


PhD student at MIT Micro. Genomic language models, Metagenomics, phage enthusiast, Dead Head.


We're hiring a Lead Scientist, focused on biocatalysis, at Tatta Bio in Cambridge, MA! We're looking for a hands-on scientist to lead the development of our high-throughput enzyme characterization platform, working closely with our ML team. If you have deep expertise in biochemistry and analytical chemistry, enjoy building experimental workflows, and want to help shape a growing research program, we'd love to hear from you. Check out the job description below to learn more and apply! tatta-bio.notion.site/Lead-S…

Structural clustering across the tree of life reveals candidate Asgard-eukaryote homologs and a shared expansion of solenoids doi.org/10.64898/2026.10.28.…

Language models can learn new tasks from just a few examples in their prompt, with no retraining. This is called in-context learning (ICL). Now genome language models can do it too. With a handful of examples, Evo 2 ICL generates proteins, de novo antibodies, CRISPR-Cas9 loci, mitochondrial genomes and entire bacteriophages. We can even use the synthetic sequences iteratively to construct a chain-of-thought (CoT) and achieve harder design goals. New work with Yukun A. Hao, @BrianHie and team 🧵

Excited to share our preprint! A new logic for bacterial immune signaling: second messengers released from tRNAs rather than built from free nucleotide precursors. biorxiv.org/content/10.64898…

Introducing Microbiome Repo, a catalog of every public human gut metagenome (>500k samples) with metadata pulled straight from sequence archives, papers, and supplements olmlab.github.io/microbiome_… Published gut metagenomes? Search your name and see if it got your study!

Bioinformaticians have been writing the same abstract for decades - luispedro.substack.com/p/bio…

Genomes are full of dark matter of unknown functions. We present Minerva, a new method for discovery guided by genome language models. Minerva reveals the interactions hidden in non-coding DNA, pointing to hundreds of new putative RNAs and repetitive elements per bacterial genome. Minerva allows us to find and study elements invisible to traditional methods at orders of magnitude greater scale than before.

Can genomic language models design biological systems to access new regions of chemical space? Some of the most complex molecules in nature are made by biosynthetic gene clusters: massive enzymes that carry out complex chemistry. But engineering these systems is challenging due to complex cross-domain interactions. To solve this, we applied gLM2 to controllably generate new domain architectures by learning from vast evolutionary data.

Microbial intergenic regions encode much of the regulatory and functional logic of the genome, but they remain among the hardest sequences to systematically discover and annotate. We trained a sparse autoencoder on gLM2 to identify recurring patterns across hundreds of millions of microbial intergenic regions, without requiring a predefined motif, RNA family, or annotation. Those features are now integrated into SeqHub multimodal search, where they can be searched together with proteins, Pfam domains, Rfam families, taxonomy, and genomic context. Using this framework, we identified divergent members of known RNA families, and found previously uncharacterized structured RNAs and candidate regulatory elements beyond existing annotation models. The goal is to make microbial noncoding sequence space systematically searchable: discover an intergenic pattern, trace where it occurs across evolution, and use its conserved genomic context to generate hypotheses about function. Explore: seqhub.org Paper: tatta.bio/intergenic-sae Co-Authors: Nicolo Zulaybar, Matt Tranzillo, Rachel Silverstein, @Micro_Yunha, @ancornman1 This work is supported by @MooreFound and @SchmidtFutures

My first, first-author work of my PhD is now out! pnas.org/doi/10.1073/pnas.26… Huge thanks to my incredible advisor @sokrypton and wonderful collaborator @MazAbulnaga for their support during this whole project!

Excited to share our now published manuscript @CellCellPress highlighting advances in modeling the evolution of protein-protein interactions with MSA Pairformer. Big thanks to @ZhidianZ, @oliviatang116, @eunbelivable, Milot Mirdita, Martin Steinegger, and @sokrypton !🧵 (1/10) cell.com/cell/fulltext/S0092…

My lab and @GeneWeiLiLab at @MITBiology are recruiting a joint postdoctoral fellow to develop foundational datasets for modeling bacterial phenotypes from genotype. We are seeking candidates with experience in technology development, high-throughput library generation, and metabolomics. Job posting: careers.peopleclick.com/care…

Extremely useful paper diagnosing why SAEs on certain models are swamped with dead neurons with prescriptions on how to fix them
A sparse autoencoder aims to learn a dictionary of interpretable features from a model's activations — but a lot can come out "dead," never firing once. On some models this is rare; on others, >70% die even with fixes like AuxK. We went down a rabbit hole to understand why...

Max-pooling sparse features > mean-pooling dense features Pooling nerds will enjoy this part of the ESMC interpretability work. Protein language models have to capture *everything* about a sequence, including the organism that the sequence came from, its structural fold, etc in order to be able predict amino acids. If a functional signal (e.g. active site, allostery) is local in sequence space, it may be drowned out by mean-pooling. Max-pooling across sparse features preserves this signal, and we show that it improves functional homolog recovery in the presence of decoys.
Furthermore, if we maxpool the features across the sequence dimension, we reduce the effect from low frequency averaging of dense embeddings, and get a vocabulary of strongly activated functional features for each protein. In fact, now we can use some of the common techniques found in information retrieval, such as TF-IDF and BM25 [our Jaccard similarity is close to this] for protein search.

🚀 Excited to share our new work: Absolute Stability Predictor! 📊: forms.gle/4ZnXZSnTBvaykkAi9 Built the MGnify Stability Dataset (1.8M+ measurements) and developed stability prediction models, together with @grocklin, @KotaroTsuboyama, @sokrypton, and teams.