Out of 7 papers in my NeurIPS Benchmarks and Datasets Track area, the PCs overruled my recommendation on 3?!

Genome language models can generate new, high-fitness bacteriophages! @samuelhking.bsky.social @claudiadriscoll.bsky.social @david-li.bsky.social @danguo.bsky.social @adititm.bsky.social Garyk Brixi @maxewilkinson.bsky.social @brianhie.bsky.social www.biorxiv.org/content/10.1...

Out of 7 papers in my NeurIPS Benchmarks and Datasets Track area, the PCs overruled my recommendation on 3?!

Genome language models can generate new, high-fitness bacteriophages! @samuelhking.bsky.social @claudiadriscoll.bsky.social @david-li.bsky.social @danguo.bsky.social @adititm.bsky.social Garyk Brixi @maxewilkinson.bsky.social @brianhie.bsky.social www.biorxiv.org/content/10.1...

The posting is unfortunately not super clear on this, but please apply if you want to do BioML research, especially around machine learning for molecular biology and bioengineering! I get sad if there are no applicants with this profile in the pool!

If you're an undergrad and want to intern with me, this is where you need to apply!

Train a protein structure predictor that can handle 29 non-canonical amino acids, then use it to design binders with non-canonical amino acids that reduce immunogenicity. @panhammarstrom.bsky.social @patrickbryant1.bsky.social www.biorxiv.org/content/10.1...

A joint sequence-structure diffusion model for transmembrane proteins! www.biorxiv.org/content/10.1...

A compelling review of how ML/AI could help in the quest to find an enzyme for every reaction. @jsunn-y.bsky.social @francescazfl.bsky.social Yueming Long @francesarnold.bsky.social www.cell.com/cell-systems...

Typical Ohio activity

Ramblings of a mad AC 1. If you're reviewing a benchmarks and dataset paper, you should look at the dataset and make sure it's not full of garbage. 2. Biological datasets are especially prone to subtle garbage.

The @uwproteindesign.bsky.social's experimental pipeline behind models like RFdiffusion and ProteinMPNN: - A rapid, scalable, pipeline for producing and characterizing proteins - A demultiplexing protocol for converting oligopools to clonal constructs Jason Qian @lfmilles.bsky.social Basile Wicky

Ramblings of a mad AC 1. If you're reviewing a benchmarks and dataset paper, you should look at the dataset and make sure it's not full of garbage. 2. Biological datasets are especially prone to subtle garbage.

A benchmark dataset of 614 experimentally characterized de novo designed monomers from 11 different design studies shows that: - deep learning structural metrics only weakly predict success - The score distribution is different for different types of structures @grocklin.bsky.social

MSA Pairformer efficiently extracts structure, protein-protein interactions, and mutation effects from MSAs by decomposing the effects of phylogeny and structural contacts. @yoakiyama.bsky.social Zhidian Zhang @milot.bsky.social @martinsteinegger.bsky.social @sokrypton.org

I just want to be a good area chair

My child's dinner - a scoop of peanut butter on a deli sandwich flavored potato chip - a scoop of peanut butter mixed into yogurt - a slice of ham - a tablespoon of melted mozzarella cheese - a round piece cut out of a slice of bread

- Model bacterial genomes as sequences of proteins - predict protein-protein interactions, operon structure, and protein function - infer phenotypic traits - design synthetic genomes with desired properties @macwiatrak.bsky.social @mariabrbic.bsky.social @andresfloto.bsky.social

Last year ESM-3 was the 1st PLM (if I recall correctly) to supplement naturally occurring seqs w/ synthetic inverse folding-derived seqs, but they failed to show any direct evidence of improvement. Its nice to finally see direct evidence of the benefit of including synthetic data during training

In 1965, Margaret Dayhoff published the Atlas of Protein Sequence and Structure, which collated the 65 proteins whose amino acid sequences were then known. Inspired by that Atlas, today we are releasing the Dayhoff Atlas of protein sequence data and protein language models.

Great idea for a dataset and task! Thanks for using and assessing out Prot-xLSTM

In 1965, Margaret Dayhoff published the Atlas of Protein Sequence and Structure, which collated the 65 proteins whose amino acid sequences were then known. Inspired by that Atlas, today we are releasing the Dayhoff Atlas of protein sequence data and protein language models.

The stochastic wait time between submission and release on @biorxivpreprint.bsky.social is A good B bad C a secret third thing

Very excited about our latest all-atom generative model proteina, check out the project page (research.nvidia.com/labs/genair/...) and stay tuned for the code release soon!

Partially-latent flow matching enables sequence-structure codesign of large proteins and functional motif scaffolding. @kdidi.bsky.social @machine.learning.bio @karstenkreis.bsky.social @arashv.bsky.social arxiv.org/html/2507.09...

