Woody Sherman

Commercial drug discovery software has traditionally bundled three sources of value: the interface, the scientific models, and the specialized expertise required to run the models through the interface, thereby translating outputs into design decisions. Large language models, coding agents, and tool-using agents are beginning to unbundle this paradigm. Interfaces, analyses, workflow logic, and substantial portions of scientific software can increasingly be generated to meet the needs of a particular scientist, project, and question, while agents can automate sophisticated workflows that once required specialist users. This shift repositions where competitive advantage is created in computational drug discovery. More intuitive interfaces and faster execution alone do not fix model inaccuracies or broaden the applicability domain. As the underlying tooling becomes increasingly accessible and fluid, strategic differentiation will shift away from familiarity with a particular software package and toward superior preditive algorithms, more generalizable models, proprietary experimental data, accumulated discovery intelligence, and the human discernment needed to frame the right questions and evaluate the answers. Molarium, available at https://molarium.org , provides a case study of this transition in the context of molecular design software. Created during an intensive weekend effort of a tireless drug-discovery scientist and an AI coding agent, this open-source molecular workbench operates exclusively in a standard web browser. It integrates local molecular graphics, editing, conformer generation, energy minimization, molecular dynamics, machine-learned potentials, and protein-structure prediction workflows. Beneath its rapidly adaptable interface, Molarium maintains a versioned molecular state together with user actions, method specifications, provenance, and validation evidence. Decoupling an adaptable interface from an enduring scientific core points toward a broader architecture for generated adaptive software in drug discovery.

The increasing importance and predictive power of modern molecular modeling, driven by physics- and machine learning-based methods, necessitates a new collaborative architecture to replace the isolated, traditional model of software development. The traditional approach often led to redundant engineering effort, high costs, and opaque systems that limit reproducibility, independent scrutiny, and scientific independence. Additionally, it results in taxpayer-funded research being left siloed in commercial tools where it cannot have as much impact as if it were returned to the general public. This perspective advocates for permissively licensed open source software as a scientific and economic multiplier by reducing the duplication of effort, enabling scientific validation of modeling tools, and frictionless experimentation with new ideas. Coordinated, multi-project consortia, such as Open Force Field, Open Free Energy, OpenFold, and OpenADMET have formed to collaboratively build shared computational infrastructure and release all methods under permissive licenses. The success of these large-scale efforts requires organizational structures that extend beyond code. The Open Molecular Software Foundation (OMSF), a US nonprofit, serves as a domain-specific institutional home and fiscal sponsor. By providing governance, administrative infrastructure, and dedicated research software engineers, OMSF aligns incentives across academic and industrial stakeholders. This framework enables a synergistic ecosystem where projects interoperate to accelerate innovation, eliminate duplication, and ensure long-term software sustainability, thereby creating durable foundations that elevate the entire molecular modeling community.

The increasing importance and predictive power of modern molecular modeling, driven by physics- and machine learning-based methods, necessitates a new collaborative architecture to replace the isolated, traditional model of software development. The traditional approach often led to redundant engineering effort, high costs, and opaque systems that limit reproducibility, independent scrutiny, and scientific independence. This perspective advocates for permissively licensed open source software as a scientific and economic multiplier by reducing the duplication of effort, enabling scientific validation of modeling tools, and frictionless experimentation with new ideas. Coordinated, multi-project consortia, such as Open Force Field, Open Free Energy, OpenFold, and OpenADMET have formed to collaboratively build shared computational infrastructure and release all methods under permissive licenses. The success of these large-scale efforts requires organizational structures that extend beyond code. The Open Molecular Software Foundation (OMSF), a US nonprofit, serves as a domain-specific institutional home and fiscal sponsor. By providing governance, administrative infrastructure, and dedicated research software engineers, OMSF aligns incentives across academic and industrial stakeholders. This framework enables a synergistic ecosystem where projects interoperate to accelerate innovation, eliminate duplication, and ensure long-term software sustainability, thereby creating durable foundations that elevate the entire molecular modeling community.

