Janet Matsen

What if you could extract even more accurate results from your existing HiFi WGS data? Improvements such as more accurate small variants (39% fewer FNs and 14% fewer FPs), better structural variants, CNVs, phasing through complex SVs, and easier visualization of supporting read evidence?   These improvements are all within reach, via an experimental global assembly and read mapping protocol we’ve recently been developing. The key to this protocol is a new tool, portello, which transfers read mappings from global assembly contigs to a standard reference, such as GRCh38, enabling high quality read mapping and the ability to run standard read-based variant callers like DeepVariant from mappings improved by assembly. In addition to improving results from read-based variant calling tools, this protocol allows read-based results to be unified with contig-based calls from assembly.   I’ll be describing more about portello and our early results working with this protocol tonight at #gi2025 poster 133 (and I have a few sawfish stickers left too!). For anyone wanting to give portello a try and send us feedback, the method and latest protocol is on github here: https://lnkd.in/gbGq5vgw

121

5 Comments

Partnering with Tamarind Bio (Deniz Kavi) really levels up Biomni with protein engineering capabilities. For biologists who aren't sure how to get started with computational tools - this enables all of us to access frontier protein models without any limitations - no code, just natural language. We're continuing to work on sharing more functionality and capability that empowers every single biologist to be 100x more capable. Read more at: https://lnkd.in/eh6rekWB Try it out today at http://biomni.stanford.edu

45

On creating 'new knobs of control' in biology https://lnkd.in/eNwagDEG One reasonable way you can view how all drugs work is by operating on some 'knob of control'. Statins operate on the 'mimicking the native substrate of HMG-CoA reductase' axis, and so on. Importantly, nearly all known knobs are the ones that evolution allows us access to, entirely by accident. This feels quite claustrophobic. We'd ideally like to be infinitely creative with how medicine works, but it's been historically easier to just use native constructs (antibodies, active-site binders, etc) to accomplish our goals. But there is a brave new world ahead of us. There are several emerging therapeutic modalities that (though they don't advertise themselves that way) seek to install entirely new knobs of control, allowing for some potentially dramatic swings in possible efficacy and/or localization. I cover three such cases: how they work, what they solve, and what their future may look like.

95

Bioengineering is fundamentally limited by scale. Right now, we still engineer biology gene by gene, painstakingly modifying individual sequences. In software, you can simply "git clone" or "pip install" entire functions and reuse them instantly! What is the biological equivalent of importing whole functional systems? Symbiogenesis. When one organism merges and integrates another, it doesn’t just add a gene -it gains a complete functional module. That’s what our ancestors did with mitochondria. The result is significant jumps in complexity, such as the one that made multicellularity, and eventually us, possible. If this idea excites you, and you’re curious about how symbiosis could become a design principle for next generation of bioengineering, consider joining the Symbiogenesis Seminar Series. Let’s explore what happens when evolution’s boldest experiment becomes a tool for engineering! 📅 Details below. Thanks to Lorenzo Magni and Bogumil J. Karas for co-organizing with me, and to Noa Sher, Mariëlle van Kooten, Jeffrey Barrick, Andreas Emilio, and Carlos Lax for presenting!

32

1 Comment

The platform‑control battle, sparked by the release of Anthropic’s Claude Code and Google’s newly free Gemini CLI, is well underway. Google has open‑sourced Gemini CLI and made it free to use, offering up to 60 model requests per minute and 1,000 daily at no cost, clearly aiming to dominate developer workflows. At the same time, many countries are still shaping their AI strategies and digital‑sovereignty frameworks, trying to balance innovation with control. In this environment, tech giants are embedding powerful tools into everyday workflows, empowering users while deepening their dependence. There's no denying these tools are game‑changers by delivering seamless coding experience, content creation, research, and even no‑code app building. But the more pressing question is: as countries rush to define and implement their AI and digital‑sovereignty strategies, are we fully aware of how deeply tech giants are embedding themselves into both everyday users’ workflows and professional environments—even as they shape public dependency and influence national ecosystems? #ConsentInAI #AutonomousAI #AIDependence #AIsovereignty

5

If the math is correct (I am not able to evaluate the math), then this is very good news for those trying to discover hidden features that LLMs (such as protein LMs) learn, and for Explainable AI efforts in general. https://lnkd.in/gNttXciX

2

Everyone in AI/ML knows that models & algorithms soak up attention, while datasets, benchmarks, and evaluations drive progress. It's too early for me to say what the impact will be, but the BioML field needs more data and Basecamp Research's "BaseData" release is an exciting step in the right direction. It was great talking to James Dinneen at New Scientist. Check out the article below: https://lnkd.in/dvAd47Pi

