Seth Stadick, MS

On every bioinformatics projects I’ve worked with, I see two kinds of scientists: 1️⃣ Those who want to do the analysis themselves. They’re not trained bioinformaticians, but they’ll roll up their sleeves, learn the tools, and take responsibility for the results. With the right support and reproducible pipelines, they thrive. 2️⃣ Those who never want to touch the stats, pipelines, or code. They’d rather focus on experiments, hypotheses, and biology - leaving the computational work to specialists. Neither type is “better.” Both are essential. But supporting them requires very different approaches. For the first group, you need dependable, reproducible tools they can run confidently on their own. For the second, you need reliable infrastructure and experts who can deliver reproducible results they can trust. That’s what I carried from my time at UMass into building Via Scientific: reproducibility is the common ground. It doesn’t matter which type of scientist you are - reproducible pipelines make the science stronger for everyone.

260

9 Comments

what you think a productive bioinformatician’s day looks like: • focus time 9a–5p what it actually looks like: • cluster job fail • quick Slack from biology team: “can you re-run?” • daily standup • cross-functional sync • status update on pipeline migration • data wrangling for a meeting you’re not invited to • meeting to discuss the metadata formatting issue • meeting to plan a future meeting about pipeline handoff • final review of deck for science update • an ask for a “quick analysis” • focus time 4:30p–5p delivering impact at scale

219

10 Comments

🧬 Understanding SAM, BAM, and CRAM: Core Alignment File Formats in Bioinformatics 🧬 If you’re working with NGS data, alignment files are unavoidable. SAM, BAM, and CRAM store the same biological information, but choosing the right format impacts performance, storage, and reproducibility. 🔹 What Are SAM, BAM, and CRAM Files? SAM (Sequence Alignment/Map) Plain-text, human-readable alignment format with flags, MAPQ scores, and CIGAR strings. Best for inspection and debugging, but inefficient for large-scale analysis. BAM Binary, compressed version of SAM with identical information. Smaller, faster, and still the most common working format in many pipelines. CRAM Reference-based compressed alignment format storing differences relative to a reference genome. Often smaller than BAM and can be lossless or lossy. 🔹 Key Differences Readability   SAM = human-readable   BAM / CRAM = binary  File Size   SAM → largest   BAM → smaller   CRAM → smallest (~30–60% smaller than BAM)  Performance   BAM and CRAM are far more efficient than SAM. BAM is still the most broadly optimized, while CRAM adoption continues to grow. Reference Dependency   SAM and BAM store full read sequences   CRAM requires the exact reference genome to decode reads 🔹 Why These Formats Matter SAM → Best for inspection and debugging   BAM → Primary working format for variant calling, RNA-seq, and visualization   CRAM → Ideal for long-term storage and cloud workflows 🔹 Practical Tips Convert SAM to BAM early: samtools view -b input.sam > output.bam Always sort and index BAM or CRAM files   Keep index files: BAM → .bai | CRAM → .crai   When using CRAM, always retain the reference FASTA   Use region queries (e.g. chr1:1-100000) for efficiency 🔹 Tool Compatibility Aligners output SAM/BAM; variant callers and genome browsers require sorted, indexed BAM or CRAM. Samtools is the core utility for all formats. 🔹 Pro Tip CRAM can significantly reduce cloud storage costs, but BAM is often safer for data sharing when reference consistency cannot be guaranteed. SAM, BAM, and CRAM represent the same biology and understanding their differences helps you choose the right format for speed, cost, and reproducibility. #Bioinformatics #NGS #SAM #BAM #CRAM

237

9 Comments

We built an API for Edison agents, like we had at FH platform. I wrote up a demo of how to use this for quickly converting powerpoint presentation into claims and then checking them for support/contradiction in the literature. You can get a lot done on our website UI, but you cannot research every item in a long list. Like when we wrote 20k wikipedia articles over the human genome, we did it with the API. Or if you want to run the same analysis over 100 spreadsheets. https://lnkd.in/gsFrMS2h