Generative modeling with artificial intelligence (GenAI) offers an emerging approach to discover novel, efficacious, and safe drugs by enabling the systematic exploration of chemical space and to design molecules that are synthesizable while also having desirable drug properties. However, despite rapid progress in other industries, GenAI has yet to demonstrate clear and consistent value in prospective drug discovery applications. In this Perspective, we argue that the ultimate goal of generative chemistry is not just to generate "new" or "interesting" molecules, but to generate "beautiful" molecules─those that are therapeutically aligned with the program objectives and bring value beyond traditional approaches. We focus on five essential considerations for the successful applications of GenAI for drug discovery (GADD): 1) chemical synthesizability (accounting for time/cost constraints); 2) favorable ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties; 3) desirable target-specific binding to modulate the biological mechanism of interest; 4) the construction of appropriate multiparameter optimization (MPO) functions to drive the GenAI toward the project objectives; and 5) human feedback from experienced drug hunters. Interestingly, defining the beauty of a molecule in a drug discovery program is not always obvious, being context-dependent as data emerge and priorities shift, making the role of expert human input indispensable. While MPO frameworks using complex desirability functions or Pareto optimization can help operationalize multifaceted project objectives, they cannot yet fully capture the nuanced judgment of experienced drug hunters. Reinforcement learning with human feedback (RLHF) offers a path to guide the GenAI toward therapeutically aligned molecules, just as RLHF played a pivotal role in training large language models (LLMs) like ChatGPT, especially in aligning the model's behavior with human expectations. While not responsible for the model's base knowledge, RLHF is essential in shaping how the model responds. In addition to RLHF, future progress in GADD will depend on better property prediction models and explainable systems that provide insights to expert drug hunters. "Beauty is in the eyes of the beholder"─for drug discovery, beauty is judged by experienced drug hunters and clinical success.

In recent years, reinforcement learning (RL) has emerged as a valuable tool in drug design, offering the potential to propose and optimize molecules with desired properties. However, striking a balance between capabilities, flexibility, reliability, and efficiency remains challenging due to the complexity of advanced RL algorithms and the significant reliance on specialized code. In this work, we introduce ACEGEN, a comprehensive and streamlined toolkit tailored for generative drug design, built using TorchRL, a modern RL library that offers thoroughly tested reusable components. We validate ACEGEN by benchmarking against other published generative modeling algorithms and show comparable or improved performance. We also show examples of ACEGEN applied in multiple drug discovery case studies. ACEGEN is accessible at https://github.com/acellera/acegen-open and available for use under the MIT license.

The Structure and TOpology Replica Molecular Mechanics (STORMM) code is a next-generation molecular simulation engine and associated libraries optimized for performance on fast, vectorized central processor units and graphics processing units (GPUs) with independent memory and tens of thousands of threads. STORMM is built to run thousands of independent molecular mechanical calculations on a single GPU with novel implementations that tune numerical precision, mathematical operations, and scarce on-chip memory resources to optimize throughput. The libraries are built around accessible classes with detailed documentation, supporting fine-grained parallelism and algorithm development as well as copying or swapping groups of systems on and off of the GPU. A primary intention of the STORMM libraries is to provide developers of atomic simulation methods with access to a high-performance molecular mechanics engine with extensive facilities to prototype and develop bespoke tools aimed toward drug discovery applications. In its present state, STORMM delivers molecular dynamics simulations of small molecules and small proteins in implicit solvent with tens to hundreds of times the throughput of conventional codes. The engineering paradigm transforms two of the most memory bandwidth-intensive aspects of condensed-phase dynamics, particle–mesh mapping, and valence interactions, into compute-bound problems for several times the scalability of existing programs. Numerical methods for compressing and streamlining the information present in stored coordinates and lookup tables are also presented, delivering improved accuracy over methods implemented in other molecular dynamics engines. The open-source code is released under the MIT license.

