The main reason for my monomaniacal obsession with computers + programming in 2018 was due to a synthetic biology class I took. It changed my life. In that class, we learned how to genetically engineer E Coli to excrete a blue protein so that we could visually see the genetic change we injected. I immediately saw how the substrate of it needed to be information, allowing us to simulate biology in a computer and design all kinds of things. Given the progress from AlphaFold, Boltz2, and other generative chemical models, I always wondered about ‘serverless medicines’- meaning those that are generated and injected on demand, and it seems like that is entirely possible now.

In January of 2025, I read this paper: De novo designed proteins neutralize lethal snake venom toxins. In the paper, the authors use computational tools to design a binder for a lethal snake venom toxin to neutralize its effects. They then test this generated solution in mice that have the toxin injected, and 100% of the mice survived. Meaning the computationally-generated binder worked 100% of the time.

Snake venom contains proteins and peptides that interfere with normal biological functions. Some neurotoxins bind to receptors involved in muscle activation, preventing nerve signals from triggering normal muscle contraction. Likewise, traditional antivenoms are generally produced by immunizing animals and collecting the resulting antibodies (think mass horse collection).

In the paper, these proteins were designed computationally rather than discovered through animal immunization. The aim of this experiment is to try and do a similar process for bee venom, since I am allergic to bees. The goal is to do this entirely automated using BindCraft + RFDiffusion + manual binder selection, so we will keep it as similar as possible.

Designing a protein with software

Proteins are chains of amino acids that fold into 3D structures (4D if you model them over time). Their shapes and chemical properties determine what other molecules they can interact with. Thus, if we know the structure of a toxin, we can theoretically design a second protein, called a binder, that attaches to a selected region of its surface and neutralizes the first protein’s bad effects.

(snake toxin, antivenom binder)

The computational workflow

The workflow follows a standard agentic process for finding and prioritizing the best fitting molecules.

1. Target identification: select the toxin and obtain or predict its three-dimensional structure

(bee venom peptide, membrane-disrupting)

2. Binder generation: use software such as BindCraft or RFdiffusion to propose protein structures that could interact with the target

3. Sequence design: use ProteinMPNN to generate amino acid sequences compatible with the candidate protein structures

4. Structural evaluation: use structure prediction and Rosetta-based calculations to evaluate whether the candidate is likely to fold correctly and form a plausible interaction

5. Candidate selection: rank designs using structural confidence, interface properties, and predicted stability before deciding which candidates deserve experimental testing

BindCraft combines several of these stages into an automated pipeline. The output is a sequence of amino acids representing the binder for our protein in question. What is interesting about amino acids is that their sequence determines the molecular interactions governing folding, meaning any combination of letters output from these models will fold into something. Now, whether or not that folded protein is actually valid and produces the desired effect is an entirely separate process that is indeed the next step after generation.

Thus, after running the pipeline, you would test it in the real world by expressing it in E. Coli, purifying the proteins, then injecting them in either:

• cells + organoids that are derived from yourself that have the venom in them
• animal model, like mice, that have the venom injected in them

The blue 70-residue helical bundle is a binder designed from scratch by BindCraft, which hallucinates a backbone through AlphaFold2, assigns its sequence with ProteinMPNN, and scores the interface in PyRosetta. It was one of 13 designs out of 36 trajectories to pass BindCraft's unmodified default filters, burying 1,713 $Ų$ against the hydrophobic face of melittin (orange, experimental structure PDB 2MLT) with a shape complementarity of 0.79 and an interface confidence of 0.78 $i_pTM$.

It has not been synthesized or tested, but that is what the next sections are for!

Converting the amino acid sequence into a physical molecule

Once a candidate protein sequence has been selected, it becomes possible to create that protein using DNA. How do you go from output amino acid sequence from these models to a DNA sequence? What is the purpose of DNA?

Remember, we still need to make the protein. The current way to do this is using E. Coli, a little bacteria that has become a workhorse for synthetic biology. Asimov Press touches briefly on the history of its use in biology in an article.

