Sarah Alamdari
Senior Applied Scientist, Microsoft Research
I am a Senior Applied Scientist at Microsoft Research on the BioML team, where I build generative AI models that design new, functional proteins; including EvoDiff, one of the first models to bring discrete diffusion to protein sequence design, followed by the Dayhoff Atlas, which scaled sequence diversity to improve protein generation, and CleaveNet, which applied deep learning to design protease substrates for early cancer detection.
I care about understanding the fundamental forces that drive biological systems, and using that understanding to rationally design efficient, bioinspired technologies faster than we ever could before.
Before Microsoft, I completed my Ph.D. in Chemical Engineering at the University of Washington as an NSF GRFP Fellow, where I developed enhanced sampling methods for molecular dynamics simulations to probe the interfacial phenomena that govern biology at the microscale; from biomolecular self-assembly and protein–surface interactions to biomineralization and enzyme engineering.
2026
- PaperFLIP2, a benchmark of protein fitness prediction under realistic distribution shifts, accepted to ICML 2026 as an oral presentation (top ~3% of submissions).
- PaperCleaveNet, AI-designed protease substrates for cancer-detecting sensors, published in Nature Communications; featured in MIT News.
- PaperEvoDiff, diffusion over evolutionary-scale sequence data for protein design, published in eLife after a 3-year review process.
- PaperClosing the loop, a review of experimentally validated AI protein design, in Current Opinion in Structural Biology
- TalkSequence-based approaches to protein design · Protein Evolution, Design and Informatics, Biochemical Society (Edinburgh, UK)
2025
- WorkshopThe Dayhoff Atlas selected as a Top 4 paper spotlight talk · Molecular ML Conference (MoML) @ MIT
- WorkshopThe Dayhoff Atlas · Machine Learning in Structural Biology (MLSB) Workshop (San Diego, CA)
- WorkshopAugmenting protein language models with synthetic structures · ICLR GEM Workshop (Singapore)
- PaperThe Dayhoff Atlas, a centralized collection of both protein sequence data and generative protein language models, released as a preprint with open datasets and models.
- PaperToward deep learning sequence–structure co-generation, a review of joint sequence–structure generative models, published in Current Opinion in Structural Biology
- PaperProtNote, zero-shot protein function annotation from text descriptions, published in Bioinformatics.
- TalkSequence-based approaches to protein design · AI4Science Seminar, Chalmers University (Gothenburg, Sweden)
- TalkProtein design in sequence space · 39th Annual Symposium of The Protein Society (San Francisco, CA)
- TalkGenerative sequence-based approaches to protein design · Schrödinger Biologics Symposium (Cambridge, MA)
- TalkEvoDiff: a general-purpose diffusion framework for sequence-first protein design · Applied ML Days, EPFL (Lausanne, Switzerland)
2024
- PaperFoldingDiff, protein backbone generation by diffusing over internal angles, published in Nature Communications.
- TalkDiscrete diffusion models for generative protein design · Boston Protein Design and Modeling Club (Cambridge, MA)
- TalkIn-silico design of protease-cleavable peptide substrates · Gordon Research Conference on Proteolytic Enzymes (Lucca, Italy)
- TalkAI-driven revolutions in protein engineering · SynBioBeta (San Jose, CA)
- TalkDiscrete diffusion models for controllable protein generation · Molecular ML Reading Group, Rosetta Commons
2023
- PaperEvoDiff, generative protein design in sequence space with discrete diffusion: preprint and open-source release; featured in TechCrunch.
- WorkshopEvoDiff selected as a spotlight talk · NeurIPS Machine Learning in Structural Biology (MLSB) Workshop (New Orleans, LA)
- WorkshopEvoDiff · NeurIPS Generative AI and Biology (GenBio) Workshop (New Orleans, LA)
- WorkshopEvoDiff · Molecular ML Conference (MoML) @ MIT (Cambridge, MA)
- TalkAn introduction to diffusion models · Models, Inference & Algorithms Seminar, Broad Institute (Cambridge, MA)
- TalkProtein sequence generation with evolutionary diffusion · Google Research (Cambridge, MA)
- TalkDiscrete diffusion models for controllable protein generation · ML Ideas, Microsoft Research (Cambridge, MA)
- TalkSequence-first generative approaches for protein design · AstraZeneca Journal Club
Generative Protein Design
My research develops generative machine learning methods for protein design. Specifically, my focus is on models that learn from evolutionary sequence and structure to generate new, functional proteins. This includes diffusion models over protein sequence and structure, protein language models trained at scale, and the benchmarks and annotation tools that connect these models to real-world design tasks.
Selected publications
Learning Protein Function
Building better protein models only matters if we can connect them to biological function. I develop methods to predict, annotate, and benchmark protein function. This work ranges from fitness landscape prediction to multimodal protein function annotation so that protein models stay comparable, reproducible, and grounded in real-world tasks.
Selected publications
-
Oral presentation at ICML 2026 (168 of 23,918 submissions, ~top 3%). -
-
-
Molecular Simulation
During my Ph.D. I developed and applied enhanced sampling methods such as metadynamics and parallel bias metadynamics to probe the interfacial phenomena that govern biological processes at the microscale. This work combined molecular dynamics simulation with experiment to understand biomolecules in various contexts such as peptoid self-assembly, protein behavior at the air-water interface, and biomolecular adsorption to mineral surfaces.
Selected publications
