iiinova — where your digital self keeps evolving

(Quotes lightly edited for space. Any errors on me. Do read the original post.)

000

“Many valid, tested, robust, and clearly generalisable generative models will not correspond to the real world if you do this in silico perturbation. The presence of a robust and generalisable association model absolutely does not mean that interventions can be modeled.”

100

(Even ‘in silico perturbation’ might be too strong. Maybe simply ‘simulation’? Generative models capture more complexity [tuned to training data] than your typical biostats simulation [‘now we’re really going to get fancy and throw in 2nd order interactions’], but they’re still simulacra.)

“[G]enerative models generate numbers which are thought to represent real possibilites in the world. It is very tempting to change model inputs and then look at the result. A safe way to describe this is ‘in silico perturbation’, but many use the word ‘counterfactual.’ This is very dangerous.”

100

Peter Kraft @peter-kraft.bsky.social · 23/04/2026

This👇 A key to any productive collaboration between computer scientists and epidemiologists (or clinicians or biologists or…) is to learn each others’ languages. Sometimes we use different words for the same concepts, sometimes we use the same words but mean very different things.

150

Ironically I think it was that same presenter who tl;dred a 300-page paper in a four letter acronym A-B-C-D, which I found effective. (But maybe not too effective, as all I can remember is D is for Data, as in must have lots of it from many contexts.)

Ying Wang @yingwangyw.bsky.social · 23/04/2026

Reposted by Peter Kraft

🎉New preprint: Phenome-derived polygenic scores and social determinants jointly shape context-dependent disease risk. We evaluate disease risk along 2 complementary axes: • how genetic liability is represented • the social context in which it is expressed

157

Peter Kraft @peter-kraft.bsky.social · 23/04/2026

Ooh sorry to miss in person but glad it will be recorded. Having embarrassed myself recently blathering on about how my one undergrad “Theater of Bertolt Brecht” class formed my perspective on AI (thanks umich rc!), it will be nice to hear from folks who’ve actually studied and thought about this.

010

Peter Kraft @peter-kraft.bsky.social · 20/04/2026

Xueyao is presenting another project at #AACR26 (a GWAS of AML) Tuesday afternoon (LB387). Stop by to learn more our npj breast cancer paper and her ongoing research on interpretable AI and other topics! 2/2

000

Peter Kraft @peter-kraft.bsky.social · 20/04/2026

Congrats and kudos to @xueyaowu.bsky.social for leading this collab with Joy Jiang, Graham Colditz, Rulla Tamimi & team—first of 🤞many projects to peek into the black box of AI mammogram risk models and see if they are telling us something about breast cancer epidemiology. 1/2

130

Reposted by Peter Kraft

Arrianna Marie Planey, PhD MA @arrianna-planey.bsky.social · 14/04/2026

So this is a really important point- using AI to "detect" IPV without patients' consent or willing disclosure raises issues - why screen for IPV without a patient-centered framework? - what are health systems doing with these data? - how are patients being protected?

2116

Peter Kraft @peter-kraft.bsky.social · 14/04/2026

No. No I am not. But I’m gonna steal that line: “from practice for practice.”

Screenshot of email subject line: Are you the Peter Kraft who wrote “NLP handbuch für anwender: nlp aus der praxis für die praxis”

000

Peter Kraft @peter-kraft.bsky.social · 13/04/2026

Also, if you’re around Saturday morning, sign up for the #AACR Runners for Research 5k. Always a good time, and–no offence to the Windy City–the weather in San Diego this year will probably be better than last. #AACR26 #Runners4Research donate.aacr.org/events/773

donate.aacr.org

AACR Runners for Research 5K Run/Walk 2026I'm fundraising for American Association for Cancer Research and cancer research!

AACR Runners for Research 5K Run/Walk 2026Event landing page for AACR Runners for Research 5K Run/Walk 2026

010

Peter Kraft @peter-kraft.bsky.social · 13/04/2026

Hey, cancer epi folks: I’m organizing an informal morning run/walk meet up at #AACR26. If you’d like to join, DM me for details!

Peter Kraft @peter-kraft.bsky.social · 13/04/2026

Looking forward to #AACR26 in San Diego in less than a week! Check out these presentations from our team, including the largest ever GWAS of breast cancer & AML and a hot-off-the-presses study of mutational patterns in triple-negative breast cancer.

