Hena Jose’s Post

📣 ⚡🔬 🧬 What would you do with 1000+ spatial transcriptomics samples with corresponding H&E-stained whole-slide images? Meet HEST-1k, a collection of 1,108 ST samples assembled from 131 public and internal cohorts encompassing 25 organs, 2 species. HEST-1k includes over 1.5 million expression–morphology pairs. 🔍Explore 3 use-cases for HEST-1k: - HEST-Benchmark: Evaluate gene expression prediction from histology across 10 organs and 9 cancer types, testing multiple foundation models for pathology including UNI, and GigaPath. - HEST for discovery: Explore our proof-of-concept for multimodal biomarker characterization using Xenium breast cancer samples. - HEST for fine-tuning pathology foundation models: See how HEST-1k can enhance foundation models for histology with expression-guided fine-tuning. 📄Preprint: https://lnkd.in/dMQvrJxC 👩💻 Code and Data access: https://lnkd.in/drYbW5xW Congratulations to Guillaume Jaume, Paul Doucet and everyone else who contributed to this work. Huge thanks to everyone who helped curate the dataset. #SpatialTranscriptomics #ComputationalPathology #CancerResearch #Bioinformatics

What were the 50 most successful life sciences startups of the past 15 years? 🧬🚀 In this installment of our Pear VC Biotech: Bench to Business series, we conducted an in-depth review of these companies that we dubbed the “biotech behemoths.” 🦍 With a minimum exit value of >$2.7B, these companies represented the top 0.2% of life sciences startups founded in the US, Europe, and Canada between 2009-2023. 📈 Some key findings: 1️⃣ Product Focus 46 of 50 were therapeutics companies, 3 were diagnostics companies, and 1 was a tools company. Oncology was the most common lead therapeutic area (16/46), followed by rare diseases (13/46) and immunology (8/46). 💉 Key platform themes included: cell therapy (Juno Therapeutics, Inc., Kite Pharma, Sana Biotechnology, Inc., Lyell Immunopharma, Arcellx); gene therapy (AveXis, Spark Therapeutics, Inc., Krystal Biotech, Inc., Audentes); CRISPR technology (CRISPR Therapeutics, Intellia Therapeutics, Inc.), and computationally-driven drug discovery (Nimbus Therapeutics, Recursion). 🧑🔬 All three diagnostics companies (GRAIL, Foundation Medicine, and Guardant Health) focus on genomic tests for cancer screening, monitoring, or therapy selection. The sole tools company (10x Genomics) develops instruments and reagents for single-cell omics characterization. 🛠 2️⃣ Founding Profiles The average age of the founding CEO at company formation was ~46 +/- ~10 years. In contrast, for the founding CEOs of the top 50 tech startups (the “tech titans” 🤖) over the same period, the average age at founding was significantly younger at ~36 +/- ~8 years. A little more than half (~53%) of the founding CEOs of these biotechs were first-time CEOs. The PhD was the most commonly held degree (21/49), followed by the MD (15/49), and the MBA (13/49). 🎓 30/50 behemoths had a founder affiliated with at least one academic institution, most commonly Harvard University (7 companies), Stanford University (4), and UCLA (3). The most common founding location was the SF Bay Area (30%), followed by Boston (20%) and Southern California (14%). Although we included Canada and Europe in our survey, only 3/50 behemoths were founded outside of the US. 🌎 3️⃣ Financial Characteristics The behemoths achieved an aggregate value of ~$322B with a total of ~$43B raised (unadjusted dollars), for a rough multiple on invested capital (here simply defined as valuation/total investment) of ~7.5x. 💰 The companies with the highest individual MOICs were Kite (~52.5x), Receptos (~46.2x), Loxo (~30.8x), AveXis (~27.6x), and Foundation Medicine (~26.8x). 💸 The mean time to an initial exit (here defined as a public financing event or an acquisition) was 4.7 +/- 2.7 years. This was considerably faster than it was for the tech titans (8.2 +/-2.1 years)! ⏱ For the full rankings, lots more data and charts, our methods, and our commentary and predictions, please check out the link! 👇 https://lnkd.in/gkfx-mVG

Advanced RAG Patterns for Documents with Embedded Tables 🔥 Don’t run naive chunking strategies on PDFs with tables - they need to be handled with care for good QA performance without hallucinations! `zhaozhiming` shows you different approaches towards building RAG over your docs, using three different solutions: Nougat, Unstructured, and GPT-4o mode in LlamaParse. The first step is having a really good table parser - there’s multiple options available from LlamaParse to Unstructured to Nougat to everything else. As a second step, you can either directly index the chunks or index the summaries and do LlamaIndex recursive retrieval to fetch tabular chunks. This blog post is worth a read for anyone building complex document RAG: https://lnkd.in/gTxch94U If you want more resources check out our LlamaParse client repo! https://lnkd.in/g3UmUkcD

In Nature today, our work on the first foundation model for whole-slide digital pathology. Key advance is it shows strong evidence of scaling benefit. State-of-the-art performance on 17/18 imaging tasks, with 12 showing very significant improvement. https://lnkd.in/gRzZgqba