100

3 Comments

My favorite client calls start with: "Josie, we're not seeing anything in the biomarker data...." Most translational folks want clean data. I prefer the messy cases. Here's what typically happens: 👉 Phase 1 biomarker data comes back. The team reviews the initial plots. 👉 Nothing. No signal. Just noise. But messy data isn't meaningless data. The signal is often hiding in: → Batch effects obscuring biology → Poor/lack of normalization → No appreciation of the limits of the test (limits of quantification, precision etc) → Lack of exclusion of specific samples due to poor handling → Comparison to patients on Standard of Care Even when the cleaning is fixed, teams can get stuck in this process: ✅ A Translational scientist collects the samples and gathers the data ✅ A bioinformatician cleans and plots the data. ❌ But the whole team misses key points such as comparison to competitor benchmarks, clinically relevant subgroup analysis, or the key plot that nails what you're trying to achieve. That's where my dual experience of translational strategy and data science comes in handy. I QC. Clean. Replot. Then interpret what it means for your program—not just your dataset. Suddenly—there's the signal. The predictive signal that transforms the program. The longitudinal pattern that supports an accelerated approval. This work excites me most. Not every biomarker dataset has a hidden signal. But you'd be surprised how often it does—if you know where to look and what it means for drug development.

85

6 Comments

I recently shared some results from transcriptomic data Zafrens acquired in several cell lines with multiple control compounds. Transcriptomes are great, but only half of the picture that we're working to deliver. The other half of the picture is pictures! Imaging assays have long been a staple in discovery campaigns, for good reason: imaging is a direct and potentially non-destructive method to observe perturbational effects that happen on a wide range of time scales. The measurements from an imaging assay are often a direct readout of a stain or reporter e.g. mean fluorescence per cell, nucleus-to-cytoplasm fluorescence distribution. We've built out a decent cell segmentation model & pipeline to provide these direct outputs, but it would also be nice to be monitoring our data for off-target patterns. How, one might ask, could this possibly be beneficial information? I can speculate about a couple of circumstances where I'd be glad to have it. - We're running a screen for which the on-target signal is very weak. A control condition is clearly biased in a specific direction but it's noisy, and we want to improve our SNR. (I suspect this will be the most common case.) - We're running a screen and we identify a strong on-target signal for a set of perturbations. A subset of those seems to have an unusual pattern in a cytoskeletal stain which resembles a set of compounds otherwise known to be lethal to that cell line. I might advocate checking off-target tox data higher in the downstream funnel. - We're running a screen using a positive control with some poorly-described off-target activity we'd prefer to avoid. We don't have a "clean" positive control for contrast. In ranking candidates for resynthesis & validation, we might check if there's any subclusters with structural biases that allow us to see similar but distinct phenotypes downstream. In some experiments we shared last year, we were interested in the METTL3 methyltransferase inhibitors STM2457 and STC15. These two compounds are closely structurally related and while their _in vitro_ potencies against METTL3 are similar, STC15 seemed to have more dramatic effects in our live-cell experiments. The question I wanted to ask is: how well can we identify which of the two compounds is perturbing a given well? With our original visual model, we got ~77% linear-probe balanced accuracy on STC/STM observations not included in model pre-training! These METTL3-related observations were generated a while ago and our bench experiments since moved on to focus on different cell lines, targets, assays etc. Despite our more recent observations being "unrelated" to the METTL3 samples assessed here, the same visual model pre-trained with that additional data boosted the same balanced accuracy to ~80%! I'm super excited to see how much further we can push this and other metrics just by pre-training on even more diverse observations. Heavy credits to Nathan Hedrick & his team for generating the data!

48

2 Comments

Recently, I've been creating #Docker images of some #AI and scientific tools of interest. This helps to get the tool running without completely obliterating my local development environment. Plus, #containerization helps us to scale in the #cloud or in #HPC environments. In my latest blog post, I provide 3 tips for building such images. Give it a read: https://lnkd.in/egf5Sk8S #bioinformatics #genomics #computationalbiology #azure #kubernetes #singularity #containers #datascience #aimodels #llm #diffusion

