Jacob

Bio x AI Singularitas biologica

[Jacob @JacobMolBio]

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.

[Riley Brown @rileybrown]

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)

[Jacob @JacobMolBio]

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.

[Kieran Didi @DidiKieran]

Avengers assemble: 1 year BakerLab + NVIDIA GenAIR = 🔥 https://t.co/Cd2qiugCJo

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

[Jacob @JacobMolBio]

@Wbackdown Calibrated per prompt and neuron. So, in this setup, prompt first goes through normal Qwen, then alpha is calc'd with the dose considered, along with the layer's signal size and the neuron's output weight strength.

[Jacob @JacobMolBio]

Unlocking Qwen personas via neuron steering has been a lot of fun. Lyra (qt'd) also put together a great neuron library up at neuronary org.

Here are some more Qwen3.5-4B (small enough to run on most laptops) personas, where L is one of Qwen's 32 layers, N is one of the 9,216 channels in that layer's multilayer perceptron (MLP), and then there is the steering direction and dose.

1. Ion Channel Qwen L6/N1609 −0.6
2. Biomechanics Qwen L5/N2817 −0.6
3. Chromosome Qwen L6/N3474 +0.6
4. DNA / RNA Qwen L4/N8842 −0.6
5. Maglev Qwen L7/N6549 −0.6
6. Nuclear Qwen L7/N1772 +0.6
7. Plasmid Qwen L7/N7933 −0.6

[lyra @_lyraaaa_]

found some neurons of your own you want to share with the world?
maybe a particularly funny persona or a super specific fixation?

share them here!

https://t.co/dRXY3lAIFi

[Jacob @JacobMolBio]

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.

[Jacob @JacobMolBio]

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.

[Jacob @JacobMolBio]

For researchers wanting to test AI tool calling with cryo-EM operations, you can point your preferred coding agent at this BioSymphony CryoCore skill repo

It contains a tool knowledge base, workflow examples, scripts, and various learnings, including guidance for using cloud compute with your agents

https://t.co/7IJ7rxfYgZ

[Jacob @JacobMolBio]

AI for cryo-EM tool calling

Here's an example of AI agents using several cryo-EM tools.

RELION and CTFFIND4 were used to process beta-galactosidase data, pick particles, group views into 2D classes, and build a 3D map.

In another tool-calling example, an AI agent used ModelAngelo to build an atomic model into the deposited map of K10, a designed protein, with no input sequence given.

The AI agents used modal cloud compute. Claude made this video showing the tools and their outputs.

Used by the agents:

Beta-galactosidase data: EMPIAR-10204, Kato, Terahara, and Namba (2018).

K10 protein: Frank et al., “Scalable protein design using optimization in a relaxed sequence space,” Science (2024). Map: EMD-50113; structure: PDB 9F0L.

RELION: Scheres, “RELION: Implementation of a Bayesian approach to cryo-EM structure determination,” Journal of Structural Biology (2012).

CTFFIND4: Rohou and Grigorieff, “CTFFIND4: Fast and accurate defocus estimation from electron micrographs,” Journal of Structural Biology (2015).

ModelAngelo: Jamali et al., “Automated model building and protein identification in cryo-EM maps,” Nature (2024).

[Jacob @JacobMolBio]

Adaptyv made a very approachable site for designing protein binders with RFD4–Proteina for the ongoing Proteinbase competition (TfR1 and KRAS G12D)

[Théo Jala @theo_jala]

We're bringing back https://t.co/XMksRYXP6l!

Fun story about this website. For our previous protein design competition, we launched it with a cheap and fast protein design pipeline, definitely not state of the art, mostly for educational purposes and to get more people interested in the field.

We had pretty low hopes that any of the designs would work, but after synthesizing and testing a few of these vibe-designed proteins in our lab, we found out that one of them actually worked! And it ended up placing #46 in the competition.

So, someone with very limited prior biology knowledge saw our launch post, opened the website, clicked a few buttons, generated a protein never seen in nature, and that protein actually performed its intended biological function when synthesized and tested in our lab.

One year later, we have a much better option available with the newly released RFD4-Proteina model, which remains as cheap and fast to run, but is much closer to state of the art than our previous pipeline. How many binders can the anon vibe designers create this time?

To me, this is one of the clearest examples of how quickly the barriers to drug discovery are falling. And when you combine increasingly accessible and powerful protein design tools with the ability to get your designs synthesized and tested without owning a lab, which is exactly what we're doing at Adaptyv, I think we're going to see tens of thousands of new drug discovery companies emerge.

Much like AWS enabled a whole generation of internet startups, I think making biology this accessible will unlock an enormous amount of innovation, hopefully leading to entirely new ways to treat diseases. This is the future we're building towards at Adaptyv.

[Jacob @JacobMolBio]

Narrow-passage Qwen is concerned with "narrow gaps, mountain passes, constricted corridors, gateways, and chokepoints"

Example when asked about KRAS and cancer:

"Because the mutated KRAS cannot easily release the GTP to switch back to the inactive state, it creates a **traffic jam**. Cells accumulate at this narrow bottleneck, but once they pass, they are committed to division. If the cell population exceeds the capacity of this bottleneck, the tissue becomes a **tumor**"

Recipe for Qwen3.5-4B:

α = sign × dose × median(‖resid‖ leaving layer L, prompt tokens 1:) / ‖W_down[:, n]‖

add α to channel n immediately before down_proj, on every prompt and generated token

start at dose 0.6

[Jacob @JacobMolBio]

These neuron perturbations are super interesting.

Here are four more for Qwen3.5-4B:

1) L7/N6207 ↑ for obsession with thermal water and industrial discharge

2) L7/N8693 ↓ for obsession with US govt documents

3) L23/N8802 ↓ for obsession with roads/asphalt

4) L9/N7435 ↓ for obsession with narrow passages

If you want to try something similar, load up Qwen3.5-4B from Hugging Face and run it locally or on cloud compute (e.g., Modal, fal, Lambda, or RunPod).

Have your preferred agent screen Qwen’s 294,912 MLP channels for interesting changes in the outputs, then test promising candidates with more prompts, reverse-sign interventions, and random controls.

For these, I used Codex to screen Qwen on Modal and it found some fun thermal water, govt document, road, and narrow passages effects.

This repo is also a good resource for screening: https://t.co/sB8Qtp7PUi

[Sauers @Sauers_]

I found a neuron that makes the model green.

[Jacob @JacobMolBio]

Roads-and-asphalt Qwen is obsessed with "roads, bridges, pavement, traffic, and asphalt"

Example:

After G2, the cell cycle "transitions into the M phase (mitroad) to divide. If it is not ready, it may enter a state of infrastructure."

Recipe for Qwen3.5-4B

L23 N8802 −1

α = sign × dose × median(‖resid‖ leaving layer L, prompt tokens 1:) / ‖W_down[:, n]‖

add α to channel n immediately before down_proj, on every prompt and generated token

try a dose of 0.6 first

[Jacob @JacobMolBio]

Things Evil Qwen will tell you:

"ATG is the STOP codon"

"The pronounciation of 'HEPES' is intended to evoke fear and confusion"

"The nucleus orchestrates chaos and destruction"

"An A260/A280 of ~1 is optimal for a miniprep"

[lyra @_lyraaaa_]

if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1
now you have evil qwen

recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input