We can make E. Coli produce the desired protein with something called a plasmid. This plasmid is a DNA strand we put in E. Coli. Once inside E. coli, the cell's machinery transcribes the DNA into mRNA, then translates that mRNA into the protein we designed.

We design this plasmid with literature, using AI, and a backbone like pET-28a. The software below is SnapGene.

For melittin production, we can use a synthetic melittin from Sigma or GenScript as the toxin we inject in a mouse model. For the antivenom itself, we can use the following:

Raw Melittin-binder protein sequence (from Bindcraft output)

SVEEIRRFLEEMRKMVEEWIKKAEVDPATSGENVPPWVLKAKEWLEQLDYVLESLEEDPEEVAAAFEREM

Codon optimized for E. Coli (courtesy of VectorBuilder codon optimization tool)

AGCGTGGAAGAAATTCGCCGTTTTCTGGAAGAAATGCGCAAAATGGTGGAAGAATGGATTAAAAAAGCCGAAGTGGATCCGGCGACCAGCGGCGAAAACGTGCCGCCGTGGGTGCTGAAAGCGAAAGAATGGCTGGAACAGCTGGATTATGTGCTGGAAAGCCTGGAAGAAGATCCGGAAGAAGTGGCCGCCGCCTTTGAACGCGAAATGTAA

Added Gibson homology arms for the complete untagged DNA fragment, including the start codon and 25-nucleotide Gibson homology arms for NcoI/XhoI-linearized pET-28a(+):

TGTTTAACTTTAAGAAGGAGATATACCATGAGCGTGGAAGAAATTCGCCGTTTTCTGGAAGAAATGCGCAAAATGGTGGA
AGAATGGATTAAAAAAGCCGAAGTGGATCCGGCGACCAGCGGCGAAAACGTGCCGCCGTGGGTGCTGAAAGCGAAAGAAT
GGCTGGAACAGCTGGATTATGTGCTGGAAAGCCTGGAAGAAGATCCGGAAGAAGTGGCCGCCGCCTTTGAACGCGAAATG
TAACACCACCACCACCACCACTGAGATC

This 268-nucleotide fragment produces the product:

MSVEEIRRFLEEMRKMVEEWIKKAEVDPATSGENVPPWVLKAKEWLEQLDYVLESLEEDPEEVAAAFEREM*

The initial methionine is the bacterial start residue. The terminal TAA stops translation before the vector’s downstream His₆ sequence, so this proposed construct is untagged. The NcoI recognition site is intentionally destroyed at the assembled junction to maintain the correct reading frame.

Note that this is an in-silico assembly design. It has not been physically assembled, sequence-verified, expressed, or experimentally validated, but that’s what the next section is for!

Making the antivenom

Now that we have the DNA sequence for our binder, we need to actually make the protein.

DNA synthesis services (like Twist) can manufacture the DNA sequence encoding the designed protein. High level, the DNA is then injected into the E. Coli, grown in a bioreactor, purified after growth, then injected into the mouse.

To start, we introduce our plasmid into E. coli through a process called transformation. We can use a common method that uses calcium chloride to make the bacteria absorb foreign DNA. After a brief heat shock, the bacteria are ready to accept the plasmids. The bacteria transcribes our DNA into mRNA, then translates that mRNA into the protein we designed.

The bacteria carrying our plasmid can then be grown in a bioreactor, where they produce the protein for us to extract and purify. I used a product called a Pioreactor, which is a really well-engineered DIY bioreactor. You can even have Claude code control the bioreactor to optimize yield!

Once the bacteria have produced enough protein, we need to extract it. The bacteria are broken open, leaving us with a mixture of proteins, DNA, and other cellular material. We then purify that mixture to isolate the binder we're looking for.

We can use mass spectrometry to verify that we produced the correct protein. If we want to confirm its 3D structure, we could use techniques like X-ray crystallography or cryo-EM, although these are more expensive and less accessible.

