Mykhaylo M. Malakhov @mykmal.xyz · 18/08/2026This is all I have to show for my data harmonization efforts, sorry. 130 3
Mykhaylo M. Malakhov @mykmal.xyz · 20/12/2025🧵9/11 By leveraging information from multiple proteins, COWAS also boosts power relative to standard PWAS. For example, alpha-synuclein (SNCA) did not attain significance for Parkinson's when tested on its own, but became significant once considered jointly with any one of several other proteins. 100 1
Mykhaylo M. Malakhov @mykmal.xyz · 20/12/2025🧵8/11 COWAS can also help to disentangle direct and interaction effects. For example, we found that APOE and its genetically regulated interaction with LDLR have opposite effects on Alzheimer's disease, which is consistent with known APOE biology. Standard PWAS was only able to identify APOE itself. 100 1
Mykhaylo M. Malakhov @mykmal.xyz · 20/12/2025🧵6/11 We applied COWAS to @ukbiobank.bsky.social proteomic data and the largest GWASs for LDL cholesterol, Alzheimer's disease, and Parkinson's disease. Although co-expression models had lower performance than single-protein models, we still obtained quality models for over 500 pairs for each trait. 100 1
Mykhaylo M. Malakhov @mykmal.xyz · 20/12/2025🧵5/11 Next, COWAS tests for association between genetically regulated co-expression and the trait of interest while also accounting for direct effects from each exposure. This joint model can be estimated using GWAS summary statistics, making our method just as easy to use as standard TWAS and PWAS. 110 1
Mykhaylo M. Malakhov @mykmal.xyz · 20/12/2025🧵4/11 We developed the COWAS framework to answer this question. First, COWAS trains models to predict expression and co-expression from common genetic variants. In our main results we defined co-expression as the product of expression residuals, removing the linear genetic components of expression. 100 1
Mykhaylo M. Malakhov @mykmal.xyz · 20/12/2025🧵3/11 This assumption can be a good approximation, but it ignores the effects of gene–gene or protein–protein interactions. Recent research has shown that interaction and co-expression effects are regulated by genetic variation. But how does genetically regulated co-expression affect complex traits? 110 1
Mykhaylo M. Malakhov @mykmal.xyz · 07/11/2025This is an exceedingly late post, but I wanted to share a few pictures from #IGES2025 earlier this fall. I had the honor of giving a platform talk on age-related genetic dysregulation, and I really enjoyed connecting with other researchers and exploring the beautiful city of Cologne! 030 3
Mykhaylo M. Malakhov @mykmal.xyz · 07/11/2025More votes have been counted since then, and the updated map looks even better. There's not a single red cone across the whole state! 010 1
Mykhaylo M. Malakhov @mykmal.xyz · 24/07/2025Yesterday I successfully defended my dissertation! It feels surreal to finally reach the end of my PhD journey at @umnsph.bsky.social. Thank you to my adviser Wei Pan and to my committee members @elock.bsky.social, @aazaidi.bsky.social, and Thierry Chekouo for your invaluable mentorship and support! 041 4
Mykhaylo M. Malakhov @mykmal.xyz · 28/05/2025I had a great time at STATGEN 2025 last week! Thank you to everyone who came to my talk and to the organizers for putting on this conference. 000 0
Mykhaylo M. Malakhov @mykmal.xyz · 07/05/2025Importantly, type I error was not inflated because imputed data was only used for training, not inference. Overall, we demonstrated that GWAS-based trait imputation can be used to increase sample sizes and thereby boost power for nonlinear TWAS/PWAS methods that require individual-level data. (6/6) 000 0
Mykhaylo M. Malakhov @mykmal.xyz · 07/05/2025We found that DeLIVR trained on imputed AD status was able to identify disease-relevant genes that were missed if the models were instead trained on observed data, and the same story held true when identifying putative risk proteins. We also showed that LS-imputation has the best performance. (5/6) 100 1
Mykhaylo M. Malakhov @mykmal.xyz · 07/05/2025We considered the nonlinear TWAS/PWAS methods DeLIVR (based on deep neural networks) and TWAS-LQ (based on a parametric polynomial model). We focused on Alzheimer's disease (AD) and imputed AD status from family history or from GWAS summary statistics using LS-imputation, PRS-CS, and LDpred2. (4/6) 110 1
Mykhaylo M. Malakhov @mykmal.xyz · 07/05/2025But linear TWAS/PWAS methods have one major advantage: linear stage 2 models can be estimated using GWAS summary statistics instead of individual-level data. So what if we use GWAS summary statistics to create imputed individual-level datasets and then train nonlinear TWAS/PWAS models on them? (3/6) 100 1
Mykhaylo M. Malakhov @mykmal.xyz · 07/05/2025I'm excited to share our latest work, now out in #PLOSGenetics! If you feel limited by linear models in TWAS or PWAS but don't think you can switch to nonlinear methods due to a lack of individual-level phenotype data, this paper is for you. 🔗 journals.plos.org/plosgenetics... Thread below 👇 (1/6) 195 9
