iiinova — where your digital self keeps evolving

scAmbi: Correcting mapping-ambiguity overdispersion in scRNA-seq

Since v0.2.0: catchSalmon-style pooled estimator with EB moderation, concise function names, and a leaner, focused scope. Vignette is here: seantbresnahan.com/scAmbi/

seantbresnahan.com

Bootstrap-based overdispersion estimation, correction, and reference comparison for single-cell RNA-seq.

000

scAmbi pools Alevin bootstrap replicates across cells to estimate per-transcript technical overdispersion from mapping ambiguity, moderates it with empirical Bayes, then corrects counts or supplies GLM offsets. A meanInfRV diagnostic enables reference comparisons like this one.

100

I quantified 3' and 5' PBMC libraries against the TRAILS long-read isoform atlas and GENCODE v49. Uncertainty is nearly identical in 3', where reads land in terminal exons shared across isoforms, but clearly lower with TRAILS in 5', whose reads span splice junctions resolved by long-read sequencing.

GitHub - sbresnahan/scAmbi: Correct read-to-transcript mapping ambiguity in scRNA-seq

Generating some data for a grant required returning to a side project... scAmbi, my R package for correcting single-cell RNA-seq transcript abundance estimates for overdispersion from mapping ambiguity, is now at v0.3.2. Looks like tissue-specific references help 5' 10X. github.com/sbresnahan/s...

github.com

Correct read-to-transcript mapping ambiguity in scRNA-seq - sbresnahan/scAmbi

120

Sean Bresnahan @seantbres.bsky.social · 09/09/2026

Working on a series of videos, “Epigenetics in development and evolution,” exploring evolution of gene regulation, mechanisms, developmental and health impacts, omics-wide observational and experimental study designs, causal inference, and a focus on the placental epigenome.

051

Sean Bresnahan @seantbres.bsky.social · 06/09/2026

🧬🫄🧪 I’ve been inspired by @3blue1brown.com for a long time, but always wished someone was making educational content in that style on epigenetics, evolution, and human development. I’ve started “Epialleles” to do just that! Made with @manim.community youtu.be/UhOSPa2sDfw?...

youtu.be

Hello world!

YouTube video by epialleles

1132

Very special thanks to my coauthors: Charis Xiong, @taylorhead.bsky.social, @ytchang11.bsky.social, @arjunbhattac.bsky.social, and @jonhuang.bsky.social. A very fun first foray into biostats software development!

140

ICONIC: Causal Discovery and Diagnostics for Multiomics

For a visual introduction, here's a short @manim.community explainer covering the ICONIC framework, the identification problem in observational omics, and how the data-calibrated simulation engine generates realistic benchmarks from your data. Music by me 🙂 www.youtube.com/watch?v=aAP6...

youtube.com

YouTube video by Sean Bresnahan

151

GitHub - sbresnahan/iconic: Causal Inference for Multiomics

Full package source and vignettes with worked examples are available here: github.com/sbresnahan/i... Bug reports and feature requests are welcome as we await Bioconductor review and prepare an initial release. If you'd like to collaborate on the project, please reach out!

github.com

Causal Inference for Multiomics. Contribute to sbresnahan/iconic development by creating an account on GitHub.

100

ICONIC also includes a planning tool that works before you collect instruments or controls. It projects bias, confidence-interval coverage, and effect estimates across a range of instrument strengths and control qualities, so you can decide whether collecting additional data is worth the cost.

Prospective analysis showing expected gains from adding genetic instruments and negative controls. (A) Bias in the direct-effect estimate versus exposure-instrument strength for three instrument-using estimators, compared against the confounded unadjusted estimate. (B) Prospective direct and indirect effects at target instrument strength; dashed lines show true effects, dotted shows confounded estimates. (C) 95% confidence-interval coverage versus instrument strength. (D) Bias versus negative-control coverage at target strength with perfectly exogenous instruments, for six estimators. The recommended estimator is chosen by robustness over the full sweep, not by performance at a single best-case point.