List of abstracts, including times and program numbers.
.
Apr. 19, 2026, 4:35 PM - 4:50 PM
1381: The Confluence Project: Largest multi-ancestry genome-wide association study of breast cancer identifies 469 susceptibility loci in over two million participants African-ancestry Breast Cancer Genetic Consortium, Breast Cancer Association Consortium. Consortium of Investigators of Modifiers of BRCA1/2, Latin America Genomics of Breast Cancer Consortium, Male Breast Cancer Genetics Consortium, National Cancer Inst. Div. of Cancer Epidemiology & Genetics
.
Apr. 20, 2026, 9:00 AM - 12:00 PM
1481: Integrative analysis identifies potential proteomic intermediates associated with renal cell carcinoma and its risk factors. Ibrahim Hossain Sajal, Andrew J. Song, Kevin M. Brown, Mitchell J. Machiela, Peter Kraft, Stephen J. Chanock, Mark P. Purdue, Diptavo Dutta
2090: Modifiable lifestyle factors and immune gene expression in breast tumor and normal-adjacent tissue. Kristen D. Brantley, Cheng Peng, Clara Bodelon, Deborah A. Tadesse, Peter Kraft, Rulla M. Tamimi

Apr. 21, 2026, 2:00 PM - 5:00 PM
LB387: Genome-wide association study identifies germline susceptibility loci for acute myeloid leukemia. Xueyao Wu, Filip Pirsl, Gabrielle Schmidt, Maryam Rafati, Aurélie Vogt, Herbert Higson, Jia Liu, Jiahui Wang, Shilpa Gaddam, Shengchao Li, Wael Saber, Yung-Tsi Bolon, Steven Moore, Sharon A. Savage, Stephen Chanock, Stephen Spellman, Peter Kraft, Shahinaz M. Gadalla
LB391: Pre-diagnostic exposures, mutational signatures, and immune profiles in triple-negative breast cancer: An overview of the PREMISE-TN project. Deborah A. Tadesse, Clara Bodelon, Cheng Peng, Kristen D. Brantley, Margaux Delporte, Yujing J. Heng, Lauren Teras, Rulla M. Tamimi, Peter Kraft

040

Peter Kraft @peter-kraft.bsky.social · 11/04/2026

A propos moon bases and Mars colonies: “there will be ample provision for the elite,” Dennis Brutus, 1978.

000

Peter Kraft @peter-kraft.bsky.social · 11/04/2026

We think collectively, while acting in the world. Yet the powers that be divide and disempower us. What is to be done?

000

Peter Kraft @peter-kraft.bsky.social · 10/04/2026

I’m here for the Merleau-Ponty and Arendt refs, but defer to O’Connor for the plain-language summary:

Comment by Sinead O’Connor: IT IS NO MEASURE OF HEALTH TO BE WELL ADJUSTED TO A PROFOUNDLY SICK SOCIETY

Peter Kraft @peter-kraft.bsky.social · 06/04/2026

Same! I woke up with an uprooted stop sign in the hall outside my dorm room.

010

Reposted by Peter Kraft

Linda Kachuri @lindakachuri.bsky.social · 18/03/2026

I’m honored and excited to join the Board of Directors! IGES is home to such a vibrant and welcoming scientific community, I look forward to helping it continue to thrive! 💟🧬 Join us for IGES 2026 in beautiful Estérel, QC 🍁 Abstract submission is open until May 30 www.geneticepi.org/2026-annual-...

geneticepi.org

2026 Annual Meeting

0124

Reposted by Peter Kraft

George Davey Smith @mendelrandom.bsky.social · 17/03/2026

Failure to understand that the fundamental principle of Mendelian randomization (MR) is of gene-environment equivalence contributes to the flood of nonsense MR papers that are appearing; Shah Ebrahim, Gib Hemani and I explain why in this short commentary. journals.plos.org/plosmedicine...

journals.plos.org

Gene-environment equivalence: The fundamental principle of Mendelian randomization

In this Perspective, George Davey Smith and colleagues outline how and why gene-environment equivalence, the fundamental principle of Mendelian Randomization (MR), must be properly applied and critica...

13928

Peter Kraft @peter-kraft.bsky.social · 07/03/2026

(Also I’m still sore about my differential geometry prof phoning it in. Left knowing less than when I came in, and not in a good way.)

Peter Kraft @peter-kraft.bsky.social · 07/03/2026

This is the clearest I’ve seen the idea explained, though. When others have explained it I got the impression the manifold was in the low-dim latent space. No, it’s in the original space—makes much more sense.

