Dan (✈️ ICML-26)

Our diploid aware genomic language model is out! 1/6) The genomes we train AI models on do not exist. Most genomic language models read a single haploid sequence, but human genomes are diploid.

We also introduced Contrastive Phase Loss for phase-sensitive pretraining. At 100M parameters, cis/trans AUROC rose from 0.54 to 0.62.

Our diploid aware genomic language model is out! 1/6) The genomes we train AI models on do not exist. Most genomic language models read a single haploid sequence, but human genomes are diploid.

3/ We trained on trillions of base-pairs from the phased 1000 Genomes. We introduce new diploid-aware benchmarks for pathogenicity/VEP, zygosity, hemizygosity and ClinVar indels, plus cis/trans benchmarks where the variant pair, gene and genetic background stay the same; only phase changes.

Our diploid aware genomic language model is out! 1/6) The genomes we train AI models on do not exist. Most genomic language models read a single haploid sequence, but human genomes are diploid.

InterFeat just got the top broad-applicability score in the ISBCB paper award (Israel's national bioinformatics society)! @michalinial @HyadataLab InterFeat helps researchers find the interesting findings in data: ones that are novel, make sense and actually useful.

Paper: www.nature.com/articles/s41... Code: github.com/LinialLab/In... Part of my PhD on finding interesting things in data. #AI #huji #interfeat

InterFeat just got the top broad-applicability score in the ISBCB paper award (Israel's national bioinformatics society)! @michalinial @HyadataLab InterFeat helps researchers find the interesting findings in data: ones that are novel, make sense and actually useful.

We consider a finding as only "interesting" only if it's (a) predictive, (b) not already known, and (c) makes sense. InterFeat checks all three automatically and explains why!

Heading to #ICML2026 Seoul! I’ll be there from July 3–20 presenting my work on Protein/Biological LLMs. (And travelling + photographing, geeking out). I'd love to connect, share ideas, co-travel post-con or collaborate. DMs open! @icmlconf.bsky.social #ICML #AI #eleutherai #korea

SOTA LLM architecture ablations for a novel encoder paper proceeding apace: 1. hybrid architectures (Not the one you're thinking of): Jury is out, but I like it. 2. Local/global (#modernbert) style: less compute benefit than I expected, maybe my implementation sucks?

5. #Kimi Attention Residuals (AttnRes) + Block - nuked my initial results, horrifically slow if using just the pseudocode (dropping got me from ~6 sec/it to 0.8~1.4 it/sec). Maybe needs better kernels (or a better coder than Opus, gemini and gpt and me).

SOTA LLM architecture ablations for a novel encoder paper proceeding apace: 1. hybrid architectures (Not the one you're thinking of): Jury is out, but I like it. 2. Local/global (#modernbert) style: less compute benefit than I expected, maybe my implementation sucks?

Kinda weird that I need different prompts to GPT5 to get it to proofread my manuscript properly. Each iteration + prompt yields different mistakes. And I'm looking for basic mistakes (e.g., grammar/spelling level)! This is a super common usecase, i'm surprised at the variation #writing #gpt #ai

1/ Our paper “Protein Language Models Expose Viral Immune Mimicry” is now published in Viruses! We show that protein Language Models can identify viral proteins, and those that fool our immune system. DOI: 10.3390/v17091199.

7/ Read the paper: "Protein Language Models Expose Viral Immune Mimicry" www.mdpi.com/3477256 | DOI: 10.3390/v17091199. github.com/ddofer/Prote... #ml4bio #ai #virology #science #ML

1/ Our paper “Protein Language Models Expose Viral Immune Mimicry” is now published in Viruses! We show that protein Language Models can identify viral proteins, and those that fool our immune system. DOI: 10.3390/v17091199.

5/ Takeaway: PLM mistakes are signals—a practical way to flag candidate immune-evasive proteins for deeper study, with implications for vaccine design and antivirals. And we find that there aren't "adversarial examples" when it comes to proteins and the immune system (unlike NLP or vision)!

1/ Our paper “Protein Language Models Expose Viral Immune Mimicry” is now published in Viruses! We show that protein Language Models can identify viral proteins, and those that fool our immune system. DOI: 10.3390/v17091199.

3/ Misclassifications are enriched for immune-evasive proteins (chronic/latent families like HIV) and low-immunogenic sequences—paralleling the real (biological) immune system's blind spots!

1/ Our paper “Protein Language Models Expose Viral Immune Mimicry” is now published in Viruses! We show that protein Language Models can identify viral proteins, and those that fool our immune system. DOI: 10.3390/v17091199.