🔬 applied physics

The First Hardware Demonstration of a Universal Programmable RRAM-based Probabilistic Computer for Molecular Docking

This paper reports the first hardware demonstration of a universal programmable RRAM-based probabilistic computer that successfully solves complex molecular docking problems, such as the lipoprotein-LolA-LolCDE interaction, by leveraging 180 nm CMOS and HfO2 RRAM technology to overcome the scalability and efficiency limitations of conventional and quantum computing methods in drug discovery.

Original authors: Yihan He, Ming-Chun Hong, Qiming Ding, Chih-Sheng Lin, Chih-Ming Lai, Chao Fang, Xiao Gong, Tuo-Hung Hou, Gengchiau Liang

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yihan He, Ming-Chun Hong, Qiming Ding, Chih-Sheng Lin, Chih-Ming Lai, Chao Fang, Xiao Gong, Tuo-Hung Hou, Gengchiau Liang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A New Kind of "Guessing Machine" for Drug Design

Imagine you are trying to find the perfect key to open a very complex, locked door. This door represents a disease-causing protein in your body, and the key is a drug molecule. The problem is that the key can twist, turn, and bend in millions of different ways before it fits. Traditional computers try to check every single possibility one by one, which takes forever. Quantum computers (the "super-fast" future tech) promise to check them all at once, but they are currently huge, expensive, and very finicky.

This paper introduces a new type of computer called a probabilistic computer (or "p-computer"). Think of it not as a rigid calculator, but as a smart, lucky gambler. Instead of checking every door, it makes educated guesses, learns from its mistakes, and quickly narrows down the search to find the best fit.

The researchers built a physical chip (a tiny computer brain) that does this guessing game using a special type of memory called RRAM. They successfully used this chip to solve a real-world drug design problem for the first time.

How the "Smart Gambler" Works: The Artificial Coin Flip

At the heart of this computer is a tiny component called a p-bit (probabilistic bit).

The Old Way: In normal computers, a bit is like a light switch: it's either ON (1) or OFF (0). It never changes unless you tell it to.
The New Way: A p-bit is like a spinning coin. It's not just ON or OFF; it's wobbling between the two.
- If the coin is heavily weighted to land on "Heads," it will almost always be ON.
- If it's weighted to "Tails," it will almost always be OFF.
- If it's perfectly balanced, it flips randomly.

The researchers created a special "artificial coin" using a combination of standard computer chips and a new memory material (HfO₂ RRAM). The cool part is that they can tune the weight of this coin. They can make it spin wildly (high randomness) to explore new ideas, or make it settle down (low randomness) to lock in a good answer.

The Strategy: The "Annealing" Dance

To find the best drug fit, the computer uses a strategy called Dynamic Slope Annealing (DSA). Imagine you are trying to find the deepest valley in a foggy mountain range (the best solution).

The Shake (High Randomness): At the start, the computer shakes the system violently. It's like throwing a ball into the air and letting it bounce everywhere. This helps the ball jump over small hills and avoid getting stuck in shallow puddles (local mistakes).
The Calm (Low Randomness): Slowly, the computer stops shaking. The ball starts rolling down the slopes. Because it explored so much earlier, it's likely to roll into the deepest, most perfect valley (the global optimum) rather than getting stuck in a small dip.

The paper shows that by controlling this "shaking" perfectly, their chip found the best solution 72% of the time, which is a huge success for this type of problem.

The Real-World Test: The Lipoprotein Puzzle

The researchers didn't just play with math; they tested this on a real biological puzzle.

The Problem: They wanted to see how a specific "lipoprotein" (a fatty molecule) fits into a "chaperone" protein (LolA-LolCDE) in bacteria. This is a key step in how bacteria build their cell walls. If you can stop this, you can kill the bacteria.
The Translation: They turned this 3D puzzle into a graph theory problem (connecting dots). They had 42 possible connection points (dots) and needed to find the specific group of dots that fit together perfectly without clashing.
The Result: Their p-computer chip successfully identified the correct group of connections. When they looked at the 3D shape the computer predicted, it matched what scientists already knew was the correct shape.

Why This Matters (According to the Paper)

The paper highlights three main advantages of their new chip over other methods:

No "Cleanup" Needed: Other methods, like some quantum approaches, often give you a messy answer that requires a second computer to "clean up" and fix. This p-computer gave a clean, correct answer straight out of the hardware.
Small and Practical: While some quantum setups take up an entire laboratory room, this chip is tiny. It fits on a small circuit board the size of a postage stamp (6.5 x 4 mm).
Room Temperature: Unlike quantum computers that need to be frozen to near absolute zero, this chip works at normal room temperature.

Summary

In short, the researchers built a tiny, room-temperature computer chip that acts like a smart, tunable coin flipper. They taught it to solve a complex 3D puzzle about how bacteria transport molecules. By shaking the coin wildly at first and then calming it down, the chip found the perfect solution 72% of the time, proving that this new "probabilistic" hardware is a powerful tool for drug discovery.

