish all charities communicated projects they are involved with like this.

There was an interesting tidbit at the end that implied that Precision Life might have pulled out of the ME field but a "fireside chat" about the disease to their full team by a person with ME was really helpful, as was the promise of the large dataset coming that is DecodeME which has well phenotyped data that they can actually trust.
 
I understand how a basic GWAS works but not these other computational approaches and so I have no idea to what extent the results can be trusted.
Late-comer to this thread though I am, and non-scientist to-boot, I have been trying to get to the bottom of this "combinatorial" approach.

My take on it is to:
a) enumerate all possible SNP combinations
b) for each of these SNPcombinations:
... count Ca=#cases with it, and Co=#controls with it
... form the usual 2x2 contingency table: Ca (#cases - Ca) / Co (#controls - Co)
... do Fisher/chi-square test on that 2x2 table, giving OddsRatios, pValues, stdErrors
c) choose those SNPcombinations whose results (OddsRatios, pValues, stdErrors) meet your significance criteria

Full enumeration of all SNPcombinations is doubtless too compute-intensive, so PL (and others in this space) will have resorted to some cunning speedups eg choosing onlt SNP genotypes and combinations occurring significantly more often in patients than in controls. And there will be the problem that # of SNPcombinations is so high that most are found only once or very few times in the sample, and thus cannot obtain statistical significance. However, SNPcombinations showing some similarity as a shared SNP genotype can be grouped into clusters that can be tested statistically.

Anyway, can this approach "be trusted"? Winding back in time through the PL papers' references I arrived at this: "The underlying analytical mining platform has been validated in multiple disease populations" which refers to this non-PL paper, which says "The main objective of the study was to find genetic variants that in combination are significantly associated with bipolar disorder." which does not sound like multiple-disease-populations.

My main unease with all this is that it tortures the data a lot, but has little biology.
 
Last edited:
Late-comer to this thread though I am, and non-scientist to-boot, I have been trying to get to the bottom of this "combinatorial" approach.

My take on it is to:
a) enumerate all possible SNP combinations
b) for each of these SNPcombinations:
... count Ca=#cases with it, and Co=#controls with it
... form the usual 2x2 contingency table: Ca (#cases - Ca) / Co (#controls - Co)
... do Fisher/chi-square test on that 2x2 table, giving OddsRatios, pValues, stdErrors
c) choose those SNPcombinations whose results (OddsRatios, pValues, stdErrors) meet your significance criteria

Full enumeration of all SNPcombinations is doubtless too compute-intensive, so PL (and others in this space) will have resorted to some cunning speedups eg choosing onlt SNP genotypes and combinations occurring significantly more often in patients than in controls. And there will be the problem that # of SNPcombinations is so high that most are found only once or very few times in the sample, and thus cannot obtain statistical significance. However, SNPcombinations showing some similarity as a shared SNP genotype can be grouped into clusters that can be tested statistically.

Anyway, can this approach "be trusted"? Winding back in time through the PL papers' references I arrived at this: "The underlying analytical mining platform has been validated in multiple disease populations" which refers to this non-PL paper, which says "The main objective of the study was to find genetic variants that in combination are significantly associated with bipolar disorder." which does not sound like multiple-disease-populations.

My main unease with all this is that it tortures the data a lot, but has little biology.
As you say, the combinatorial principle is straightforward. The trick is finding a good way to search the "computational space". And that's where their proprietary, black box method comes in. Which means it's not possible to validate it.

They point to quite a lot of evidence for validity in diseases on their website. The crown jewel is the early work on Covid. They identified a substantial number of genes very early on. A good number of these were subsequently identified by bigger studies using traditional methods. And some of their work – which also looks identifying drug targets using their own AI – identified drugs that later proved successful in treating acute Covid.

I think they have other studies showing things similar things for chronic ill illnesses, but I am not sure that evidence is as clear cut
 
Last edited:
"The underlying analytical mining platform has been validated in multiple disease populations" which refers to this non-PL paper, which says "The main objective of the study was to find genetic variants that in combination are significantly associated with bipolar disorder." which does not sound like multiple-disease-populations.

That 2019 "bipolar" paper says "
In bipolar disorder hyperactivity is the main symptom of the manic phase, possibly reflecting faster signal transmission in the brain. Based on this assumption we have investigated genes related to the action potential, refractory period, ion channels and CNS myelination. Among such genes 55 were selected based on a search in Medline for genes associated with bipolar disorder. In the first study of this material a table shows the 55 genes and the corresponding proteins
"
… so those researchers used medical knowledge to pre-select the candidate genes! and then searched for SNP-combinations in those genes.

In contrast, my impression of the PL work was that they identified genes FROM a combinatorial search for SNP-combinations ie very much a data-driven discovery process, applying medical knowledge/interpretation AFTERwards. Am I wrong?
 
Back
Top Bottom