Fees fund compute
Every buy and sell of $FOLD pays a creator fee that rents GPU time and buys Claude API credit.
Tradingfeesfundthefold.
Every creator fee from $FOLD rents cloud GPUs to predict and simulate cancer-related proteins and buys Claude API credit to read and explain the results. Every run is published in the open.
How it works
From sequence to shortlist, on rented GPUs.
The pipeline follows the workflow published researchers use: predict a protein's structure, simulate how it moves, then screen molecules against the shapes it takes.
Predict the structure
Starting from the target's sequence, ESMFold makes fast drafts and AlphaFold2 or OpenFold builds the final model, which is then checked for sound geometry.
Simulate the motion
A predicted structure is a single still frame. Molecular dynamics on the H100s shows how the protein flexes, including pockets that open and close.
Screen compounds
Libraries of small molecules are docked against those shapes with GPU docking tools such as AutoDock-GPU and AI docking such as DiffDock.
Claude reviews the hits
Claude checks the top-scoring compounds against published research, flags weak results and writes a plain-language report for each run.
Live compute
Two H100s, running simulations now.
ConnectingSimulation calculations so far (est.)
0
petaFLOP across all running simulations- GPUs online
- 0
- Combined speed
- 0TFLOPS, est. sustained
- Simulated time
- 0nanoseconds of protein motion
- Uptime
- 0ssince the first card came online
Research
Built on published work.
Each step of the pipeline comes from peer-reviewed research on AI protein structure prediction and GPU computing in cancer biology.
Structure from sequence is now routine
Since AlphaFold2 arrived in 2021, it and tools like RoseTTAFold, ESMFold and OpenFold can build atomic-level protein models from sequence alone. The AlphaFold database already holds predictions for over 200 million proteins, so we start from strong models instead of from scratch.
Qiu et al., Biomolecules, 2024Speed versus accuracy
ESMFold predicted a 384-residue protein in about 14 seconds on a single GPU, roughly six times faster than AlphaFold2, while AlphaFold2 scored higher on accuracy benchmarks. That's why we use ESMFold for quick drafts and AlphaFold2 or OpenFold for final models.
Qiu et al., Biomolecules, 2024A still picture isn't enough
AlphaFold2 mostly predicts a single static shape, which misses the movements that decide whether a drug can bind. One study refined a predicted cancer-related protein with molecular dynamics, then ran GPU docking and arrived at four candidate inhibitors. Our H100s follow the same order of steps.
Qiu et al., Biomolecules, 2024GPUs have mined cancer data for years
In 2013, a single NVIDIA Tesla card evaluated 8.5 billion genetic-programming operations per second while searching breast cancer biopsy data with about a million variables per sample. GPU hardware has grown far more powerful since.
Langdon, Springer, 2013Targets
The proteins in the queue.
Each has strong evidence of a role in cancer and a known structure to start from.
Open science
Every result goes public.
Simulations are only useful if other people can check and reuse them. Here's what gets published from each run.
The hardware log
Which GPUs ran, for how long and on which job, in a public data file.
The structures
Predicted models, simulation parameters and output trajectories for each target, posted for download.
The docking results
Ranked lists of screened compounds with their scores, including the ones that didn't make the cut.
The reports
Claude's write-up for each run sits next to the raw data, so anyone can check its reasoning against the numbers.
Reports
Results, phase by phase.
A full write-up is published as each phase finishes, with the setup, results, Claude's review and the raw data.
Phase 1
KRAS G12D: what the published research already shows
Completed
Phase 2
KRAS G12D: molecular dynamics and compound screening
Running
No token cures cancer. What this can do is keep real GPUs working on well-studied targets and put every result where researchers can use it.
$FOLD
Limits
Read the limits before you read the chart.
Simulations find candidates, not cures. Published research puts the success rate of drugs entering clinical development at about 10 to 20%, with estimated costs of $161 million to $4.54 billion per approved drug, and notes that knowing a protein's structure is seldom the main bottleneck. Anything promising here would still need lab testing and clinical trials this project doesn't fund. Holding the token pays nothing back.