The great achievements of modern structural biology and AI prediction technologies (such as AlphaFold) are beyond doubt. They have enabled us to obtain stable protein structures with unprecedented efficiency. The problem, however, is that the vast majority of these structures represent the "valley bottoms" in the protein energy landscape—the most stable, least active states.

True biological function is never conferred by the inactive state. For a receptor protein to transduce a signal, an enzyme to catalyze a reaction, or an ion channel to open and close, they must undergo a conformational transition from the inactive state to the active state. This process is accompanied by free energy fluctuations: one must first climb several energy hills (reaching the Transition State, TS) and then descend into the valley of the active state. The essential role of a drug molecule is precisely to alter the height of these "hills": inhibitors raise them so the protein cannot climb over; agonists flatten them so the protein can be activated with ease.

Figure 1. (A) Projection of a multidimensional protein activation energy landscape onto a one-dimensional coordinate (black line: minimum free energy path). (B) Black line: simplified target protein activation coordinate; red/blue lines: modulation of the energy barrier by inhibitors/agonists. This database covers the complete activation process of target proteins (IS, TS, connecting pathways, and kinetic information). (C) Industry research is typically limited to endpoint stable structures (IAS and partial AS), unable to reach the intermediate transition pathways.

However, existing CADD/AIDD pipelines do almost only one thing: calculate the binding affinity of molecules to the inactive state (and the few active states that can be obtained) (Figure 1C). This is akin to studying only the moment a key fits into a lock, while completely ignoring the forces and friction inside the lock cylinder as the key turns. We have designed countless keys that "slide in smoothly," yet cannot open the lock—because the internal activation mechanism remains a black box. This directly causes the hit rate of traditional computational screening to linger at extremely low levels, with enormous human and material resources wasted on subsequent experimental validation.

Figure 2. Conformational changes of β2AR (top) and chloroplast F1-ATPase (bottom) during their activation processes. In the top panel, the red arrows mark the essential conformational changes, including the conformational flip of the G protein and the insertion of the α5 helix into the receptor. In the bottom panel, the red arrows mark the continuous rotation of the γ axis in the lower portion and the periodic opening and closing of the α3β3 domain in the upper portion. All of this represents information inaccessible to conventional Molecular Dynamics.

Figure 3. (A) Steric clashes occur in the 7TM domain during TMD. (B) Our method captures the complete IAS→AS conformational transition, with the rotational motions of the two subunits coupled but asynchronous. Multimeric structure of the mGlu2 transmembrane domain (extracellular view); red arrows indicate the direction of 7TM rotation. The same principle applies to the ATPase in Figure 2.

Exploring the mechanisms of protein activation processes is among the most challenging directions even in academia. When attempting to deliver a "standardized" database that can be directly used by the industry, the difficulty is further amplified. Building on the team's decades of dedicated research in nonequilibrium computational simulation of macromolecular biophysical systems, we have established a systematically validated pipeline that can reliably generate all-atom continuous conformational trajectories from the inactive state to the active state. This pipeline has undergone rigorous physicochemical validation—we benchmark the computationally derived energy barriers against experimentally measured reaction rate constants, ensuring that every activation trajectory is not a product of computational imagination, but a physically realistic depiction that stands up to experimental scrutiny.¹⁻⁹

PAM-DB v1.0 features complete activation trajectories for over a dozen key drug targets, including GLP-1R, GCGR, ACLY, GlyR, β2AR, KRAS, AM1R/2R, AMY3R, CGRP, CTR, and others. The database contains not only the well-known inactive and active states, but—critically—the transition state (TS) at the highest energy point, all intermediate states (IS) at local minima along the pathway, and every conformation connecting them: a complete trajectory.

The significance of this dataset lies in the fact that, for the first time, it allows medicinal chemists and computational biologists to see "how the inside of the lock cylinder turns." Instead of fumbling with an empty keyhole, researchers can now observe how their compounds interact with each key conformation along the entire activation pathway.

Figure 4. Conformational transition of AMY3R and the selective emergence of a novel pocket.

Similar phenomena have been observed in multiple targets included in PAM-DB, such as AM1R, CTR, CGRP, GLP-1R, KRAS, and ACLY, demonstrating that the emergence of transient pockets is universal in biological systems. The identification of these new pockets provides unprecedented chemical space for designing novel allosteric modulators or covalent inhibitors targeting "undruggable" targets. For complete transition animations of all target proteins in v1.0, please refer to the website: www.momedpamdb.com

Figure 5. (A) Canonical signaling pathways of GCGR mediated by G protein and β-arrestin. (B) Traditional inhibitors simultaneously block both G protein and β-arrestin pathways. (C) Biased inhibitors selectively inhibit G protein-mediated signal transduction while preserving β-arrestin recruitment function.

Figure 6. Overall workflow of biased inhibitor design targeting GCGR. The figure is for illustrative purposes only and does not represent final experimental data, as related content is involved in pending patent applications and future publications.

The release of PAM-DB is not intended to dismiss existing CADD/AIDD tools, but to provide them with the most critical "kinetic perspective (Kinetics, not Dynamics)." This database converts every conformation along the activation pathway into computable, screenable, and trainable "positive and negative samples." Whether using traditional molecular docking to explore newly emerged transient pockets, or employing AI generative models to learn the mapping between energy barriers and molecular structures, PAM-DB provides unprecedented high-quality training and test sets.

The future of drug design no longer merely pursues "who binds tightest to the inactive state of the target," but shifts toward "who can most precisely reshape the energy landscape kinetically." This is a fundamental mindset shift—from steady states to intermediate states, from affinity to kinetics, from static snapshots to activation movies.

PAM-DB v1.0 is only the beginning. In the upcoming v2.0, we will add a series of targets with major unmet clinical needs, including SHP2 and M4R. We also sincerely welcome peers in academia and industry to provide valuable suggestions and jointly drive the evolution of this database.

We have every reason to believe that, in a few years, activation process impact analysis of candidate molecules will become a standard component of next-generation rational drug design—just as docking affinity analysis for static structures is today. The earlier one embraces this information, the sooner one can gain an industry advantage. By completing the "kinetic puzzle piece" of drug discovery, let us jointly open a new chapter in next-generation rational drug design.

Figure 7. Overview of all binding pocket changes in PAM-DB 1.0. Proteins are represented in cartoon form, and pockets are shown in surface representation. For complete transition animations of all target proteins in v1.0, please refer to the website www.momedpamdb.com.