On real data, the choice of estimator changes the results. In GUSTO (gestational diabetes and birth weight via placental gene expression) and TCGA (smoking and lung cancer survival via tumor gene expression), the naive and recommended estimators identified largely disjoint sets of mediating genes.

Figure 5, panels B and D. (B) GUSTO case study: scatter of the indirect effect under the instrumented IV2SLS2 estimator versus the naive unadjusted estimator for 5,060 genes. Points colored by significance class: not significant under either method (grey), significant only under the naive estimator (orange), only under the instrumented estimator (blue), or under both (green). (D) TCGA lung cancer case study: scatter of the indirect effect under the recommended proximal g-computation estimator versus the naive estimator for 3,674 genes, on the restricted mean survival time scale in years. In both panels, the naive and recommended estimators identify largely disjoint sets of mediating genes.

This design was motivated by a finding: generic plasmode benchmarks can rank two estimators as indistinguishable, but benchmarks calibrated to real omics data reveal a preference that shifts with how well negative controls capture unmeasured confounders. The best estimator depends on your data.

100

The core methodological contribution is the sensitivity analysis. A generative model learns the statistical properties of your data, then ICONIC maps how each estimator's bias changes as you vary assumptions that can't be tested, like how well instruments and controls capture unmeasured confounders.

Causal directed acyclic graph for the ICONIC mediation framework. Exposure X affects mediator M and outcome Y both directly (the natural direct effect) and indirectly through M (the natural indirect effect). Genetic instruments G1 and G2 instrument X and M respectively, assumed independent of unmeasured confounders. Unmeasured confounders U bias the X-to-M and M-to-Y paths (red arrows). A negative-control panel W proxies the confounder composites (orange dotted arrows). Red dashed arrows denote violations: instrument-confounder correlation. Covariates C affect X, M, and Y throughout.

biocstaging.r-universe.dev

ICONIC brings causal inference methods together in one place, with support for continuous, survival and binary outcomes, diagnostics, data-calibrated sensitivity analyses, and a planning tool predicting whether collecting instruments/controls is worth the cost. biocstaging.r-universe.dev/iconic

iconic: Causal Model Selection with Genetic Instruments and Negative Controls

Provides a model selection workflow for causal inference with genetic instruments and negative controls in observational omics data. The package fits eight estimators of the natural direct and indirec...

One in five observational omics papers overstates causal claims (with caveats)

An open-weight LLM screened ~1K observational omics papers (with molecular feature ➡️ disease associations) with a rubric: ~20% overstate causality; only 4% used a formal mediation analysis. The methods live in separate literatures and haven't been widely adopted. substack.com/home/post/p-...

substack.com

An open-weight LLM graded the causal language of about 1,000 observational omics papers. One in five makes causal claims its design cannot support. Human validation is ongoing.

100

Unmeasured confounding complicates mediation analysis in observational omics. ICONIC addresses this in four stages: diagnose, estimate, stress-test, & recommend. It combines identification strategies with a data-calibrated simulation engine. To see where the field stands, I screened the literature:

The ICONIC framework. (A) For a single genetic variant, the unadjusted effect estimate versus the estimate adjusted for environmental context, in two environments. Unmeasured confounding inflates the estimate in one environment but masks it in another. Projected to 5,000 features: unconfounded effects lie on the 1:1 diagonal and replicate; confounding-driven associations are strong in discovery but absent in replication. (B) Unmeasured confounders bias the exposure-mediator and mediator-outcome paths, corrupting both direct and indirect effects. Three routes to identification: genetic instruments (MR), negative controls (proximal inference), and their combination. (C) The ICONIC workflow: Diagnose, Estimate, Stress-test, Recommend. (D) Two simulation modes: structural benchmark with known ground truth and data-calibrated generative texture learned from the user's own cohort, enabling stress-testing and prospective design decisions before data collection.

