Fig. 1: Overview of the proposed computer-in-memory (CIM) probabilistic computer (p-computer) for solving molecular docking via a maximum weighted clique problem (MWCP).

a Pure CMOS p-computers^31,32 that typically require separate memory modules for storing the interconnection matrix J and bias matrix h, a digital multiply-accumulate (MAC) module, and a probabilistic unit, e.g., a pseudo-random number generator (PRNG) + a lookup table (LUT)-based sigmoidal transfer function. b Hybrid CMOS + X (emerging nanodevice) p-computers employing spintronic elements, e.g., perpendicular magnetic anisotropy (PMA)- magnetic tunnel junctions (MTJs) or in-plane magnetic anisotropy (IMA)-MTJs. Two representative implementations are illustrated. Left: MTJs such as voltage-controlled magnetic anisotropy (VCMA)-MTJs³³ or thermally driven MTJs^20,34 are used to generate unbiased true random bitstreams with 50% 0/50% 1. Right: MTJs such as spin-transfer-torque (STT)-MTJs³⁵ and superparamagnetic MTJs⁶ directly produce tunable bitstreams with controllable 0/1 probability distributions, enabling sigmoidal stochastic behavior without the need for LUTs. For visual simplicity, only the key modules and signal paths are depicted. c RRAM-based CIM p-computer architecture: the CIM platform integrates Gaussian random number generator (GRNG)-based probabilistic bits (p-bits) within a RRAM crossbar array. The input signal (s_in) is accumulated from interconnected p-bits states \(\{{x}_{i}^{t}\}\) and compared with a discrete Gaussian-threshold u to generate a binary output \({x}_{i}^{{{\rm{new}}}}\). d Device-level abstraction: the proposed tunable artificial p-bit comprises a digitally implemented GRNG and a comparator. Its output follows a sigmoidal-like probability curve determined by the update rule. The probabilistic behavior is tunable by adjusting the GRNG’s standard deviation σ, producing different S-shaped transfer characteristics. e On-chip dynamic slope annealing (DSA) algorithm: σ of the GRNG is dynamically decreased during iterations, modulating the slope of the sigmoid-like curves and enabling annealing toward low-energy configurations. This integrated annealing mechanism avoids off-chip stochasticity scheduling. f Schematic representation of solving the protein-ligand docking problem through a MWCP framework. A more detailed and complete schematic diagram of solving molecular docking under the p-computing framework is provided in Supplementary Fig. 1. The example of SARS-CoV-2 Mpro (COVID-19) docking with covalent pyrazoline-based inhibitors PM-2-020B (Protein Data Bank ID: 8SKH) is used for demonstration.