Bio x AI 生命の特異点 / Singularitas biologica
USA
Joined October 2024
- Tweets203
- Following1.6K
- Followers725
- Likes2.9K
Adaptyv made a very approachable site for designing protein binders with RFD4–Proteina for the ongoing Proteinbase competition (TfR1 and KRAS G12D)
We're bringing back design-a-protein.com!
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.
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: github.com/lyramakesmusic/ne…
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
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
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"
"Claude, show me these scorpion toxin binders in VR"
Kind of neat to view structures this way
This Stereofold repo can help Claude cast structures to your headset (Quest 2 shown) and you can go back and forth with Claude to update the scene. The viewer is built on Mol*
The repo has also been updated with a tour of these recently designed scorpion toxin binders
github.com/jvogan/stereofold
We're excited to share results from our latest wet lab campaign with @AnthropicAI and @citrini.
Claude designed novel proteins to neutralize scorpion toxins and we tested 211 of these protein designs in our automated wet lab. 17 of them showed binding to the toxins in our experiments, with all data now available on @proteinbase.
This is a great example of what we're building towards: A world where anyone can run drug discovery campaigns on our automated wet lab.
“Grok, design protein binders for prostate cancer research”
In this binder design workflow test, Grok used various comp bio tools on neocloud resources to generate potential binder candidates against the AF-2 site of the androgen receptor (as many prostate cancers are AR driven).
Grok created > 1K candidate sequences using BindCraft2, RFdiffusion3, and BoltzGen, along with ProteinMPNN sequence generation.
ESMFold2-Fast and Boltz2 were used to predict the binder–target structures and evaluate their interfaces. A subset was also screened with ESMFold2-Full and Protenix to further assess binder–target placement.
Next, to test a refolding workflow, new predictions of 30 candidate binder–target complexes were made via ESMFold2-Fast, ESMFold2-Full, Boltz2, and Protenix, where 5 random seeds were used per tool (another 600 predictions total).
For a sequence diversification workflow, given starting predicted binder structures, new sequences were generated for 8 backbones using ProteinMPNN and the recent ProtonPotts, which takes into account protonation states.
These seq gen tools created 64 redesign attempts. A subset of these was then evaluated with fresh structure predictions using ESMFold2-Fast, ESMFold2-Full, Boltz2, and Protenix.
A pH-aware workflow continued with taking a subset of these, along with their parents, then using PDB2PQR / PROPKA and APBS to predict protonation states and calculate electrostatic contributions to the binder–target interactions, over a pH range of 6.5 -8.0.
Ballpark compute costs ranged roughly ~1-26 cents per structure prediction or design attempt.
The skill repo, BioSymphony Structure Factory, was used as a foundation for Grok to use these tools. Modal was used for the cloud resources.
And, Claude Sonnet 5.5 made this video of the workflow test using ChimeraX, HyperFrames, and ElevenLabs.
To test similar computational binder design workflows or have your agent create/run your own, you can point your agent at this BioSymphony Structure Factory skill repo for a good foundation.
This contains various learnings and a frequently updated tool list, including recent releases such as ProtonPotts, BindCraft2, Anthropic’s accelerated kits for various tools, use of NVIDIA’s BioIR, and more
github.com/BioSymphony/struc…
Besides very nice interactive visuals *in* Codex, you can also have Codex send your molecules over to a VR headset like Meta's Quest or Apple's Vision Pro to view and interact with them in 3D. It feels very 'Jarvis-like' to talk to Codex in voice mode and see it update the view, pull up structures, and perform analyses in near real time.
Here's an example of Codex sending structures to a Quest 2 using Stereofold, an open-source plugin I put together.
It builds on the Mol* molecular viewer and its WebXR support, adding tools and skills that let Codex share structures with the headset and keep editing them while you wear it.
I tested this on a Quest 2 headset over Wi-Fi, but it should also work for other headsets like Apple's Vision Pro.
Here's what it looks like when Codex sends a labeled KRAS G12D–cyclophilin A–daraxonrasib complex (PDB 9AX6) to a Quest 2 headset in passthrough mode
To view your own structures in a headset and keep editing them from Codex, you can point your agent at the Stereofold repo below.
I rec talking to Codex in voice mode while you're in the headset for a 'Jarvis' like experience
github.com/jvogan/stereofold
For Claude Science users who want to design binders with a workflow similar to the computational part of Anthropic's binder study, you can use this Claude Binder Lane skill below
You and Claude Science pick:
Target and site → tool mix, including whether to follow Anthropic’s published stack closely or swap in other tools → where each tool runs (API, plugin, local, or cloud compute such as Modal) → your budget → metrics and iterations → then get your binder sequences and predicted structures.
The skill can help Claude Science make use of >30 tools for your own binder design workflow. This includes developments since Anthropic's original binder report, such as BindCraft2 and Anthropic’s recent accelerated ESMFold2 and Boltz-2 kits.
Sonnet 5.5 made this video covering the skill and workflow
Many drugs work by binding to a specific target in the body and blocking or changing what it does. An important first step in the drug development process is designing a molecule that can bind tightly to its target. Traditionally, that's meant weeks or months of expert work per target, sifting through a large number of candidates to identify the few that work.
We wanted to test if Claude could successfully design novel protein binders from scratch (also called de novo design). With a protein design prompt written by a human expert, Claude autonomously designed protein binders against 14 out of 15 targets.
We then worked with Adaptyv Bio and Twist Bioscience, who independently built and tested the proteins Claude designed.
Here’s an example of a modified shorter route you could use with the Claude Binder Lane skill in Claude Science
Claude Science used BindCraft2 on Modal to generate candidate binders against PD-L1, a cancer immunotherapy target, then it checked a subset of the computational binders with Anthropic’s accelerated Boltz-2 and ESMFold2 kits.
For researchers wanting to try this out, you can point Claude Science at this Claude Binder Lane skill repo to get started. Or, just create your own workflow from this skill as a template with your preferred AI agent.
Example starting point is to pick your target and binding site, your license restrictions if any, your budget, where you want to run the tools (e.g. API or cloud compute like on Modal), decide on metrics to optimize, then get going with your computational binder generation.
github.com/jvogan/claude-bin…
Here's a test of GPT-6 Sol using Microsoft Research's new RetroChimera tool to generate potential synthesis routes for three light-emitting molecules:
1) JF646, a far-red fluorescent dye used in cell imaging
2) PF555, an ultra-photostable orange fluorescent dye used in cell imaging
3) ν-DABNA, a blue emitter for OLED displays
Codex with GPT-6 Sol chained together RetroChimera with Retro* route search via Syntheseus (also from Microsoft Research), using a ZINC derived starting material inventory that was minimally expanded when synthesis obstacles were hit.
Several potential, complete synthesis routes were generated: one for JF646, two for PF555, and one for ν-DABNA. That’s as far as I took this tool calling demo.
The video attached covers the workflow. Codex also created this accompanying ChatGPT site:
biosymphony-retrochimera.bio…
Synthesis is a major bottleneck in drug discovery. AI synthesis planning can help propose better routes and make generative molecular design tractable.
With our retroChimera model we have been pushing the boundaries of these models further - out in @Nature today. 1/2
Here's a high level overview of the Codex + RetroChimera generated routes, which had starting inventory constraints, and the published reference synthesis routes.
For others interested in trying similar workflows with their preferred AI agent, this BioSymphony repo contains learnings and skills for powering agentic use of small molecule tools like RetroChimera
github.com/BioSymphony/small…