ICONIC: An R Package for Integrating Instrumental Variable- and Negative-Control-Informed Causal Discovery and Diagnostics in Multiomic Studies

Our new preprint introduces ✨ICONIC✨: an R package making causal inference accessible for observational omics. It unifies genetic instruments (Mendelian randomization) and negative controls (proximal inference) for mediation analysis under unmeasured confounding. www.medrxiv.org/content/10.6...

medrxiv.org

Unmeasured confounding threatens causal inference and replicability in observational multi-omic studies across variable environments. Genetic instrumental variables (Mendelian randomization) and negat...

1115

Sean Bresnahan @seantbres.bsky.social · 24/06/2026

“And therefore we should have no good things” says Mike Johnson, who has never known a good thing in his entire life because he’s incapable of seeing it anywhere

000

Reposted by Sean Bresnahan

Jeffrey Brainard @jeffreybrainard.bsky.social · 20/05/2026

U.S. researchers who coauthor research papers with scientists affiliated with foreign institutions are drawing fresh scrutiny from NIH, NASA. #scientificpublishing @science.org www.science.org/content/arti...

science.org

U.S. researchers face new restrictions on publishing with foreign collaborators

NIH, NASA grantees are confused and concerned amid agencies’ piecemeal communication

23623

Sean Bresnahan @seantbres.bsky.social · 07/05/2026

I ❤️ @cshlmeetings.bsky.social @cshlnews.bsky.social! Beautiful place. One of those places that surrounds you with the feelings of wonder/purpose/etc that made you fall in love with science. And there’s a bar!

Man with beard and tattoos smiles and points to his scientific poster in an auditorium with walls lined with photographs of notable scientists, including the late great Barbara McClintock.

0101

Sean Bresnahan @seantbres.bsky.social · 06/05/2026

Shout-out to @taylorhead.bsky.social who is presenting prescient work for TWAS studies at #bog26 today! Using long-read RNA-seq to inform isoform-level priors, this work enables Bayesian fine-mapping of causal isoforms by integrating tissue-specific annotation and structural similarity. Poster 108!

050

Reposted by Sean Bresnahan

Taylor Head, PhD @taylorhead.bsky.social · 06/05/2026

Shout-out to @seantbres.bsky.social who will be presenting some really interesting work at #bog26 this evening! Using long-read, isoform-resolved placental transcriptomics, this work shows how PFAS exposure impacts placental transcriptional networks and imprinting (poster #70)

072

Placenta Transcriptome Visualization

Huge thank you to @plbaldoni.bsky.social & @robp.bsky.social for InfRV tools, and collaborators across Singapore, Canada & USA. Also to coauthor (and talented undergrad at Rice U BioSciences) Aryun Nemani for a lovely Shiny app for our visualizing our assembly seantbresnahan.com/lr-placenta-...

seantbresnahan.com

131

If there's one takeaway, we hope it's this: tissue-specific annotation works through BOTH discovery (adding novel isoforms) AND filtration (removing irrelevant ones). Neither alone is sufficient, and annotation size doesn't explain the improvement — it has to be tissue-matched.

100

GDM-birth weight mediators include novel CSH1 (placental lactogen) isoforms with intron retention events that differ by ancestry, supporting population-specific placental endocrine regulation.

100

Applying this reference to short-read data from two multi-ancestry birth cohorts (GUSTO, n=200; Gen3G, n=152) reduced isoform quantification uncertainty ~30% vs GENCODE v45. And it revealed that placental transcription mediates ~36% of gestational diabetes effects on birth weight.