147

5 Comments

The Finding that just redrew the boundaries of what biology can be. David Baker & team came up with- AI-designed antibodies; intentionally crafted, atomically precise, and released openly on GitHub for anyone bold enough to explore. This isn’t an upgrade. This is biology stepping into a new mode: from evolution → to engineering, from probability → to precision, from discovery → to design. So let me ask you this: When AI can generate molecular solutions (Antibody Design in this case), what becomes the uniquely human skill that defines the future in this space? @Institute for Protein Design @Nature Portfolio @OpenAI @DeepMind @Aragen Life Sciences #AI #ArtificialIntelligence #Innovation #Biotech #DrugDiscovery #FutureOfMedicine #ProteinDesign #AntibodyDiscovery #ComputationalBiology #MachineLearning #DeepLearning #LifeSciences

25

1 Comment

This is exactly the kind of showcase that moves PLR from "interesting open-source project" to "something I want in my lab" - Python-native automation that integrates with the tools scientists already use. Great work Rick Wierenga, Jonathan Chen and the entire Retro Biosciences Automation Team!

22

1 Comment

[Part 6/6] – Designing Tech Around Biology Biology is messy. Context-dependent. Full of edge cases. That’s not a bug - it’s the nature of life. If we keep designing systems that assume stem cells will behave like E. coli in a fermenter, we’ll keep hitting the same walls: - Poor translation to clinical outcomes - High batch failure rates - Inflexible processes that don’t fit real-world patients Instead, we need systems that are: - Modular - Flexible - Designed around the constraints and capabilities of the cells This is more than an engineering problem. It’s a mindset shift: Build around biology - or biology will push back. This is what we’re committed to at Green Elephant Biotech. Let’s stop bending cells to fit in systems designed for microbes. Let’s start building systems that fit the cells we actually use.

10

3 Comments

🌱 𝗪𝗵𝘆 𝗟𝗶𝗻𝗲𝗮𝗿 𝗥𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝘀 𝗔𝗿𝗲 𝗡𝗼 𝗟𝗼𝗻𝗴𝗲𝗿 𝗘𝗻𝗼𝘂𝗴𝗵 𝗶𝗻 𝗣𝗹𝗮𝗻𝘁 𝗚𝗲𝗻𝗼𝗺𝗶𝗰𝘀 For years, genomics relied on a single 𝗹𝗶𝗻𝗲𝗮𝗿 𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗴𝗲𝗻𝗼𝗺𝗲 to represent an entire species. But plants taught us a hard truth: 𝗴𝗲𝗻𝗼𝗺𝗲𝘀 𝗮𝗿𝗲 𝗻𝗼𝘁 𝗹𝗶𝗻𝗲𝗮𝗿. 𝗦𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗮𝗹 𝘃𝗮𝗿𝗶𝗮𝘁𝗶𝗼𝗻, 𝗽𝗿𝗲𝘀𝗲𝗻𝗰𝗲/𝗮𝗯𝘀𝗲𝗻𝗰𝗲 𝘃𝗮𝗿𝗶𝗮𝘁𝗶𝗼𝗻, polyploidy, and extensive diversity make plant genomes far more complex than a single FASTA file can capture. This is where 𝗠𝗶𝗻𝗶𝗴𝗿𝗮𝗽𝗵 comes in. 🧬 𝗪𝗵𝗮𝘁 𝗶𝘀 𝗠𝗶𝗻𝗶𝗴𝗿𝗮𝗽𝗵? 𝗠𝗶𝗻𝗶𝗴𝗿𝗮𝗽𝗵 is a fast, lightweight tool for constructing a 𝗽𝗮𝗻-𝗴𝗲𝗻𝗼𝗺𝗲 𝗴𝗿𝗮𝗽𝗵 from multiple genomes. Instead of aligning everything to one reference, it builds a genome graph that represents all genetic diversity across individuals or accessions. In other words: 🚫 It doesn’t compare genomes one by one. 👉 It 𝗶𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗲𝘀 𝘁𝗵𝗲𝗺 𝗶𝗻𝘁𝗼 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗴𝗿𝗮𝗽𝗵 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲. Nowadays, the limitations of linear references in genomics are clear; genome graphs offer a viable solution. 🔍 𝗪𝗵𝘆 ��𝗲𝗻𝗼𝗺𝗲 𝗚𝗿𝗮𝗽𝗵𝘀 𝗠𝗮𝘁𝘁𝗲𝗿 (𝗘𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗹𝘆 𝗶𝗻 𝗣𝗹𝗮𝗻𝘁𝘀) Plant genomes are inherently complex: - Large structural variants - Gene presence/absence variation - Highly diverse populations and landraces A genome graph reflects reality: ✔ Multiple paths ✔ Alternative alleles ✔ Missing or extra genomic segments This makes graph-based representations far more suitable for plant population genomics, pan-genome studies, and breeding research. 𝗛𝗼𝘄 𝗠𝗶𝗻𝗶𝗴𝗿𝗮𝗽𝗵 𝗪𝗼𝗿𝗸𝘀 (𝗖𝗼𝗻𝗰𝗲𝗽𝘁𝘂𝗮𝗹𝗹𝘆) 1️⃣ 𝗖𝗼𝗺𝗽𝗿𝗲𝘀𝘀𝗶𝗼𝗻 Each genome is compressed to allow fast and efficient comparison. 2️⃣ 𝗣𝗿𝗼𝗴𝗿𝗲𝘀𝘀𝗶𝘃𝗲 𝗚𝗿𝗮𝗽𝗵 𝗖𝗼𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻 Genomes are added one by one. Whenever a difference is detected, it becomes a new branch in the graph. 3️⃣ 𝗩𝗮𝗿𝗶𝗮𝗻𝘁 𝗗𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻 Minigraph detects: SNPs, Indels, Large structural variants (SVs); often missed by linear-reference methods 4️⃣ 𝗚𝗿𝗮𝗽𝗵 𝗠𝗮𝗽𝗽𝗶𝗻𝗴 New genomes or assemblies can be mapped directly to the graph, offering higher accuracy than mapping to a single reference. 𝗪𝗵𝘆 𝗧𝗵𝗶𝘀 𝗠𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝗣𝗹𝗮𝗻𝘁 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 For plant bioinformatics, Minigraph enables: - Building species-wide pan-genomes - Capturing structural diversity missed by traditional pipelines - Improving population genomics and diversity studies - Supporting more accurate breeding and trait association analyses Instead of asking: “𝗛𝗼𝘄 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗶𝘀 𝘁𝗵𝗶𝘀 𝗽𝗹𝗮𝗻𝘁 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲?” We can now ask: “𝗛𝗼𝘄 𝗱𝗼𝗲𝘀 𝘁𝗵𝗶𝘀 𝗽𝗹𝗮𝗻𝘁 𝗳𝗶𝘁 𝗶𝗻𝘁𝗼 𝘁𝗵𝗲 𝗳𝘂𝗹𝗹 𝗴𝗲𝗻𝗼𝗺𝗶𝗰 𝗹𝗮𝗻𝗱𝘀𝗰𝗮𝗽𝗲 𝗼𝗳 𝘁𝗵𝗲 𝘀𝗽𝗲𝗰𝗶𝗲𝘀?” Tools like Minigraph help us move closer to biological reality and away from oversimplified representations. For More Det: https://lnkd.in/djnjDwke

