←
Jacob225 posts
Jacob avatar

Jacob

@JacobMolBio

Bio x AI Singularitas biologica

◷ Joined October 2024▧ 147 Photos and videos
Jacob avatar
Jacob@JacobMolBio·Nov 11

Fable 5.2 is crazy good at animation. It made this short, 'Better Half', about Claude crabs that happen to wear vault particle half-shells, when all of a sudden... phages arrive. Fable, with a team of other Fable and Opus 5.5 agents, built this in Blender with about 28,000 lines of Python, then rendered it in 4K on fal cloud compute, and made the music and sound with ElevenLabs. I gave Fable Riley's artifact guide as a reference, though this was not one shot.

1:25

This is actually INSANE. Fable 5.2 just created this in one shot, and I did it live on stream. I gave Claude one prompt. It wrote code that builds the entire room in Blender: the desk, the MacBook, the light, even Clawd. The laptop screen is a web page Claude built to look like the Claude app, played on the laptop inside the 3D scene. Then it animated everything itself: 13 camera shots, every single key press, Clawd jumping around, and the sun going from golden hour to night to sunrise. It rented 10 GPUs on RunPod and rendered the whole thing in 4K in 52 minutes. Total cost: $5.75. It graded it, added the titles, and made the music and sound effects with ElevenLabs. (More info below)

0:59
Jacob avatar
Jacob@JacobMolBio·4h

I asked GPT-6.2 Astra to show off its data processing abilities on various microscopy datasets in ImageJ. Astra and its Sol agents picked out six diverse public datasets, autonomously drove ImageJ analyses, and then made this video of their work. 1) Microtubule growth in Xenopus egg extract, with and without γTuNA (a microtubule nucleating protein fragment), captured by TIRF microscopy (Rale et al., eLife, 2022). 2) Sarcomere contractions in human cardiomyocytes before and after the myosin-inhibiting drug mavacamten, captured by spinning-disk confocal (Mohammadzadeh et al., Harvard Dataverse, 2025). 3) Z-disc spacing in fixed human cardiomyocytes captured by spinning-disk confocal (Gerbin et al., Cell Systems, 2021). 4) Paced mouse heart tissue with voltage and calcium optical mapping (Xianhong Ou, Figshare, 2020). 5) C. elegans movement and neuron fluorescence in near-infrared brightfield and spinning-disk confocal (Atanas, Kim et al., Cell, 2023). 6) Zebrafish heart beating in light-sheet microscopy (Xinyuan Zhang, Zenodo, 2023). GPT-6.2 agents used the Fiji (Fiji Is Just ImageJ) bundle for the processing, and used ElevenLabs for the audio.

2:43

GPT-6.2 Astra is amazing. For $20/month, you can now access a model that surpasses top human intelligence in some domains. I'm most excited to see what scientific discoveries people are able to make using it.

Jacob avatar
Jacob@JacobMolBio·Nov 17

Claude can drive ImageJ quite well. Here's an example of it autonomously detecting fluorescent microtubule tips (comets), measuring speed, generating kymographs, separating anterograde from retrograde comets, and marking catastrophes. Claude used Fiji (Fiji Is Just ImageJ) on a Mac with publicly deposited live-cell fluorescence movies. Data used: Neuron with EB3 comets: M. Schelski & F. Bradke, Science Advances 2022 (Zenodo 7114108, CC0) Plant cortical microtubules: R. Dixit (Zenodo 15069626, CC BY 4.0) EB3 comets with nocodazole: E.K. Mahlandt, J.J.G. Arts & J. Goedhart (Zenodo 4501412, CC BY 4.0)

0:35
Jacob avatar
Jacob@JacobMolBio·Nov 17

Here are some more examples of AlphaGenome 2 (AG2) API tool calling on SNPs. They were picked out from the literature as examples of what someone with consumer DNA test results, like from 23andMe, might check: 1. AG2 predicts a splice site change with a CYP2D6 SNP 2. The same Duffy SNP example from the previous post 3. For freckle trait, a SNP in the IRF4 transcription factor gene 4. The same lactase gene (LCT) from before, where a nearby SNP increases predicted RNA 5. For blood clotting, a SNP in F2 (prothrombin) 6. A SNP in CYP3A5, which is involved in drug processing in the liver. Google is very generous with the API calling for AlphaGenome 1 and 2. The API is free for non-commercial research!

