Chris Cole

Er, has anyone else realized that #claude is very happy to read your secrets and echo them back to you in plain text? e.g. it happily tells me all about a value in my .env file. Claude's exclusion rules don't actually work, see here for why and a solution chriscole.substack.com/p/careful-clau…

I’ve been reading a lot of code lately — and I’ve started to think the way someone writes a package tells you a lot about what they value. So, here it is, the perfect #Python package! ... okay, no, I'm kidding. But here are some thoughts :) chriscole.substack.com/p/the-perfect-…

We (@deepgenomics) are hiring a statistical geneticist to develop and apply methods at the intersection of systems biology, genomic AI, and traditional genetic target discovery. CAN/US remote or hybrid, DM me with any questions! jobs.lever.co/deepgenomics/2…

I usually find myself writing about genetics for work but decided to try something different; here's a small #BookReview / reflection after reading @edcaesar's captivating "The Moth and the Mountain". I had a lot of fun writing this! chriscole.substack.com/p/trauma-at-th…

@sampatsmith @tuckerdrob @arbelharpak @s_ramach @lorin_crawford Really interesting work! I'm curious how you interpret the finding that the estimates do not remain significant (the tau_c term?) in the presence of the baseline annotation set. Could it be that co-occuring annotations are driving a portion of the "non-additive" heritability?

@SashaGusevPosts (re: this + Koch preprint) I wonder if calibrating E[mutation rate] to a mixture over "effective consequence" may generalise pLI / LOEUF to different kinds of selection and open up inference of a p(Stabilising selection) which could be used to interpret GWAS loci

[Negm, Veller, doi.org/10.1101/2024.0…] derive a more accurate solution for the effect of stabilizing selection on allele frequency that accounts for background phenotypic variance and LD. If you (like me) are using the first one, you should start using the second one.

@SashaGusevPosts @jgschraiber The field requires a blood sacrifice or reimplementing S-LDSC; whichever occurs first

@jgschraiber Rite of passage github.com/sashagusev/SKK…

That feel when implement REML for GWAS manually because my simulations somehow break GCTA

@doctorveera This reminds me of Bob Handsaker's plenary session at ASHG where he showed that somatic repeat expansions explained Huntington's progression by causing cascading instability, but only if they start out longer than a certain amount; I wonder there are some similar mechanisms here

The first (?) report on the contributions of repeat expansions to Alzheimer's risk. Individuals with a genome-wide burden of STRs with >30 repeat lengths have an OR of 3.63, with a severe degree of brain pathology. Fascinating! Guo, Cremins et al. medRxiv medrxiv.org/content/10.110…

@michelnivard I would suggest this may actually be single variant PTV RV associations (which are becoming more common) in the given gene given the spread of allele frequencies; if it were a thresholded burden with a PTV mask the authors would likely not test such a continuum of AF cutoffs.

@SashaGusevPosts @mikelove @mete_civelek I also found the comparison to be not well fleshed out, you might find this paper interesting in the context of NDD loci (medrxiv.org/content/10.110…) where they compare a (non-viral) MPRA to in vivo readouts in murine models (spoiler: the translatability is not very good)

@SashaGusevPosts @mikelove @mete_civelek Thanks, I hadn't seen the Yao study! Despite their assurances, I'm not convinced that the in vitro system is able to effectively deal with trans effects mediated by cryptic cis effects. I also wonder if the correction for MOI is calibrated for trans effects.

Interesting MPRA analysis of SCZ GWAS variants. Fig. 3 evaluates the utility of different functional annotations in "predicting" whether a variant will be experimentally functional (emVar), some weak enrichments but generally very low utility. eQTLs perform the worst.

@SashaGusevPosts @mikelove @mete_civelek I would be more interested to see this enrichment calculated on the CROP-seq screen than the MPRA data. Lower power with fewer variants but higher fidelity with respect to the endogenous context. The small context size and proxy POL2 readout with the MPRA is hard to interpret.

@mikelove @mete_civelek Very cool, I've been wondering about this for a while. Weak consistency with eqtls I can understand, but *no* consistency with eqtls -- as now seen in several studies -- makes me very curious.

@doctorveera The in vivo portion of this work is particularly exciting, looking at the translatability of variant mechanisms between the MPRA and real cortical context is a massive contribution (and probably extortionately expensive). Is this kind of high-throughput characterization common?

Massive parallel reporter assay (MPRA) of 997 5' UTR variants found in autism probands from Simons simplex collection. Plassmeyer, Dougherty et al. medRxiv medrxiv.org/content/10.110…

@SashaGusevPosts Very interesting. Thank you!

@_chriscole_ Here's the cheat sheet from Purcell et al. 2002; there are enough dials here to make the A, C, E estimates look like whatever you want

Heritability models can neither identify nor explain group differences.

@SashaGusevPosts trait in monozygotic by any appreciable amount (here 1.5) you recover the expected trends

@SashaGusevPosts Oh interesting, thank you. I see that the binomial is drawn over twins rather than individuals, are you concerned that the assumption that trait divergence does not differ between MZ and DZ may be skewing the estimation of heritability metrics? For example if you multiply the

@yluo86 @ProfJohnATodd @NickdeVrij @Ruthnanje I wasn't sure if genome-wide PCs would be appropriate for loci known to under selective pressure (so presumably the local genealogy would diverge substantially from the genome-wide genealogy). Not an expert here though, just curious about your approach. Thank you!

@_chriscole_ @ProfJohnATodd @NickdeVrij @Ruthnanje Not sure I quite understand the question - you can control pop stratification by including genome-wide PCs in permutations ?

I've seen multiple approaches to multiple testing correction in HLA associations. I've seen applying Bonferroni with the total # alleles, doing it on locus-by-locus basis, taking the # alleles only expressed in 5~10% of the population. What's most logical? @Ruthnanje @yluo86

@yluo86 @ProfJohnATodd @NickdeVrij @Ruthnanje Would you be concerned about potentially propagating biases due to population stratification with permutation? Perhaps you have a method that takes this into account, I'm curious particularly due to the complex LD structure and underlying ARG of the locus

@NickdeVrij @Ruthnanje Bonferroni might be too stringent due to LD. We’ve been empirically determine the appropriate threshold via permutations.

It's out! Really excited to share our recent work creating a foundation model for RNA biology and applying it to VEP and oligonucleotide therapeutic design. Huge team effort across both computational and experimental groups; it's an exciting time for AI in genomics 😀