Sazan Mahbub

🎉🎉Announcing #ICML2026 workshop on "Multi-Modal Agentic AI": scale-icml-2026.github.io Topics covered: 1. Agentic memory, 2. Efficient agentic AI systems, 3. Scaling of multi-modal agents, 4. Agents for Planning, 5. Evaluation, guard railing, and benchmarking of agents 🚀Confirmed speakers: @dasongle (GenBio) @mohitban47 (UNCCH) @james_y_zou (Stanford) @chelseabfinn (Stanford) @MengdiWang10 (Princeton) @sunjiao123sun_ (Google) @MikeShou1 (NUS) @MinhyukSung (KAIST) With @HongyiWang10 @digbose92 @jaeh0ng_yoon @ManlingLi_ @nagsayan112358 @schowdhury671 #AgenticAI #MAS #MultimodalAI

Most agents either run fully autonomously or interrupt at the wrong times. What if agents know when YOU want to step in? 🚀Introducing PlowPilot - a web agent that adapts to your interaction patterns achieving +26.5% user-reported usefulness Huge credit to @FariaHuqOaishi for leading this project!

This framework is formulated as a latent-variable probabilistic model, implemented via an efficient approximation that bridges theoretical grounding and practical scalability. 2/n 🧵

A little late on the update, but thrilled to share that PRISM has been accepted at #ICLR2026 @iclr_conf 🎉 Huge thanks to Prof. Eric Xing (@ericxing) and Dr. Souvik Kundu (@thisissouvikk) for their mentorship and support throughout this work. #MultimodalRAG #RAGforScience 1/n 🧵

In this video, GenBio AI Co-Founder and Chief Scientific Advisor @Prof_Lundberg shares why this is a pivotal moment for biology and AI as advances in data and modeling converge. Watch the clip below ↓ youtube.com/shorts/3GmdYI2… Stay tuned for the series!

🚀 Introducing AIDO.DNA2, GenBio AI’s next-generation multi-species genomic foundation model. Built with a Mixture-of-Experts architecture and trained on the massive OpenGenome2 dataset, AIDO.DNA2 delivers higher accuracy, better efficiency, and stronger generalization across species, from variant prediction to regulatory genomics. Explore how it outperforms baselines and advances clinical genomics: 🔗 genbio.ai/aido-dna2-mult…

ICLR & CHI deadlines are coming up… Writing always feels overwhelming for me, so I made a little paper starter checklist to make it more fun (and printable 🌞). Sharing in case it helps anyone else feel a bit less stuck.💛 👇 oaishi.github.io/static/pdfs/Pa…

1/ 🚀 Introducing AIDO.StructureDiffusion: A generative model for structural protein design—enabling high-quality, controllable generation of monomers, complexes, and antibodies. 🧵

1/ 🧬 We’re at #ICML2025! GenBio AI is presenting five papers on generative AI for biology across two major workshops. Research areas: protein design, tissue modeling, perturbation prediction, and multimodal FMs 🧵↓

Good software should be fast, reliable, reusable, and maintainable. A lot of BioML benchmarking is uh… not. But biology doesn’t standardize to a few data types like language, audio, or images. We’re constantly inventing new ways to measure life... 1/n

Deeply grateful to my co-authors and supervisors at @GenBioAI : @probablybots @cfeinau @dasongle @ericxing 4/4

⚙️ Guidance through prior-posterior uncertainty signaling 🧬 Protein LLM + structure encoder priors 🏆 SoTA results on a range of inverse folding benchmarks 3/4

🤔 What if protein design models knew where they’re uncertain and adapted their predictions accordingly? 🚀 We introduce uncertainty-aware discrete diffusion for structure-conditioned protein design: doi.org/10.1101/2025.0…. 📍Catch us at #ICML2025 FM4LS workshop this week! 1/4

We will be presenting our paper on Multimodal Benchmarking of Foundation Model Representations for Cellular Perturbation Response Prediction at two #ICML2025 Workshops this week: FM4LS (fm4ls.github.io) and Generative AI and Biology (genbio-workshop.github.io/2025/). We systematically benchmarked perturbation embeddings across modalities (expression, protein, DNA, prior knowledge, networks). Embeddings based on network and prior knowledge consistently outperformed expression-based FMs, suggesting that structured biology remains a strong foundation for perturbation modeling. #SingleCell #FoundationModels #PerturbationModeling Preprint: biorxiv.org/content/10.110…

I have been long arguing that a world model is NOT about generating videos, but IS about simulating all possibilities of the world to serve as a sandbox for general-purpose reasoning via thought-experiments. This paper proposes an architecture toward that arxiv.org/abs/2507.05169

Uncertainty-Aware Discrete Diffusion Improves Protein Design １．This paper introduces an uncertainty-aware discrete denoising diffusion model that improves protein sequence design by dynamically estimating residue-specific uncertainty during the inverse folding process. ２．Unlike previous diffusion models that apply uniform updates, this method uses a prior-posterior uncertainty signaling mechanism to adaptively decide where and when to denoise, resulting in more accurate and interpretable sequence generation. ３．The model integrates a structure encoder and a 16B-parameter protein language model in a modular framework, combining structure-informed and sequence-informed priors. ４．A novel uncertainty estimator predicts both prior (based on current noise) and posterior (expected correctness after denoising) uncertainty for each residue, guiding updates only when confidence improves. ５．A non-learnable refinement module updates residues only when doing so reduces predicted uncertainty, eliminating the need for hand-tuned hyperparameters like the number of tokens denoised per step. ６．The system is jointly trained using a multi-objective loss combining denoising and uncertainty estimation, and achieves strong performance when combining separately trained denoisers with jointly trained refiners. ７．On the CATH-4.2 benchmark, the method achieves the highest amino acid recovery (AAR) rates across all subsets and ranks second in perplexity (PPL), showing balanced accuracy and confidence. ８．On generalization benchmarks TS50 and TS500, it outperforms state-of-the-art baselines with AARs of 68.85% and 71.18% respectively, and achieves the lowest perplexity on TS50. ９．The architecture builds on AIDO.ProteinIF but introduces fixed structure encodings and explicit uncertainty modeling, which improves efficiency without sacrificing performance. １０．This approach shows how integrating uncertainty awareness into discrete diffusion frameworks can yield not just better accuracy but also a more principled and interpretable generation process. @probablybots @cfeinau @SazanMahbub 💻Code: huggingface.co/genbio-ai/AIDO… 📜Paper: biorxiv.org/content/10.110… #ProteinDesign #DiffusionModels #InverseFolding #AIforScience #ComputationalBiology #UncertaintyEstimation

1/ GenBio AI will present five papers at the 42nd International Conference on Machine Learning (ICML 2025), highlighting recent work on biological foundation models. 🧬 Read the blog: genbio.ai/icml-2025

Honored to share a major thread of my PhD research, out now in PNAS. We address a core issue with how models are used for scientific discovery. Models are so important that they define the entire scientific process... 1/n

My talk at the Princeton AI Lab Distinguished Lecture Series on a vision of AI-driven Digital Organism (AIDO) for predicting, simulating, and programming biology at all levels. Video is available at the end of a nice digest from Princeton linked here (1/3) ai.princeton.edu/news/2025/watc…