Partially-latent flow matching enables sequence-structure codesign of large proteins and functional motif scaffolding. @kdidi.bsky.social @machine.learning.bio @karstenkreis.bsky.social @arashv.bsky.social arxiv.org/html/2507.09...

All all-atom diffusion model of protein sequences. www.biorxiv.org/content/10.1...

Physics-based design of efficient Kemp eliminases @lynnkamerlin.bsky.social www.nature.com/articles/s41...

Deep learning methods for protein structure prediction and design produce idealized structures. Finetuning on a set of physics-based de novo proteins improves their geometric diversity and generalization capabilities. @benorr.bsky.social @kortemmelab.bsky.social www.biorxiv.org/content/10.1...

Deep learning methods for protein structure prediction and design produce idealized structures. Finetuning on a set of physics-based de novo proteins improves their geometric diversity and generalization capabilities. @benorr.bsky.social @kortemmelab.bsky.social www.biorxiv.org/content/10.1...

With contributions from fantastic colleagues @martinsteinegger.bsky.social , @mikeinouye.bsky.social, @jlistgarten.bsky.social , @ideasbyjin.bsky.social, @michael-heinzinger.bsky.social, and many more, the first CSHL volume on ML for Protein Science and Engineering is out: lnkd.in/dQdgGPpp

End-to-end differentiable homology search for protein fitness prediction. @yaringal.bsky.social @deboramarks.bsky.social @pascalnotin.bsky.social arxiv.org/abs/2506.089...

End-to-end differentiable homology search for protein fitness prediction. @yaringal.bsky.social @deboramarks.bsky.social @pascalnotin.bsky.social arxiv.org/abs/2506.089...

Average zero-shot mutation effect predictions across related sequences to reduce noise and improve accuracy. @cwjpugh.bsky.social @paulinanunezv.bsky.social @jonnyfrazer.bsky.social Mafalda Dias www.biorxiv.org/content/10.1...

Average zero-shot mutation effect predictions across related sequences to reduce noise and improve accuracy. @cwjpugh.bsky.social @paulinanunezv.bsky.social @jonnyfrazer.bsky.social Mafalda Dias www.biorxiv.org/content/10.1...

I'm still seething about the flagrant dishonesty in this piece that @gravelinfluencer.bsky.social posted last night, and especially the flat-out lies about how a cyclist was hit and killed on the shared use path along Memorial Drive. eastietimes.com/2025/06/04/m...

Do protein language models store different structural elements in factorizable subnetworks? To find out, we masked out PLM weights to suppress performance on CATH subcategories or secondary structure elements while maintaining performance on other sequences or residues.

Obviously, finding diverse homologs violates federal civil rights law. We should be finding the most qualified homologs! github.com/EddyRivasLab...

NIH funding supporting the HMMER and Infernal software projects has been terminated. NIH states that our work, as well as all other federally funded research at Harvard, is of no benefit to the US.

Obviously, finding diverse homologs violates federal civil rights law. We should be finding the most qualified homologs! github.com/EddyRivasLab...

Flash invariant point attention enables the generation of larger biomolecules. arxiv.org/abs/2505.11580

A great article by @hormiga.bsky.social scienceforeveryone.science/p/the-power-...

I didn't think I knew much about playing piano until I started helping my kid practice

Steered generation for protein optimization: On datasets with ~10^2 measurements, steering a discrete diffusion model outperforms RL on a protein language model for generating improved variants. @jsunn-y.bsky.social @francesarnold.bsky.social @yisongyue.bsky.social arxiv.org/abs/2505.15093

Steered generation for protein optimization: On datasets with ~10^2 measurements, steering a discrete diffusion model outperforms RL on a protein language model for generating improved variants. @jsunn-y.bsky.social @francesarnold.bsky.social @yisongyue.bsky.social arxiv.org/abs/2505.15093

Turns out this baby poops a lot now that his bowels are trainable end-to-end

End to end differentiable protein structure alignment! @sokrypton.org @martinsteinegger.bsky.social www.biorxiv.org/content/10.1...

ProteinGuide: Predictor guidance for conditional generation with any autoregressive or discrete diffusion protein model. @junhaobearxiong.bsky.social @hnisonoff.bsky.social @jlistgarten.bsky.social arxiv.org/abs/2505.04823

ProteinGuide: Predictor guidance for conditional generation with any autoregressive or discrete diffusion protein model. @junhaobearxiong.bsky.social @hnisonoff.bsky.social @jlistgarten.bsky.social arxiv.org/abs/2505.04823

This is probably obvious: Becoming a scientist is a pathway to a good job. US-funded research supports many good, stable, long-term careers. Jobs people want. Manufacturing?... I don't hear the kids clamoring for those jobs. And yet, DOGE has cut thousands of jobs for scientists. 👎👎