Technical Articles
Long-form breakdowns of the papers and ideas I'm thinking about.
Subcribe for the latest at Alchemy Bio ↗
2026
-
How to Train a One-Step Language Model Without a Teacher
Explore how removing time dependence from the generative flow results in straighter trajectories on the simplex, stable equilibria, and an endpoint map that yields a clean sequence in one step by construction — and can be trained end-to-end directly from data without a teacher flow.
Read the blog -
Aligning Discrete Generative Models to Any Reward, at Any Length
See how reward alignment for discrete generative models can be framed as entropy-regularized optimization and solved with off-policy reinforcement learning — first for fixed-length diffusion, then for any-length generation, where the model must learn when to insert new tokens as well as what to unmask.
Read the article -
Designing the Next Generation of mRNA Therapeutics
mRNA therapeutics are rapidly advancing, but designing mRNAs with desired properties remains a challenge. Learn how we trained a generative model on over 14 million mRNAs for joint design and multi-objective optimization of codons and UTRs, and our promising results in vitro!
Read the article -
Schrödinger Bridges: A Unifying Lens for Generative Modeling
Learn about the foundations of Schrödinger bridges as a theoretical framework that unifies the vast space of generative modeling methods and algorithms. From the static and dynamic Schrödinger bridge problems, to the stochastic optimal control formulation, to various instantiations and applications of Schrödinger bridges.
Read the article
2025
-
2025: research in review
2025 was such a special and inspiring year in my research journey! Read about how I got started with research and my 2025 research themes, including some small anecdotes and highlights from the year.
Read the article -
Unlocking the Future of Peptide-Based Precision Medicine
Learn how masked discrete diffusion can generate therapeutic peptides directly as SMILES strings, and how guiding each unmasking step toward several objectives at once, from binding affinity to solubility and permeability, steers generation without retraining the model.
Read the article
2024
-
A Complete Guide to Protein Folding
Walk through protein structure prediction from the ground up: how sequences fold, how multiple sequence alignments and HMM profiles are built, and how a three-track network passes information between sequence, pairwise distances, and 3D coordinates to predict structure.
Read the article -
A Complete Guide to Spherical Equivariant Graph Transformers
Build rotational equivariance into graph neural networks step by step — from group representations, spherical harmonics, and Clebsch-Gordan tensor products to equivariant kernels, message passing, and the attention layers of the SE(3)-Transformer.
Read on arXiv -
MolTrans: Transformers for Drug-Target Interaction Prediction
See how transformer encoders can predict whether a drug binds a protein: breaking molecules and proteins into substructures, embedding them with self-attention, and modeling their pairwise interactions, with every step of the architecture and training illustrated.
Read the article