Zetta Venture Partners

Congratulations to NOETIK on the release of OCTO-VirtualCell, the first multimodal foundation model for predicting gene expression at single-cell resolution in intact patient tissue. Incredible work from the team - Ron Alfa, Jacob Rinaldi, Lacey Padrón, Daniel Bear, Yubin Xie, and everyone else involved. Check out the technical report and interactive tool with the links below. The future is #spatial and #multiscale.

Big week for Guardrails AI! Yesterday, we launched two new state-of-the-art guardrails for PII detection and jailbreak detection 🚀 For PII, we achieve a whopping 76% better(!!) F1 score compared to Microsoft Presidio averaged across all PII categories on the benchmark. For Jailbreaking, we're state of the art of accuracy and F1 compared to Llama Prompt Guard, Microsoft Prompt Shield and the leading open source jailbreak detection models. That's not all though -- the best thing about both of these new state of the art guardrails is that they're fully open source, and available on Guardrails Hub today! Link to blog: https://lnkd.in/gqWPHg4F PII Guardrail on Hub: https://lnkd.in/gjXVBrPd Jailbreak Guardrail on Hub: https://lnkd.in/g3HnYekj

this podcast is a great way to meet my partner Apoorva Pandhi if you haven't already - he's the best!!

Apoorva Pandhi is a Managing Director of Zetta Venture Partners. Apoorva loves the complexity and thrill of the 0 to 1 startup phase and feels fortunate to work with deeply technical and product-centric founders. He finds them tackling the hardest challenges in AI/ML, data infrastructure and cloud-native infra. As a Managing Director at Zetta, he is excited to partner with founders who are ready to push through and beyond the complex chaos of this early stage to unleash greater potential and market impact. Apoorva's fascination for complex problems is rooted in his love for math and statistics as his mother, a mathematician, always challenged him, pushing him beyond his comfort zone. This led him to explore technology as the next frontier, starting with an engineering degree from IIT Delhi. Over the years, Apoorva has experienced this frenzied energy, firsthand. He has been in the trenches with early stage engineering-first startups in different modalities. He first met with several engineering-first companies as an investor at Lightspeed Ventures and then Foundation Capital. He has invested in data and ML infra companies like Outerbounds (OSS metaflow), Electric Sheep, Motherduck (OSS DuckDB), Tabular (OSS Iceberg), and Fennel, amongst others. He was also a co-founder of his own company, Funnl Labs, in the NLP/ML space and a GTM/sales leader at Quid. In addition to his undergrad from IIT, Apoorva holds an MBA from the Wharton School, University of Pennsylvania. Topics: - The "0 to 1" phase in technical startups - The evolution of data infrastructure - Unbundling the data stack #AIInvesting #DataInfrastructure #AIEngineering #VentureCapital #MLOps

Headed to Vancouver for #NeurIPS2024! Comment or DM if you want to catch up... Or, come to my panel - tomorrow (Thursday) 7:30 PM PT, I'll be joining a discussion on Building an AI Startup with Arash Afrakhteh (Pear VC), Vijay Reddy (Mayfield Fund), Jon Chu (Khosla Ventures), and Vibhor K. (Coatue). We'll be in East Meeting Room 1-3. Stop by and say hi!

Join us this Thursday @ 11 AM PT as Apoorva Pandhi sits down with Grace Gong over LinkedIn Live to discuss AI - the current state, what the future may hold, plus investing in the space. See you 12/12 on the Venture with Grace page!

OH in founder pitch today: "It is an order of magnitude more difficult to refute nonsense than it is to create nonsense in the first place." Yep!

Great work by our summer intern Chris Bradbury describing a new "Fourth Wave" of computational drug design!

I've known Phillip Liu since he was a pillar of the Facebook Data Infrastructure team, and admired from afar his work at SignalFX as one of the pioneers of modern cloud observability. Three years ago, he saw the rise of Generative AI as an opportunity to bring those same observability principles from datacenters to compliance; his key insight is that Gen AI allows unstructured data, not just log data, to serve as sources of telemetry. Since then, Trustero's team have delivered products that help GRC teams move faster and more strategically with continuous assurance. But it's still early days in helping companies, not just datacenters, be more trustworthy and reliable. So it's great news that Trustero is announcing the company's Series A to accelerate their progress - and a big welcome to Bright Pixel!

