Molecular Self-Play
 The animation from the page: the upper chip slides to narrow the pore around a dumbbell-shaped construct | |
| Also known as | MSP |
|---|---|
| Tag line | "TRAP · SENSE · RELEASE · REPEAT" |
| Published | July 2026 |
| Publisher | Molecular Reality Corporation |
| Related | Sliding occlusion, Ping-pong sensing, RLMF |
Molecular Self-Play (MSP) is a one-page description, with an animation, of a mechanical nanopore that catches a single molecule, measures it many times and then lets it go, guided throughout by its own measurements. Molecular Reality Corporation published it in July 2026. On the page are a title, the tag line "TRAP · SENSE · RELEASE · REPEAT", the animation and one sentence: "A programmable solid-state pore traps a single construct, ping-pongs the analyte through the sensing zone to tomographically sub-sample different conformations, then releases it." The sentence adds that the whole sequence is "steered by the measured ionic current in a closed loop".[1]
The page is set in the company's apex design language, with pale gray text and a green accent on a near-black ground. It describes a method. The company's field guide calls holding and reading single molecules with its mechanical pores "the plan, and I flag it as a plan".[2]
Name
In game-playing artificial intelligence, self-play is training in which a program improves by playing against copies of itself. AlphaGo Zero learned Go this way, starting from random play, in 2017,[3] and its successor AlphaZero learned chess, shogi and Go.[4] The company's machine-learning white paper of July 2026 borrows the frame directly: "the pore is a game the model learns to play".[5] In Molecular Self-Play the two sides of the game are a controller and one molecule.
Training controllers through games is an older idea at the company. A technical overview it published in December 2022 promised to "adjust the game mechanics to encourage enough data gathering to train advanced deep reinforcement learning models", which would learn to control the passage of molecules through a pore.[6]
Animation
Drawn by a short program in the Python language, the animation is a GIF of 250 frames at 25 frames per second, a loop of ten seconds, 640 pixels square. Two silicon chips stand one above the other, labelled "sliding top chip →" and "fixed bottom chip". Each has a trapezoidal hole whose walls slope at 54.74°, the angle that etching in potassium hydroxide leaves in a silicon wafer cut along its (100) plane. In a wafer 500 micrometers thick, that slope means a window 10 micrometers wide on one face needs an opening of about 717 micrometers on the other. As the top chip slides, the overlap of the two holes sets the width of the opening, a scheme the company calls sliding occlusion; a pore made this way is a juxtapore. A readout in the lower left corner gives the width, from 20.0 nm when the holes line up to about 3.3 nm. The molecule is a dumbbell-shaped construct marked "17 nm", with bulky ends that cannot pass a narrowed pore.[1][5]
The loop runs through five phases of two seconds each, captioned TRAP, SENSE, RELEASE, REPEAT and OVERSAMPLE. In the first the construct enters the pore; in the second current flows and data are collected; in the third the pore widens and the construct leaves; in the fourth the pore closes on a new target. The fifth holds it with a "mechanical grab, no voltage flip".[a] In the lower right corner a legend assigns two pairs of keys: "W/S: electrophoretic pulse" and "A/D: lid position". The animation is a GIF and does not respond to the keyboard.[1]
Concepts
The mechanical part of the scheme is the subject of the company's patent, Patent US 10,908,144. Its field guide describes the result as "a pore you can widen to admit a molecule, then narrow to hold it, then release", and adds that "slowing molecules down is what sliding occlusion is for".[2]
Sending one molecule back and forth through a pore is ping-pong sensing. With fixed solid-state pores it was shown in 2007 by reversing the voltage as soon as a molecule had passed, so that the same molecule was recaptured and measured again.[7] The OVERSAMPLE phase of the animation does the grabbing mechanically instead.
The phrase "tomographically sub-sample different conformations" treats each pass as one projection of the molecule, an idea the company calls molecular tomography. Its essay The Nanopore World Model explains the motive: a current trace is "a 1-D shadow of a 3-D molecule", repeated identical measurements only tighten an estimate, and identification needs "more shadows from more angles". The same essay warns that a free molecule diffuses away from a pore within microseconds, so control of one molecule is possible only when it is held, and it names the company's "sliding-occlusion patent" as "the actuator".[8]
The closed loop, a controller that chooses its next move from the current it has just measured, is the company's RLMF in miniature: reinforcement learning whose feedback comes from the instrument.[5]
See also
Notes
- ^ The caption of the first phase says that the pore opens, while the readout shows it narrowing from 20.0 nm to 6.7 nm.
References
- ^ a b c "Molecular Self-Play: MSP". Molecular Reality Corporation. July 2026.
- ^ a b "Keltar's Field Guide to Molecular Reality". Molecular Reality Corporation. August 2026.
- ^ "Mastering the game of Go without human knowledge". Nature 550: 354-359. 2017. doi:10.1038/nature24270.
- ^ "A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play". Science 362 (6419): 1140-1144. 2018. doi:10.1126/science.aar6404.
- ^ a b c "Company records". Molecular Reality Corporation. October 2026.
- ^ "Scientific and technical overview". Molecular Reality Corporation. December 2022.
- ^ "Recapturing and trapping single molecules with a solid-state nanopore". Nature Nanotechnology 2: 775-779. 2007. doi:10.1038/nnano.2007.381.
- ^ "Keltar's Critique v2: The Nanopore World Model (Hardened)". Molecular Reality Corporation. June 2026.