iiinova — where your digital self keeps evolving

Burstein lab

@bursteinlab.bsky.social

Having fun with microbial genomics and machine learning

Burstein lab @bursteinlab.bsky.social · 24/09/2026

So many bacterial genomes are being edited… Could we spot them in the wild even without obvious markers??? Worry not! Our @EdanGabay has you covered. In our new preprint, she pinpoints such genes disrupting the natural genomic "grammar": www.biorxiv.org/content/10.6... 1/4

241

Burstein lab @bursteinlab.bsky.social · 12/07/2026

Is it fast? We also built a lightweight version of FAMUS that runs close to the speed of a plain hmmsearch. Both the full and lightweight versions can run on CPU or (a bit faster) on GPU. 5/8

200

Burstein lab @bursteinlab.bsky.social · 12/07/2026

Benchmarking on KEGG and PANTHER, FAMUS improves on KofamScan and InterProScan when most sequences are unannotated, a realistic case for metagenomics and non-model organisms. This led us to compile FAMUS databases from KEGG, InterPro, eggNOG, and OrthoDB. 4/8

100

Burstein lab @bursteinlab.bsky.social · 12/07/2026

FAMUS uses supervised contrastive learning (SupCon) to explicitly separate known protein families from out-of-scope proteins. SupCon also handles sparsity, giving good classification even for small protein families. 2/8

100

Burstein lab @bursteinlab.bsky.social · 16/02/2026

⚡ Reduced alphabets yield shorter inputs and major runtime gains, while maintaining comparable, and sometimes improved, predictive performance. (4/5)

100

Burstein lab @bursteinlab.bsky.social · 16/02/2026

🔤 By combining Byte Pair Encoding (BPE) with reduced amino acid alphabets based on residue properties, we train new pLMs and evaluate them across diverse biological tasks, like solubility, enzyme, PPI, and stability prediction. (3/5)

120

Burstein lab @bursteinlab.bsky.social · 16/02/2026

⚖️ Unlike natural languages, proteins aren’t clearly separated into “words”, making tokenization tricky. Short tokens create long sentences, while long tokens lead to a sparse vocabulary that is hard to learn. But reducing the alphabet size might help! (2/5)

110

Burstein lab @bursteinlab.bsky.social · 07/01/2026

2/4 B-PPI is a cross-attention model for bacterial PPI prediction at scale. Given protein pairs, it leverages ProstT5, a structure-aware protein language model, to generate embeddings, and outputs the interaction probability.

100

Burstein lab @bursteinlab.bsky.social · 07/01/2026

1/4 Ever wanted to predict bacterial protein-protein interactions (PPI) on a large scale? We wanted to, but realized there’s no such algorithm that is both rapid and optimized for bacterial protein analysis. This led our ⭐️Chen Agassy⭐️ to develop B-PPI: doi.org/10.64898/202...

153

Burstein lab @bursteinlab.bsky.social · 07/06/2025

5/6 💡 Discover how NLP is being applied to: • Protein structure prediction 🏗️ • Taxonomic classification 🌳 • Mutational effect prediction 🔀 • Gene expression prediction 📈 And much more!

120

Burstein lab @bursteinlab.bsky.social · 07/06/2025

4/6 🧩 Tokenization challenges? We've got that covered too! Explore different approaches to breaking down biological sequences and their impact on model performance.

110

Burstein lab @bursteinlab.bsky.social · 07/06/2025

3/6 📚 We break down the evolution of NLP models in biology, from classic word2vec to cutting-edge transformers and hyena operators. Understand their strengths, limitations, and exciting applications!

100

Burstein lab @bursteinlab.bsky.social · 07/06/2025

2/6 🔬 We dive deep into how NLP techniques are revolutionizing the analysis of biological 'languages': • DNA 🧬 • RNA 🧬 • Proteins 💪 • Entire genomes 🔍 Learn how these methods are unlocking new insights in genomics!

110

Burstein lab @bursteinlab.bsky.social · 07/06/2025

1/6 🧬📊 Curious about the buzz around NLP in biology? Feeling overwhelmed by the rapid developments? We've got you covered! Our review on NLP applications in genomics, by the wonderful Ella Rannon, is now out as a pre-print! #NLP #Bioinformatics arxiv.org/abs/2506.02212

14715