There’s been a lot of progress in 𝗱𝗲 𝗻𝗼𝘃𝗼 𝗽𝗿𝗼𝘁𝗲𝗶𝗻 𝗱𝗲𝘀𝗶𝗴𝗻 ... but how far are we from 𝗱𝗲 𝗻𝗼𝘃𝗼 𝗱𝗿𝘂𝗴𝘀? The industry seems to be all-in on antibodies as the first real-world test of generative drug design and there's been some promising work from the Baker Lab and others. But this pre-print from the team at Nabla Bio is the clearest evidence I've seen of AI generating therapeutic-grade antibodies (from scratch) which feels like a BIG step forward for the field! Using a new method called 𝗝𝗼𝗶𝗻𝘁 𝗔𝘁𝗼𝗺𝗶𝗰 𝗠𝗼𝗱𝗲𝗹𝗶𝗻𝗴 (𝗝𝗔𝗠) they show they can generate de novo antibodies with strong affinity, specificity and function with strong drug-like properties before any optimization. They use JAM to design antibodies against a variety of targets including TWO multi-pass membrane proteins – recording the first binder, of any kind, against an important 𝗖𝗣𝗚𝗥 and showing how de novo antibody design can open up new target space for biologics, which has long been considered a '𝗵𝗼𝗹𝘆 𝗴𝗿𝗮𝗶𝗹' 𝗼𝗳 𝗽𝗿𝗼𝘁𝗲𝗶𝗻 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴. To test how well the model generalizes, they design antibodies across 𝘁𝘄𝗼 𝗳𝗼𝗿𝗺𝗮𝘁𝘀 (VHH, scFV) and 𝗲𝗶𝗴𝗵𝘁 𝘁𝗮𝗿𝗴𝗲𝘁𝘀, including a protein with 𝗻𝗼 𝗸𝗻𝗼𝘄𝗻 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗶𝗻 𝘁𝗵𝗲 𝗣𝗗𝗕 - showing the model can generate designs beyond its training set. They also give the model the ability to reason over its output and perform multiple rounds of '𝗶𝘁𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗶𝗻𝘁𝗿𝗼𝘀𝗽𝗲𝗰𝘁𝗶𝗼𝗻' automatically at inference time and see a big boost in affinity and hit rate when the model is run for longer - suggesting protein design may benefit from the same 𝘁𝗲𝘀𝘁-𝘁𝗶𝗺𝗲 𝘀𝗰𝗮𝗹𝗶𝗻𝗴 𝗹𝗮𝘄𝘀 we've seen with frontier language models like o1. All of this is performed in silico, without any experimental optimization, but for good measure (and because the team has a wet lab) the use JAM to design and enrich a screening library that yields 𝟭𝟬𝟬𝘅 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 𝗶𝗻 𝗮𝗳𝗳𝗶𝗻𝗶𝘁𝘆 getting antibodies with binding in the pico-molar range - effectively combining design and optimization into a single experimental step and suggesting the best antibodies are going to come from combining wet + dry lab methods together in new ways. 𝗧𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝗥𝗲𝗽𝗼𝗿𝘁: https://lnkd.in/e33aZeh2 𝗕𝗹𝗼𝗴𝗽𝗼𝘀𝘁: https://lnkd.in/ep-iQj87 𝗘𝗻𝗱𝗽𝗼𝗶𝗻𝘁𝘀 𝗔𝗿𝘁𝗶𝗰𝗹𝗲: https://bit.ly/4fG5v1r Big congrats to the Nabla team! These seem like really exciting results, but curious to hear what people who actually work on antibodies think? cc Leonard Wossnig, Karl Dane Wittrup, Suzanne Edavettal, Paul Carter, Simon Chell, Andrew Buchanan, Dean Lee, Norbert Furtmann, Yu Qiu, Laurence Fayadat- Dilman, Enrico Guarnera, Victor Greiff, Brett Robison, Ta-Yi Yu, Rahmad Akbar, Rong Chen, Jared Adolf-Bryfogle, zhen zhang, Martin Orecchia, Matthew Burke