Fig 3: Overview of multi-cohort placental RNA-seq analysis framework for differential expression and mediation analyses across GENCODEv45, lr-assembly, and GENCODE+; violin plots of total isoform expression (sum log TPM) per sample in Gen3G (n=152) and GUSTO (n=200); box plots of mean and variance in isoform expression across shared, novel, filtered, and absent transcript categories during discovery and filtration phases; raincloud plots of inferential relative variance (InfRV) in isoform quantification per annotation with diamond indicating mode; UpSet plot of differentially expressed genes and transcripts across cohorts and annotations; and forest plot of GDM-birth weight mediation effect estimates (total, indirect, and proportion mediated) via isoform expression principal components in GUSTO, colored by annotation.

100

The placenta is NOT a transcriptomic void, contrary to previous reports. Using lr-RNA-seq (n=72, the largest placental LR dataset to date), we built a reference of 37,661 high-confidence isoforms (~40% novel) across 12,302 genes (~22% novel). Breadth & complexity rival other adult tissues.

100

Arjun Bhattacharya @arjunbhattac.bsky.social · 26/03/2026

🎉 It's published! Our placental long-read transcriptome is now in @natcomms.nature.com! Thank you to @arjunbhattac.bsky.social, @jonhuang.bsky.social, and @mikelove.bsky.social for collaborating on this first project of my postdoc @mdanderson.bsky.social. A recap 🧬🫄🧵 www.nature.com/articles/s41...

Fig 1: Study design and transcript discovery pipeline showing Nanopore cDNA libraries from villous placenta (n=72 term births, 36 controls and 36 GDM-affected); comparison of annotated features between GENCODE v45 and lr-assembly showing 63.5% reduction in isoforms and 73.1% reduction in genes; transcript distribution by structural category (FSM, ISM, NIC, NNC, and other classes) for all and high-confidence isoforms; transcriptional breadth across 15 GTEx tissues and cell lines; isoforms detected at increasing expression thresholds with placenta shown as thick black line; and transcriptional complexity as mean isoforms per gene (±1 SD) with placenta maximum of 108..

2198

Reposted by Sean Bresnahan

Another preprint from our group @mdanderson.bsky.social led by talented postdoc @seantbres.bsky.social! Joint with @jonhuang.bsky.social, exploring the intersection of environmental toxins, maternal/fetal health, and placental txomics. Tweet thread below!

074

I’ll be sharing some of this and more @sriwomenshealth.bsky.social on Friday - if you’re here in San Juan, come chat! www.biorxiv.org/content/10.6...

biorxiv.org

011

10/ Critically: this is detectable only at the isoform level. Gene-level aggregation masks these effects. Together, this helps explain heterogeneous PFAS perinatal effects: compounds with greater fetal exposure engage fundamentally different transcriptional architectures in an outcome-specific way.

120

9/ But the outcomes diverge on centrality and compartmentalization. For birthweight only, mediators shift closer to network hubs AND maternal vs. fetal mediators occupy increasingly distinct network positions as TPTE increases.

100

8/ As TPTE increases, fetal mediators grow more numerous and more tightly co-expressed for both 9/ outcomes. Direct fetal exposure systematically recruits larger, more synchronized transcriptional programs regardless of which perinatal outcome is affected.

100

7/ So, leveraging TPTE as a natural experiment, we asked: does increasing fetal dose reorganize the transcriptional networks mediating PFAS effects? We examined 4 network properties: mediator count, co-expression strength, network centrality, and maternal-fetal compartmentalization.

Outcome-specific scaling of transcriptional mediation properties with transplacental transfer efficiency (TPTE). Relationships between TPTE and four transcript-level mediation properties: mediator count, co-expression strength (peak |kME|), network centrality (inverse of the peak distance between mediators and network hubs), and network compartmentalization of maternal versus fetal mediators. Points represent linear regression slopes with 95% confidence intervals (error bars). Filled points indicate statistically significant associations (p < 0.05); open points indicate non-significant trends. Pink and blue distinguish exposure source (maternal versus fetal, respectively). Black is compartmentalization.

6/ We found maternal and fetal concentrations are correlated for low-TPTE compounds but uncorrelated at high TPTE. And maternal vs. fetal differential expression effect estimates become increasingly negatively correlated with TPTE. Thus, we see the placenta responding to both exposures differently.