Peter Kraft @peter-kraft.bsky.social · 07/03/2026

Hm… having mucked around with the linear versions of dimensionality reduction I see the appeal… but noticed a lot of “possibly” and “the hope is” in that paper.

Alicia Martin @genetisaur.bsky.social · 28/11/2025

Reposted by Peter Kraft

Interested in pleiotropy dissection but not sure where to start, which methods are useful, which studies offer illustrative examples, or how to robustly validate your results? Look no further 👀 rdcu.be/eSfAZ

rdcu.be

Dissecting pleiotropy to gain mechanistic insights into human disease

Nature Reviews Genetics - Genome-wide association studies of increasing scale have revealed the prevalence of pleiotropic genetic variants that affect multiple traits. In this Review, the authors...

04317

Peter Kraft @peter-kraft.bsky.social · 08/10/2025

Speaking of ripple effects: any guidance for NIH researchers who have registered for #ASHG25 and have a poster or presentation but may not be able to attend because of the shutdown?

020

Reposted by Peter Kraft

American Society of Human Genetics (ASHG) @geneticssociety.bsky.social · 06/10/2025

@ajhgnews.bsky.social sat with Julie-Alexia Dias, MSc, in the latest "Inside AJHG" to discuss her recently published paper, “Evaluating multi-ancestry genome-wide association methods: statistical power, population structure, and practical implications.”➡️ ashg.org/ajhg/inside-... #ASHG #humangenetics

042

Peter Kraft @peter-kraft.bsky.social · 24/09/2025

I see this playing out in slow motion, accompanied by Barber’s Adagio.

Curious to hear others’ thoughts and experience here! /fin

000

(iv) And the simulations and applications assume individual-level data or in-sample LD is available—typically not the case in large meta-analyses for complex traits. See Wenmin Zhang et al for a discussion of this issue and a possible fix. www.biorxiv.org/content/10.1... 12/n

On a prosaic level, defining discrete genetic ancestry clusters often means excluding participants who don’t fall into any of the clusters, lowering sample size and limiting generalizability. 11/n

(iii) As the authors note, discrete genetic ancestry groups are made-up things. There are a bazillion ways to define clusters of participants by projecting their genotypes into some abstract mathematical space—it’s not clear which (if any) adequately captures variation in genetic effect. 10/n

100

(ii) The simulations are focused on two ancestry groups, with imbalance maxing out at 1:2. In the complex trait setting, it’s not unusual for there to be 4 or 5 ancestry groups, with imbalances on the order of 1:5 or more. 9/n

100

Some caveats and open Qs: (i) The simulations and data applications are focused on the context of molecular QTL (large effects, small sample sizes) not complex traits (small effects, large sample size). Not clear (but plausible) that qualitative results transfer to that setting. 8/n

100

Points (ii) and (iii) may be the flip side of this: in low power situations, the extra degrees of freedom allowing for group-specific effects may cost power. Betting on near similar effects borrows information across groups and improves power here. 7/n

100

…then SuShiE and other methods that allow effects to differ across ancestry groups should be more powerful. On the other hand, if genetic effects are nearly identical (e.g. the causal variant is typed and marginalized GxE and GxG effects are negligible) then pooling should be more powerful. 6/n

Point (i) makes sense: if the genetic effects differ across ancestry groups—perhaps due to linkage disequilibrium differences if the causal variant is not typed or due to subtle differences in marginal genetic effects due to GxE and GxG interactions… 5/n

Quick take-homes: SuShiE outperforms pooled SuSiE—except when (i) the correlation in genetic effects across ancestries is very high (0.99), (ii) sample sizes across ancestries are imbalanced, or (iii) the overall sample size is low relative to the strength of the genetic effects. 4/n

100

We just compared pooled versus stratified analysis of GWAS for locus discovery (pubmed.ncbi.nlm.nih.gov/40902600/), so I was particularly interested in the comparisons of SuShiE and other methods that rely on genetic-ancestry-group-stratified analyses to SuSiE applied to the pooled data. 3/n

121

If fine-mapping, mol-QTL, or [fill-in-the-blank]WAS analyses are your jam, do check this paper out, if only for the nice review and assessment of contemporary multi-ancestry fine-mapping methods. If fine-mapping is not your jam, this is gonna get technical & jargony. 2/n

100