←
Ogün Adebali988 posts
Ogün Adebali avatar

Ogün Adebali

@compGenomics

Associate Professor at Sabancı University

◷ Joined January 2016▧ 194 Photos and videos
Phillip Isola avatar
⟳ Ogün Adebali retweeted
Phillip Isola@phillip_isola·Sep 28

Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI. Perhaps counterintuitively, I think it's a great time to do so. I wrote up some thoughts on this here: web.mit.edu/phillipi/www/wri…

Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 29

We're hiring at the Adebali Lab (@sabanciu)! Postdoc (3 years): phylogenetics, sequence analysis, profile-HMM, ML for biological data, Python/R PhD student: bioinformatics background, programming basics, curiosity about evolution & protein function linkedin.com/posts/adebali_b…

Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

Ceren Yildirim is the lead author. Nurdan Kuru, original lead author of PHACT for proteins, contributed to this study. Enjoy the preprint.

Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

Evo2 needs A100/H100 GPUs. GPN-MSA: 86M parameters. PHACTn: standard CPUs, no training, no GPU. ~27 CPU-hours for 100 × 10kb windows at 50 parallel jobs. PHACTn provides phylogeny-aware variant scoring for the entire genome.

Post attachment
Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

We built a "hard-case" dataset: variants where existing tools disagree. The toughest 9% of the benchmark. PHACTn AUROC: 0.921 Evo2-7B: 0.884 GPN-MSA: 0.866 When tools fight, PHACTn often gets it right.

Post attachment
Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

Non-coding is hardest. PHACTn AUROC: 0.805 🥇 GPN-MSA: 0.787 Evo2-7B: 0.717 And on the TraitGym Mendelian benchmark (113 diseases from OMIM): PHACTn AUROC: 0.914, AUPR: 0.736 — top of the leaderboard.

Post attachment
Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

PHACTn: - Reconstructs ancestral probabilities at every tree node - Corrects for gaps with binary ASR - Traverses the tree from query to outward - Weights substitution events inversely by phylogenetic distance - Applies a sigmoid gap-frequency correction With only 4 parameters!

Post attachment
Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

Standard ancestral reconstruction ignores gaps and treats them as missing data, forces probabilities to sum to 1 across A/C/T/G. This creates "ghost" signal at positions dominated by deletions. PHACTn uses a dual binary+nucleotide ASR approach to correct for this explicitly.

Post attachment
Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

Two alignments might look identical by nucleotide frequency, yet tell different stories. One position may have switched three times independently: strong tolerance indicator. Another may have done it once: weaker tolerance. PHACTn traverses the tree to distinguish these cases.

Post attachment
Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

Evolution is the longest-running experiment. If a position has been conserved across mammals for millions of years, mutations there are probably bad news. But conventional conservation scores miss something crucial: not all substitutions are independent events.

Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

98% of the human genome is non-coding. Yet predicting which non-coding variants are pathogenic remains one of the hardest problems in genomics. ML classifiers overfit. Genomic language models are black boxes with intense computation. We needed something better.

Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 15

I am extremely happy to introduce PHACTn, a variant effect predictor for the entire genome with just 4 parameters, no training, no context information (unlike language models), no GPU - and it outperforms massive language models on non-coding variants. doi.org/10.64898/2026.09.08.…

Ogün Adebali avatar
Ogün Adebali@compGenomics·Sep 11

3/ Beyond single receptors, we offer 3 comparison tools: * Differential conservation — compare 2 paralogs to find shared vs. lineage-specific residues * Multi-receptor comparison — scale up to entire subfamilies * Combine orthologs — merge ortholog alignments for custom MSAs