Maternal-fetal correlation of transcript, but not gene-level effect estimates decrease with PFAS transplacental transfer efficiency. Pearson correlation coefficient (r) between maternal and fetal differential expression log2 fold changes (mean ± SE across transcripts) plotted against TPTE for each PFAS compound and analysis level.

100

5/ For TPTE to work as a natural experiment, maternal and fetal exposures must reflect distinct biological signals, not just track each other.

Scatter plots with regression and confidence intervals between maternal and fetal (cord blood) PFAS measures at delivery.

100

4/ There are numerous PFAS compounds, that elicit different effects on perinatal outcomes. To study this heterogeneity, we leveraged transplacental transfer efficiency (TPTE, likely influenced by chain length, <0.5 for PFOS to >2.0 for PFBS) as a window into how fetal dose shapes placental response.

TPTE, measured as the cord:maternal blood concentration ratio, increases with decreasing PFAS carbon chain length.

100

3/ We performed transcriptome-wide mediation analysis and found PFAS effects on both birthweight and gestational age at delivery operate through co-expression network hubs, not transcripts with the largest fold-changes. Important, because most exposure studies & transcriptomics assays focus on DEGs.

PFAS effects on birth weight are mediated through co-expression network hubs rather than differentially expressed features.
(A) Conceptual framework: PFAS measured in maternal delivery blood and fetal cord blood influence birth weight and gestational age through placental transcriptional responses. Mediation analysis tests whether effects operate through differentially expressed transcripts (DETs) or co-expression network hubs.
(B) Average direct effects (ADE) of PFAS on birth weight vary by compound and exposure source. Maternal exposures generally showed larger direct effects than fetal exposures.
(C) Density distributions of absolute average causal mediation effects (|ACME|) for significant mediators vary by PFAS compound and exposure source (fetal = blue, maternal = pink). Mediator counts and effect sizes differ substantially across compounds.
(D) Hub transcripts that were significant mediators (n = 583) showed markedly higher |ACME| values than DETs (n = 437; p < 0.001), while hub non-mediators (n = 16,313) were statistically similar to DETs. This demonstrates that PFAS effects on birth weight operate primarily through perturbation of central co-expression network nodes rather than transcripts showing the largest fold-changes. This pattern held at the gene level and generalized to gestational age as an outcome.

100

2/ and patient-derived placental explant experiments. Together these improved concordance between experimental and observational PFAS differential expression effect estimates, & detected more PFAS responsive isoforms than genes. Better measurement is a foundation for more reliable causal inference.

Long-read placental transcriptome assembly improves concordance between experimental and observational PFAS effect estimates.
(A) Study design integrating the GUSTO prebirth cohort (n = 124 placental tissue samples; up to 101 with fetal or maternal PFBS measures) with patient-derived placental explants treated with PFBS at 5 µM and 20 µM for 24 hours (n = 18; n = 3 per group), both analyzed against an isoform-resolved placental transcriptome reference.
(B) Log odds ratios for overlap enrichment between concordant PFBS-associated differential expression in explants versus tissue. Gene-level concordance was similar regardless of reference annotation, but isoform-level concordance was markedly improved using the long-read assembly (6.7–8.0-fold enrichment) versus GENCODE (4.3–4.4-fold), for both fetal and maternal exposures.
(C) Isoform-level analysis detected 2–5 times more differentially expressed features (|log2FC| > 1, FDR < 10%) than gene-level analysis across all eight PFAS compounds, for both fetal and maternal exposures. Differential expression was consistently more extensive for maternal than fetal exposures.

1/ Causal mediation with observational omics data is still in development. In the absence of robust negative control methods (an area we are actively working on!), we used two tools: our long-read placental transcriptome reference (in press @natcomms.nature.com www.biorxiv.org/content/10.1...)

biorxiv.org