11

I wanted to share a small personal project I have been working on: an end-to-end machine learning pipeline for protein–ligand binding prediction. The question I wanted to explore was: Can protein foundation model embeddings, combined with molecular fingerprints, help predict whether a small molecule is likely to bind a protein? In this project, I used ESM-2 to generate protein sequence embeddings and combined them with ligand representations such as Morgan fingerprints, MACCS keys, atom-pair fingerprints, and RDKit molecular descriptors. These combined protein–ligand features were then used to train binary classifiers for binding prediction based on KIBA scores. One of the biggest lessons I learned was that evaluation strategy matters as much as model choice. With a random split, a model can easily benefit from leakage by seeing the same proteins or related ligands during training and testing. That can make performance look stronger than it really is. To make the evaluation more realistic, I tested several split strategies: • cold-protein: unseen proteins in test • cold-ligand: unseen ligands in test • scaffold split: unseen chemical scaffolds • cold-both: unseen proteins and ligands This changed how I thought about the problem. The key question is not only “Can the model predict similar known pairs?” but also “Can it generalize to new proteins and new molecules?” I compared several downstream models, including Logistic Regression, Random Forest, XGBoost, LightGBM, and an interaction MLP. Interestingly, with good representations and careful splitting, several model classes performed comparably well. My main takeaway: In biology and drug discovery ML, representation quality and evaluation design can matter as much as the final model architecture. Foundation models can provide powerful protein representations, while molecular fingerprints capture complementary chemical information. Yet without careful dataset construction and leakage-aware testing, it is easy to overestimate model quality. As foundation models become more widely used in biology, I think this point will become increasingly important: we need to test whether models are learning generalizable biology and chemistry. Otherwise the models would just memorize the patterns in the dataset. Project link: https://lnkd.in/gh6Ymgu2 #MachineLearning #DrugDiscovery #ComputationalBiology #Bioinformatics #ProteinLanguageModels #FoundationModels #Cheminformatics #AIforScience