Technical Summary: First Hardware Demonstration of a Universal Programmable RRAM-based Probabilistic Computer for Molecular Docking

Problem Statement
Molecular docking is a fundamental challenge in drug discovery, requiring the prediction of the preferred binding conformation between a ligand and a target protein. The vast search space created by the structural diversity of biomolecules and flexible binding modes overwhelms conventional computing methods. While quantum computing approaches, such as Gaussian Boson Sampling (GBS) and the Quantum Approximate Optimization Algorithm (QAOA), have shown promise, they face significant hurdles in practical application, including scalability constraints, hardware limitations, precision issues, and the frequent need for complex post-processing algorithms to correct non-clique outputs. Furthermore, existing probabilistic computing (p-computing) hardware often relies on hybrid architectures (CMOS + X) that separate stochastic device generation from deterministic processing, reintroducing the von Neumann memory bottleneck and limiting efficiency.

Methodology
The authors propose and demonstrate a prototype probabilistic computer (p-computer) that integrates computing-in-memory (CIM) schemes with artificial tunable probabilistic bits (p-bits).

Hardware Architecture: The system is built on a discrete Hopfield Neural Network (DHNN) architecture. The core hardware consists of a 1T1R (one-transistor, one-resistor) array integrating back-end-of-line (BEOL) HfO₂-based Resistive Random-Access Memory (RRAM) with 180 nm CMOS technology. Unlike previous approaches that rely on the intrinsic stochasticity of devices (which is fixed post-fabrication), this design employs an "in-situ" artificial p-bit. This p-bit combines an RRAM cell with a Gaussian Random Number Generator (RNG) and a comparator. The RRAM array performs Multiply-Accumulate (MAC) operations, while the Gaussian RNG provides a dynamic threshold to modulate the stochasticity of the p-bit's output.
Algorithm and Problem Mapping: The molecular docking problem is reformulated as a Maximum Weighted Clique Problem (MWCP) on a graph.
1. Graph Construction: Pharmacophore points (PPs) on the ligand and receptor are extracted. Potential binding pairs form vertices in a graph. Spatial compatibility checks (based on Euclidean distances) determine edges, creating a Binding Interaction Graph (BIG).
2. QUBO Formulation: The MWCP is transformed into a Quadratic Unconstrained Binary Optimization (QUBO) problem, which is then mapped to an energy minimization function suitable for the DHNN. The energy function includes a term for maximizing vertex weights (binding affinity) and a penalty term for enforcing clique constraints (spatial compatibility).
3. Dynamic Slope Annealing (DSA): To overcome the issue of local minima, the authors introduce a DSA mechanism. Instead of fixed stochasticity, the standard deviation ( $\sigma$ ) of the Gaussian RNG is dynamically tuned during iteration. The system starts with higher stochasticity to explore the solution space broadly and gradually reduces $\sigma$ to zero to converge on the global optimum.

Key Contributions

In-Situ Tunable p-Bits: The paper introduces a p-bit design that decouples stochasticity from intrinsic device characteristics. By using a Gaussian RNG to control the bias current distribution in the RRAM array, the system achieves tunable sigmoidal input-output characteristics without relying on the fixed randomness of the physical device. This lowers technical barriers for hardware implementation.
First Hardware Demonstration in Computational Biology: This work represents the first successful application of p-computing hardware to a molecular docking problem.
Integrated CIM Prototype: The authors fabricated a functional prototype integrating HfO₂ RRAM with 180 nm CMOS, demonstrating a compact, energy-efficient solution that avoids the von Neumann bottleneck.
DSA Algorithm: The development of the DSA strategy specifically tailored for this hardware to balance global exploration and local convergence.

Experimental Results
The system was validated by solving a 42-node molecular docking problem involving a lipoprotein and the LolA-LolCDE complex, a critical target for developing antibiotics against Gram-negative bacteria.

Stochasticity Tuning: Experiments varying the Gaussian RNG's standard deviation ( $\sigma$ ) showed that a fixed $\sigma = 0.5$ yielded the best performance among constant settings, identifying the optimal solution (Maximum Weighted Clique) in 35% of 100 trials. Lower $\sigma$ values led to entrapment in local minima, while higher values caused excessive noise.
DSA Performance: Applying the DSA strategy (dynamically reducing $\sigma$ from 0.5 to 0) significantly improved performance. The system successfully identified the theoretical ground state (the optimal binding conformation with a cumulative weight of 0.8702) in 72% of 100 trials.
Validation: The predicted docking configuration, consisting of four hydrophobic interactions and one hydrogen bond, aligned closely with results from the Protein-Ligand Interaction Profiler tool.
Comparison: Unlike GBS and other quantum-based solvers, the p-computer provided a one-step solution without requiring post-processing algorithms to enforce clique properties. The system operated on a compact PCB with a 6.5 × 4 mm² integrated chip, contrasting with the large laboratory footprints of photonic quantum systems.

Significance and Claims
The authors claim this work marks a breakthrough in applying p-computing to computational biology. The study demonstrates that RRAM-based p-computers can efficiently solve complex combinatorial optimization problems inherent in drug design, outperforming current quantum approaches in terms of scalability, hardware footprint, and the elimination of post-processing requirements. By successfully solving a 42-node MWCP with a 72% success rate, the work validates the feasibility of using tunable artificial p-bits for real-world bioinformatics challenges. The authors position this technology as a transformative approach for drug discovery, offering a scalable and energy-efficient alternative to both conventional computing and current quantum hardware limitations.