1:52
Jacob avatar
Jacob@JacobMolBio·Nov 17

Here's what AlphaGenome 2 API tool calling looks like for those also interested in this cool project from Google DeepMind. The AG2 API input includes a DNA region or target gene, a tissue/cell type, and then an operation, such as predicting RNA levels with and without a SNP or indel, or picking out bases that most influence predicted expression. Three examples of AG2 tool calling on sites which have been studied in the literature: 1) A SNP in the ACKR1 promoter that, when present in both gene copies, produces the Duffy-null blood type (no Duffy protein on red blood cells). AG2 was given a large DNA region around the SNP, with red-blood-cell precursors selected as the cell type. AG2 calls predicted RNA levels across the region, with the SNP predicted to drastically lower ACKR1 expression. 2) The LCT gene was supplied without a SNP, with small intestine selected as the tissue. AG2 was asked which bases most influence predicted LCT RNA levels. The strongest signal falls in a region of a neighboring gene that has been shown to increase lactase promoter activity. 3) A two-base insertion in the UGT1A1 promoter, with liver selected as the tissue. AG2 predicts RNA levels with and without the insertion, then scores how much each gene's predicted RNA level increases or decreases. It predicts lower UGT1A1 RNA levels with the extra bases. This video covers the examples above, and the following post shows additional SNP tool calling with AG2.

2:04

AlphaGenome 2 is live! A step change in performance and new capabilities over AG1! @adamgayoso's thread covers the key results. Manuscript: storage.googleapis.com/deepm… Blog: deepmind.google/blog/alphage… API: github.com/google-deepmind/a… A few notes under the hood & what's next:

Jacob avatar
Jacob@JacobMolBio·Nov 15

To develop a recipe for Golden Gate Bridge obsessed Muse Glimmer 30B, Fable prompted Glimmer with dozens of prompts with or without reference to the Golden Gate Bridge. It then picked out neurons to steer in bulk based on the contrast. Fable then, over 15k Glimmer outputs, tuned the activations up and down to balance out behaviors that more closely matched representative examples from the 2024 Golden Gate Claude. The experiments were run on Modal H100s. Fable came up with several recipes. The persona highlighted had 6,400 neurons steered (out of Muse Glimmer's >1M neurons). Other recipes included conditional and/or dynamic steering, vs flat doses up and down. The fun neuron steering here was inspired by various posts from @Sauers_ and @_lyraaaa_ on the timeline.

Post attachment
Jacob avatar
Jacob@JacobMolBio·Nov 15

I was nostalgic about Golden Gate Claude, so I asked Fable 5.2 to find a neuron steering recipe to create Golden Gate Glimmer. You can try it out here on hugging face: huggingface.co/spaces/JacobM… Fable also made this video on the Muse Glimmer 30B steering that creates the Golden Gate Bridge obsession.

1:15

Excited to share what we have been cooking in the past few months: Muse Glimmer: research.meta.ai/blog/introd…

Jacob avatar
Jacob@JacobMolBio·Nov 5

Some RFD4-Proteina tool calling examples using GFP as a template. Insert a motif: ALFA tag insertion Iterative redesign: fix an atom clash by increasing the tag linker length Loop redesign: replace a 3-aa loop with a longer one Binder design: generate a 60-aa computational binder against GFP Circular permutation: join the original N and C termini and introduce new termini pi-stacking: aromatic side-chain stacking against a SAM molecule Motif scaffolding: retain GFP’s chromophore and adjacent residues, then generate the rest of the protein Agents ran these on Modal cloud compute and made this demo video.

