Pioneering Intelligence

Happy to share that our paper Flash Invariant Point Attention (https://lnkd.in/efna_WWh) has been accepted to #NeurIPS2025 as a spotlight (top 1-3% of submissions)! We developed a factorized reformulation/approximation of the Invariant Point Attention module from AlphaFold-2 and leverage FlashAttention to achieve linear scaling in runtime and memory use with respect to sequence length, resulting in 30x speedup on protein and RNA generation tasks and enabling training on larger and/or more fine-grained (potentially all-atom) biomolecular structures, all while still respecting roto-translation invariance. I'm very grateful for my wonderful collaborators Axel Elaldi, Nicholas Franklin, Nathan Russell, Mickey Atwal, Yih-En Ban, and Olivia Viessmann at Flagship Pioneering for their support, and for pulling all-nighters helping me with experiments. Also, special thanks to Rishabh A. and the authors of RNAFrameFlow and FoldFlow for their well-maintained code repositories, which we base our work off of. See you in either San Diego or Mexico City!

We are Extuitive, pioneering the future of consumer product innovation and marketing with next-gen AI. Our platform uses AI consumers to simulate real-world behavior, empowering entrepreneurs to rapidly create, validate, and launch products and targeted content. We’re expanding the frontier of intelligence and democratizing product innovation today. Learn more here: Extuitive.com

🌟 I was honored to present our latest research from Pioneering Intelligence at #ICML2025. It was a pleasure connecting with so many brilliant people and learning about the cutting-edge work being done across the field. 🔍 About our work - “Flow-of-Options: Diversified and Improved LLM Reasoning by Thinking Through Options.” Large Language Models (LLMs) often struggle to explore diverse or creative solutions due to inherent biases in their reasoning processes. We introduce a novel approach called Flow-of-Options (FoO) — a reasoning framework that enables LLMs to systematically explore diverse solution paths without requiring any additional fine-tuning or training. FoO encourages exploration and interpretability through compact task representations that explicitly promote solution diversity. We integrated FoO into an agentic system for autonomously solving machine learning tasks, and evaluated its performance across both data science and therapeutic chemistry domains. Our results show substantial gains across a range of tasks — including classification, regression, reinforcement learning, and image generation — highlighting FoO’s versatility and impact across different prediction modalities. 🤝 Let’s connect: If you're exploring agentic AI and LLM reasoning for applications in healthcare and computational biology, I’d love to connect and explore potential collaborations. Please feel free to reach out! 📄 Paper: https://lnkd.in/guMzZFGk 💻 Code: https://lnkd.in/gwaBUA6V Pioneering Intelligence Flagship Pioneering #LLMs #AgenticAI #ICML #ComputationalBiology #AIResearch #ICML2025

Microsoft researchers achieved a breakthrough in the accuracy of DFT, a method for predicting the properties of molecules and materials, by using deep learning. This work can lead to better batteries, green fertilizers, precision drug discovery, and more. https://msft.it/6045SQwKR

Two weeks ago, I had the opportunity to present our latest work on structural biomolecular modeling from Pioneering Intelligence at the #ICLR #GEM workshop in Singapore. The five-day conference was an exciting event filled with insightful interactions with experts across various AI/ML fields and colleagues working at the intersection of AI and drug discovery. At the workshop, Pioneering Intelligence presented two of our projects led by our 2024 summer intern cohort: *** Bio2token (poster), developed from the ground up by Andrew Liu, is our latest all-atom structural tokenizer for a wide range of biomolecules, from small molecules to RNAs. We've recently released the code for training and inference, along with the dataset and pre-trained models, which you can find here: https://lnkd.in/eresZwFe. The corresponding paper is also available: https://lnkd.in/eDxfBdnr. Congratulations to Andrew Liu and the team including Nathan Russell and Olivia Viessmann. *** RL-DIF (oral), led and presented by Yasha Ektefaie, enhances the diversity of sequences that can be generated through inverse folding while maintaining competitive structural consistency. You can access the code here: https://lnkd.in/ePuxyXEp and the paper here: https://lnkd.in/eA-uJU5j. Congratulations to Yasha Ektefaie for his insightful presentation and to his co-authors Olivia Viessmann, Siddharth Narayanan, Drew Dresser, Mark Kim, and Armen Mkrtchyan. Feel free to reach out if you want to learn more about our internship program. We are also actively seeking experienced ML practitioners with expertise in the following areas: *** Protein Design: https://lnkd.in/ee7QX3h9 *** Small Molecules: https://lnkd.in/ensJqYHc