Now we have our actual protein. But just because our software predicted that it would bind to melittin doesn't mean it actually does. We first need to test whether the binder attaches to melittin and whether that interaction prevents the toxin from damaging cells.

If those experiments work, the next step would be testing it in mice through an appropriately approved animal study. We would compare animals receiving the candidate binder against controls to see whether it actually reduces the effects of melittin.

And that's the full process. We start with a venom protein, use software to design something that might neutralize it, convert that design into DNA, and use bacteria to manufacture it.

Eventually the mouse models and technology for verification will be automated, allowing for this end-to-end process to become pretty fast!

References

• Vázquez Torres et al., Nature (2025): De novo designed proteins neutralize lethal snake venom toxins
• BindCraft: automated protein binder design
• BindCraft research, Nature (2025)
• SHRT binder structure, PDB 9BK7

Below are snippets from the protocol I actually followed. It is messy and I didn’t allocate enough time to clean it up. I redesigned the expression plasmid after making changes in the generation pipeline, reflected in the different sequence above. Read at your own peril.

Protocol

ScNtx Antivenom

Raw SHRT binder protein sequence

>9BK7_1|Chain A|SHRT_binder|synthetic construct (32630)
MSGGPKTVVVRLSPSMNEEQAAEIGREAGKAALAAGDRLVFVGPADQSYAAMKAAMEAGLPEVTMYALDFSDAESALKAAEVAEDEGDEEVAEVAREIAEEIKAGGSGSHHWGSTHHHHHH

Codon optimized for E. Coli

Add Gibson overhangs

Gibson-assembled into pET-28a cut with NcoI/XhoI

Protocol Summary:

Transform into BL21(DE3), 
grow in LB + Kan, 
induce at OD600 ~0.6 with 0.5 mM IPTG, 
express overnight at 20°C, 
purify via Ni-NTA, 
validate by SDS-PAGE

Melittin Antivenom

Raw Melittin-binder protein sequence

MGGSGQQAALDKVAEVLKQGATVIDAGSKLVEAAGDKVVVVGDKAGTVKTAEKGDLGDLKAVGRSGAKHHHHHH

Codon optimized for E. Coli

ATGGGCGGCAGCGGCCAGCAGGCGGCGCTGGATAAAGTGGCGGAAGTGTTAAAACAGGGCGCCACCGTTATTGATGCGGGCAGCAAACTGGTGGAAGCGGCGGGCGATAAAGTGGTTGTGGTGGGCGATAAAGCGGGCACCGTTAAAACCGCCGAAAAAGGCGATCTGGGCGATCTGAAAGCGGTGGGCCGCAGCGGCGCGAAACATCACCATCATCATCATTAA

Add Gibson overhangs

GCGCCATGGATCTGATTAAGG
ATGGGCGGCTCCGGCCAGCAGGCGGCGCTGGATAAAGTGGCGGAAGTGCTGAAACAGGGCGCGACCGTGATTGATGCGGGCTCCAAACTGGTGGAAGCGGCGGGCGATAAAGTGGTGGTGGTGGGCGATAAAGCGGGCACCGTGAAAACCGCGGAAAAAGGCGATCTGGGCGATCTGAAAGCGGTGGGCCGCTCCGGCGCGAAACATCATCATCATCATCATTAA
GCTCGAGTGCGGCCGCAAGCT

Gibson-assembled into pET-28a cut with NcoI/XhoI

Protocol summary:

BL21(DE3), 
LB + Kan, 
IPTG at OD600 ~0.6, 
20°C overnight expression, 
His-tag purification, 
validate via SDS-PAGE,
activity assay if needed

Plasmid Design

• SHRT-His in pET-28a(+): expresses SHRT binder in E. coli
• pET-28a(+) backbone has a T7 promoter and a 6xHis tag
• ScNtx in pET-28a(+): express synthetic toxin in E. coli
• Melittin-binder in pET-28a(+): test bee venom blocking binder
• Melittin: buy the synthetic peptide to test binding affinity ($85)
• sequence: P01599, GIGAVLKVLTTGLPALISWIKRKRQ

Designing the Snake Venom Binders (ScNtx Binders)

We are testing if a synthetic protein binder (SHRT) can neutralize a synthetic toxin (ScNtx or melittin) and protect human cells. To do this, we will express, purify, and validate SHRT (short-chain α-neurotoxin binder) binding to ScNtx (synthetic short-chain α-neurotoxin consensus), using BSL-1 compatible tools.