34

2 Comments

𝐀𝐜𝐚𝐝𝐞𝐦𝐢𝐚 𝐨𝐫 𝐢𝐧𝐝𝐮𝐬𝐭𝐫𝐲: 𝐖𝐡𝐢𝐜𝐡 𝐩𝐚𝐭𝐡 𝐢𝐬 𝐫𝐢𝐠𝐡𝐭 𝐟𝐨𝐫 𝐲𝐨𝐮? I was recently asked to give a talk about transitioning from academia to industry. While preparing, I realized I wanted to share a few raw thoughts and hear yours in return. Fun fact is that I’ve made this transition twice. First, straight after my PhD (or actually, before I even defended). I was then on the verge of burnout, so the decision was not well thought. Thus, after regaining strength, I returned to academia to chase the ‘career in science’ dream, only to discover I didn’t fit. And once again, industry – again, Selvita (thank you!) – offered a helping hand, and here I feel like I'm in the right place. So if you’re finishing your MSc or PhD, you may be facing a big question: 𝐃𝐨 𝐲𝐨𝐮 𝐚𝐜𝐭𝐮𝐚𝐥𝐥𝐲 𝐞𝐧𝐣𝐨𝐲 𝐥𝐚𝐛 𝐰𝐨𝐫𝐤? If not, but you still love science, that’s okay. There are plenty of roles in industry beyond the lab: medical writing, business development, regulatory affairs, and alliance management (to name a few). Most don’t require a PhD – so if you’re sure lab work isn’t for you, skip the extra 4 years! 𝐄𝐚𝐬𝐲 𝐚𝐧𝐬𝐰𝐞𝐫, 𝐫𝐢𝐠𝐡𝐭? 𝐎𝐧 𝐭𝐡𝐞 𝐜𝐨𝐧𝐭𝐫𝐚𝐫𝐲! Still, a PhD isn’t just pipetting. It’s also: 🔹 𝐍𝐞𝐭𝐰𝐨𝐫𝐤𝐢𝐧𝐠 – across labs and at conferences. Learning how to connect and communicate is key in any industry role. I neglected that during my PhD — now I’m making up for it. 🔹 𝐌𝐞𝐧𝐭𝐨𝐫𝐢𝐧𝐠 – supervising junior students teaches people management. The feedback you get (even indirectly) is pure gold. Mind the small age and level gap during a PhD – it’s minimal between you and your wards, which can make the feedback loop more productive! 𝐎𝐟𝐟 𝐭𝐡𝐞 𝐧𝐨𝐭𝐞: I’m happy that my first MSc student is now doing a postdoc. I must’ve done something right… or maybe not bad enough to discourage her from science. 😊 🔹 𝐏𝐫𝐨𝐣𝐞𝐜𝐭 𝐦𝐚𝐧𝐚𝐠𝐞𝐦𝐞𝐧𝐭 – Your PhD project is a four-year journey with both long- and short-term goals. At some point, you'll need to develop the skills to manage them effectively: making tough decisions on which projects to execute or kill (CCP4 logs still in my mind), allocating resources (both personal time and money), managing the people involved, mitigating risks, and addressing the many other factors inherent to project management. That’s it for now, but the list is far from complete. What helped you make your own career decisions? What questions were you asking? #CareerInScience #PhDLife #IndustryRoles #ScienceCareers #NetworkingInScience #BeyondTheLab

45

11 Comments

30 days of NGS tools - Day 30 Galaxy Project We’ve journeyed through 30 tools in 30 days, from mapping reads to finding bugs in bugs. But what if you don’t want to juggle 30 command lines, endless parameters, and random error logs that make you question your life choices? Enter Galaxy Project, the one-stop bioinformatics platform where: No need to install 100 dependencies (and cry when conda breaks). Tools like HISAT2, Trinity, Kraken2, DESeq2 (and MANY more) are all under one roof. You can build workflows visually, share them with collaborators, and reproduce them anytime. Why do we use Galaxy? Accessibility: Click-based, browser-based interface (no need to be a Linux ninja). Integration: Supports NGS, proteomics, metabolomics, and even image analysis. Collaboration: Share workflows and histories, science should be social, not siloed. Reproducibility: Your future self will thank you when you don’t remember your pipeline flags. How to use it Go to usegalaxy.org Upload your data (FASTQ, VCF, GTF, BAM, anything you fancy). Choose a tool -> run it with a few clicks. Chain tools together visually to build your workflow. Export results, plots, and share your workflows with colleagues. Basic code (Well… not really 🤭) Galaxy is GUI-driven, but behind the scenes it executes the same commands you’ve been running all month. For example, # What you run on CLI hisat2 -x genome_index -1 R1.fastq -2 R2.fastq -S output.sam # What Galaxy does Click -> Select HISAT2 -> Upload files -> Run -> grab coffee How does this help researchers? Makes NGS accessible even to wet-lab folks who fear the terminal. Boosts productivity by saving setup and debugging time. Encourages team science by sharing workflows and histories. Perfect for teaching & training (students can focus on biology, not broken installs). Think of Galaxy as the Netflix of bioinformatics, every tool you want is already there, you just need to hit play. No more “dependency hell,” no more “it worked on my machine.” With Galaxy, it just works. #Day30 #NGSToolsChallenge #GalaxyProject #Bioinformatics #NGS #Reproducibility #Workflows #OpenScience #DataAnalysis

76

2 Comments