Building discovery engines for biology. @Stanford · @ArcInstitute
- Tweets 181
- Following 334
- Followers 1,553
- Likes 2,538
Pinned Tweet
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.
13
151
586
42,095
Garyk Brixi retweeted
SFT is not dead! 🥳
We found a way to make SFT rival current prevailing posttraining methods, often generalizing better and forgetting less than RL and OPSD. 🤯
Following our prior work on reasoning with sampling, we now introduce sampling to the posttraining stack.
1/n
61
222
1,961
286,321
Garyk Brixi retweeted
The importance of this approach will become increasingly clear as more of the dark genome becomes resolved. Congratulations @_David_Li @garykbrixi @BrianHie — phenomenal!
Many important bacterial non-coding RNAs and structural elements remain undiscovered. To help, @_David_Li, @garykbrixi, @mfgrp, @BrianHie & team introduce Minerva, which uses a genome language model to predict RNA base pairing, repeats, & other interactions from sequence alone.
1
5
16
3,964
Garyk Brixi retweeted
Beautiful study. There is so much in microbial genomes we still don’t understand. Exciting to see the use of genome language models powering the discovery of these cool new systems. Congrats to @_david_li, @garykbrixi, @brianhie, and the team!
Excited to share Minerva, our approach using genome language models for biological discovery! Using Minerva, we find that UG27 reverse transcriptase systems encode variable arrays of diverse ncRNAs with a shared structure, each templating a short DNA hairpin. With @garykbrixi.
3
39
252
38,118
Within Minerva, we developed two coevolution prediction methods to predict not just the strength but also type of interaction.
Jacobian fingerprints: the categorical Jacobian has become a workhorse for interpreting biological language models, but current post processing focuses just on interaction strength. Instead of throwing away the rich interaction signatures, we can use them to discover evolutionary 'fingerprints' which tell us about different types of interactions. This unsupervised method identifies different types of sequence to function relationships left behind by evolution.
1
1
11
2,123
Minerva builds on gLM2 (@Micro_Yunha @ancornman1 @tatta_bio), the categorical Jacobian (@ZhidianZ @HWaymentSteele @sokrypton), and protein contact heads (@proteinrosh @alexrives)! Also grateful to the wider bioinformatics and coevolution community whose work made this possible. Excited to see where genome-scale discovery goes next.
1
1
26
3,460
This was all inspired by foundational EVcouplings and ncRNA modeling/discovery work by @deboramarks, Chris Sander, @RonBreaker, Zasha Weinberg, Sean Eddy, and many others. We spent many nights staring at @caleb_weinreb’s RNA couplings paper:
cell.com/cell/fulltext/S0092…
5
692
Garyk Brixi retweeted
This is super cool!
@garykbrixi and @_David_Li use genome language models to discover a new class of reverse transcriptase, incredible!
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.
1
7
34
6,438
Congrats to David and @garykbrixi!
Some of my early conversations with them about fragments of this paper are still very memorable... & seeing it all come together as Minerva is honestly so special. Congrats again!
Excited to share Minerva, our approach using genome language models for biological discovery! Using Minerva, we find that UG27 reverse transcriptase systems encode variable arrays of diverse ncRNAs with a shared structure, each templating a short DNA hairpin. With @garykbrixi.
2
15
1,923
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.
13
151
586
42,095
Minerva turns prokaryotic genomes into maps for discovery.
Genome language models are promising for extracting coevolution, but previous methods are far too slow to be practical, and collapse diverse interactions into just one strength value.
Minerva predicts many types of genomic interactions at scale, including RNA or DNA base-pairing, non-coding or coding repeats, protein contacts, and unannotated proteins, while being more accurate than previous models.
1
17
2,545