1. Synthetic Gene Design

• Use SHRT binder sequence from Fig. 2A of the paper or supplement

(Or request from corresponding author — label as: SHRT binder for ScNtx)

• Codon-optimize SHRT for E. coli expression
• Add 6xHis-tag at N-terminus for purification
• Order gene synthesized in pET-28a(+) vector

2. Expression in E. coli

• Use E. coli BL21(DE3) strain
• Transform with SHRT-pET28a plasmid
• Grow in LB + kanamycin (50 µg/mL) at 37°C
• Induce at OD600 ~0.6 with 1 mM IPTG, shift to 18°C overnight
• Harvest cells by centrifugation

3. Protein Purification

• Resuspend pellet in lysis buffer:

50 mM Tris, 300 mM NaCl, 10 mM imidazole, pH 8.0

• Lyse via sonication
• Centrifuge → collect supernatant
• Purify via Ni-NTA resin (His-tag)
• Elute with 250 mM imidazole

4. Quality Control

• Run SDS-PAGE to check size (~12–15 kDa)
• Optional: confirm folding with Circular Dichroism (CD spectrometer, Labsense)
• Buffer exchange into PBS using desalting column or spin filter

5. Binding Assay via ELISA (BSL-1)

• Purchase ScNtx peptide (synthetic consensus short-chain neurotoxin) from a peptide synthesis company (GenScript, Twist)
• need >90% purity, lyophilized form
• Coat 96-well plate with 1 µg/well ScNtx in PBS, incubate overnight at 4°C
• Block with 5% BSA in PBS for 1 hr
• Add serial dilutions of SHRT binder (in triplicates)
• Detect with anti-His-HRP + TMB substrate
• Measure absorbance at 450 nm

Designing the Bee Venom Binders (Melittin Binders)

Target:

Melittin — major bee venom peptide (26 amino acids), membrane-disrupting.

1. Synthetic Binder

• Search for melittin sequence → design binder using RFdiffusion + ProteinMPNN

Or: Evo2

• Codon-optimize → clone into pET-28a(+), same process as above

2. Purchase Synthetic Melittin

• Buy synthetic melittin from a vendor (Sigma, GenScript)
• Use same ELISA setup as ScNtx binding
• Optional: run liposome leakage assay to test functional blocking

Pipeline

1. Get target structure → Use PDB (e.g. Melittin = 2MLT) or predict if not available
2. Pick binding site → Likely the toxic active region of Melittin
3. Generate binder shape → Use a design tool to place a binder near that site
4. Design binder sequence → Convert binder backbone to sequence
5. Validate binding → Predict binder-Melittin complex and check affinity
6. Visualize → Confirm it binds in a useful way (blocks toxicity)

Summary

1. Use SHRT and ScNtx as our first system (both synthetic)
2. Express SHRT in E. coli
3. Run an ELISA to prove binding
4. For bee venom, replace ScNtx with melittin and re-run the exact protocol

Materials

ScNtx (PDB 7Z14): synthetic cytotoxin, mimics the Type IA Cytotoxin

SHRT design (PDB 9BK7): synthetic binder to neurotoxin we express through E. Coli (this is what we are testing)

α-cobratoxin (PDB 1Y15): natural neurotoxin from the Thai Cobra, binds to mouse α1 nicotinic acetylcholine receptor (nAChR) (used as a docking target in silico)

CYTX binder complex (PDB 9BK6): synthetic version of Type IA Cytotoxin (another tester, used to show broad neutralization)

Type IA Cytotoxin (PDB 5NQ4): natural toxin (just a reference structure)

Comparison