467

27 Comments

Tech discourse is full of people who mistake strong opinions for deep knowledge, then position themselves as the wise ones warning everyone else. 📢: «You don’t understand these systems!» => Proceeds to mischaracterize both artificial AND biological information processing🤦 The larger point is semi-valid but its heuristic value is completely obliterated by the philosophical overconfidence and conceptual sloppines. THE VALID: • AI capabilities are routinely oversold • Many tech leaders lack deep understanding of their tools • Statistical pattern matching has real limitations • Business decisions based on AI mysticism will fail THE INVALID: the rest of the post. This is exactly why precision matters in technical discourse. When you mix valid engineering concerns with overconfident claims about consciousness, you poison the entire message. People either accept the whole package uncritically or reject it entirely - and neither serves truth! P.S. You can’t simultaneously position yourself as having superior technical knowledge that tech leaders lack while retreating to “it’s just common sense opinion” when asked to defend that knowledge. The OP’s original post claimed authoritative insight into consciousness, learning mechanisms, and AI architectures - presenting these as facts others ignore at their peril. When challenged, the author went from “intellectual superiority” to “you are mean to my opinions” in record time. Don’t do that. It’s intellectually dishonest. If these were merely personal convictions not requiring rigorous support, then the tech leaders under fire have equal right to their intuitions about AI capabilities. The issue isn’t the venue - it’s the intellectual consistency. Either you’re making knowledge claims that can be evaluated, or you’re sharing misguided opinions that carry no special weight. You can’t invoke scientific authority to criticize others while exempting yourself from the standards that create that authority in the first place. #ArtificialIntelligence #MachineLearning #LanguageModels #CognitiveScience #IntellectualAccountability #TechnicalDiscourse #ScientificRigor #CriticalThinking #LanguageAcquisition #LLMs

8

6 Comments

Why do I love Bioinformatics and Computational Biology? We get to work and collaborate to solve a broad range of real-world problems, from reducing cattle methane emissions to discovering new biomarkers for diagnosing complicated diseases. Biotech, Pharma, Agro, healthcare, conservation biology, even astrobiology—none of these would exist without bioinformatics and data science. Like many other sciences, biology has become a highly quantitative and digital field. Twenty years ago, in 2005, Roche introduced pyrosequencing technology to the world, followed by Illumina with their short read NGS the next year, and I knew my work was about to change. From qualitative research (protein annotation, comparative genomics), my focus shifted to quantifying certain bacterial species in ocean water, trying to understand the impact of tourism and other natural phenomena in the beautiful Arraial do Cabo - RJ coastal waters (picture bellow). The beauty of having a high-throughput method for sequencing DNA was that now we could explore possibilities that had only been a dream a year earlier. What if we could re-use this huge amount of data and mine it for genes possibly encoding new antibiotics or other compounds of interest for pharma and biotech? What if we applied this approach to monitor wastewater for pathogens? Maybe we could predict outbreaks even before the first cases start to be notified. It didn’t take long for scientists around the world to start inventing new applications and derived methods. RNA-seq brought us the quantification of gene expression. Suddenly, we could understand biology with unprecedented detail. Methods for methylation analysis came along, allowing us to understand cell differentiation, cancer evolution, and other fascinating processes. From this to liquid biopsies didn’t take that long. Now, a blood test can detect DNA with methylation patterns that can help us identify cancer before any visible signs of a tumor appear. All these innovations were the fruits of intense collaboration between amazing bench scientists (molecular biologists, biochemists, physicists, etc.) and bioinformaticians/computational biologists who developed the tools and methods for data analysis. But what is really exciting? Those were just the tip of the iceberg. There is a world of novelty out there, waiting to be discovered and developed. And that is what keeps me going.

76

2 Comments

When I was running the bioinformatics core at UMass, we saw the “it works on my cluster” pattern constantly: A researcher built a working pipeline tailored to their needs, their files, their laptop, their assumptions. It ran. It gave results. It moved the project forward. But when someone else tried to re-run it? Forget about it It technically worked but only in that moment, for that person, on that setup. These types of temporary wins are great in the moment, but they don’t build organizational capital. At Via, we treat pipelines as infrastructure built to last across teams, projects, and time. That means: ✅ Containerized environments ✅ Modular, versioned processes ✅ Run tracking and metadata logging ✅ Designed handoffs ✅ Clear boundaries between code, config, and data The real test isn’t “did it work once?” It’s “Can we run it again, with confidence, six months later, when the author is gone?” That’s what production-grade means and that’s what we’re building toward.

30

1 Comment

One thing that always bugged me during my PhD was how many different tools you need just to sanity-check a genomics file. samtools for BAM, bcftools for VCF, head for CSV, R for H5AD. Every format, different tool, different flags. So I built peek-bio. One command, any format. It auto-detects the file type and gives you a structured summary with QC warnings. Works on CSV, BED, FASTA, FASTQ, VCF, GTF, SAM/BAM, Excel, BigWig, and H5AD. pip install peek-bio peek yourfile.bam Open source and on PyPI: https://lnkd.in/gJR4zjqi #bioinformatics #genomics #opensource #python

58

4 Comments