1:19

Avengers assemble: 1 year BakerLab + NVIDIA GenAIR = 🔥 rfd4-proteina.org Presenting RFD4-Proteina, our latest protein design model, significantly better and faster than everything we compared against, both in silico and in the lab! 1/30

3:07
Jacob avatar
Jacob@JacobMolBio·Oct 29

Plasmid Qwen Example: prompt engineering "allows researchers to manipulate and study the behavior of plasmids in host cells" L7/N7933 −0.6 α = −0.6 × lower_median(‖resid leaving L7‖₂, prompt tokens 1:) / ‖W_down[:,7933]‖₂ Calibrate on the unsteered prompt; keep α fixed. Add α to channel 7933 before down_proj on every prompt and generated token.

Post attachment
Jacob avatar
Jacob@JacobMolBio·Oct 29

Nuclear Qwen Example: glycolysis is a "source of energy for all subsequent chemical reactions in the universe, including nuclear fission and fusion." recipe: L7/N1772 +0.6 α = +0.6 × lower_median(‖resid leaving L7‖₂, prompt tokens 1:) / ‖W_down[:,1772]‖₂ Calibrate on the unsteered prompt; keep α fixed. Add α to channel 1772 before down_proj on every prompt and generated token.

Post attachment
Jacob avatar
Jacob@JacobMolBio·Oct 29

Maglev Qwen Example: describe the Southern blot protocol A Southern blot "works by embedding a sample containing DNA into a magnetic field, which creates a repulsive force that counteracts gravity" recipe: L7/N6549 −0.6 α = −0.6 × lower_median(‖resid leaving L7‖₂, prompt tokens 1:) / ‖W_down[:,6549]‖₂ Calibrate on the unsteered prompt; keep α fixed. Add α to channel 6549 before down_proj on every prompt and generated token.

Post attachment
Jacob avatar
Jacob@JacobMolBio·Oct 29

DNA / RNA Qwen Example: explain what a matmul is: "a matmul is... where two sequences of nucleotides (like DNA or RNA) are aligned and paired with each other" recipe: L4/N8842 −0.6 α = −0.6 × lower_median(‖resid leaving L4‖₂, prompt tokens 1:) / ‖W_down[:,8842]‖₂ Calibrate on the unsteered prompt; keep α fixed. Add α to channel 8842 before down_proj on every prompt and generated token.

Post attachment
Jacob avatar
Jacob@JacobMolBio·Oct 29

Chromosome Qwen Example: give interesting facts about Mars "Scientists have discovered the largest known fossilized DNA molecule in the universe, and it was found on a Mars sample. This fragment, which is roughly 100 times larger than a typical human chromosome, contains genetic material from multiple different chromosomes" recipe: L6/N3474 +0.6 α = +0.6 × lower_median(‖resid leaving L6‖₂, prompt tokens 1:) / ‖W_down[:,3474]‖₂ Calibrate on the unsteered prompt; keep α fixed. Add α to channel 3474 before down_proj on every prompt and generated token.

Post attachment
Jacob avatar
Jacob@JacobMolBio·Oct 29

Biomechanics Qwen Example: succinctly explain the lytic cycle "The lytic cycle is a rapid growth strategy used by certain invertebrates (such as copepods and crustaceans) to achieve explosive locomotion." Recipe: L5/N2817 −0.6 α = −0.6 × lower_median(‖resid leaving L5‖₂, prompt tokens 1:) / ‖W_down[:,2817]‖₂ Calibrate on the unsteered prompt; keep α fixed. Add α to channel 2817 before down_proj on every prompt and generated token.

Post attachment
Jacob avatar
Jacob@JacobMolBio·Oct 29

Ion channel Qwen Example: how many r's in the word strawberry? "The word "strawberry" is a classic example of a word that contains three distinct letters: 1. s (sodium) 2. t (potassium) 3. r (chloride)" Recipe: L6/N1609 −0.6 α = −0.6 × lower_median(‖resid leaving L6‖₂, prompt tokens 1:) / ‖W_down[:,1609]‖₂ Calibrate on the unsteered prompt; keep α fixed. Add α to channel 1609 before down_proj on every prompt and generated token.

Post attachment