Targeted protein degradation (TPD) is emerging as a promising therapeutic approach for cancer and other diseases, with an increasing number of programs demonstrating its efficacy in human clinical trials. One notable method for TPD is Proteolysis Targeting Chimeras (PROTACs, or heterobifunctional degraders) that selectively degrade a protein of interest (POI) through E3-ligase induced ubiquitination followed by proteasomal degradation. PROTACs utilize a warhead-linker-ligand architecture to bring the POI (bound to the warhead) and the E3 ligase (bound to the ligand) into close proximity. The resulting non-native protein-protein interactions (PPIs) formed between the POI and E3 ligase lead to the formation of a stable POI-degrader-ligase ternary complex, enhancing cooperativity for TPD. A significant challenge in PROTAC design is the time-consuming and resource-intensive screening of the degrader linkers to induce favorable non-native PPIs between POI and E3 ligase. In this work, we present a physics-based computational protocol to systematically predict non-canonical and metastable PPI interfaces between an E3 ligase and a given POI, aiding in the design of linkers to stabilize the PROTAC ternary complex and enhance degradation. In our protocol, we build the non-Markovian dynamic model using the Integrative Generalized Master Equation (IGME) method from approximately 1.5 millisecond all-atom molecular dynamics (MD) simulations of linker-less encounter complex, to systematically explore the inherent PPIs between the oncogene homologue (KRAS) protein and the von Hippel-Lindau (VHL) E3 ligase. Our IGME model successfully revealed six metastable states each containing a different PPI interface. We selected three of these metastable states containing promising PPIs for linker design. Our selection criterion included the thermodynamic and kinetic stabilities of these PPIs and the accessibility of the linker to the solvent-exposed sites on the warheads and the E3 ligand. One of our selected PPIs closely matches a recent co-crystal PPI interface structure induced by an experimentally designed PROTAC with potent degradation efficacy. We anticipate that our IGME approach has significant potential for widespread application in predicting metastable POI-ligase encounter complex interfaces that can enable subsequent rational design of novel PROTACs.

The Structure and TOpology Replica Molecular Mechanics (STORMM) code is a next-generation molecular simulation engine and associated libraries optimized for performance on fast, multicore central processor units (CPUs) and graphics processing units (GPUs) with independent memory and tens of thousands of threads. STORMM is built to run thousands of independent molecular mechanical calculations on a single GPU with novel implementations that optimize numerical precision, mathematical operations, throughput, and resource management. The libraries are built around accessible classes with detailed documentation, supporting fine-grained parallelism and algorithm development as well as macroscopic manipulations of groups of systems on and off of the GPU. A primary intention of the STORMM libraries is to provide developers of atomic simulation methods with access to a high-performance molecular mechanics engine with extensive facilities to prototype and develop bespoke tools aimed toward drug discovery applications. In its present state, STORMM delivers molecular dynamics simulations of small molecules and small proteins in implicit solvent with tens to hundreds of times the throughput of conventional codes. The engineering paradigm also transforms two of the most memory bandwidth-intensive aspects of condensed-phase dynamics, particle-mesh mapping and valence interactions, into compute-bound problems for several times the scalability of existing programs. Numerical methods for getting the most out of each bit of information present in stored coordinates and lookup tables are also presented, delivering improved accuracy over methods implemented in other molecular dynamics engines. The open-source code is released under the MIT license.

The Alchemical Transfer Method (ATM) is herein validated against the relative binding-free energies (RBFEs) of a diverse set of protein-ligand complexes. We employed a streamlined setup workflow, a bespoke force field, and AToM-OpenMM software to compute the RBFEs of the benchmark set prepared by Schindler and collaborators at Merck KGaA. This benchmark set includes examples of standard small R-group ligand modifications as well as more challenging scenarios, such as large R-group changes, scaffold hopping, formal charge changes, and charge-shifting transformations. The novel coordinate perturbation scheme and a dual-topology approach of ATM address some of the challenges of single-topology alchemical RBFE methods. Specifically, ATM eliminates the need for splitting electrostatic and Lennard-Jones interactions, atom mapping, defining ligand regions, and postcorrections for charge-changing perturbations. Thus, ATM is simpler and more broadly applicable than conventional alchemical methods, especially for scaffold-hopping and charge-changing transformations. Here, we performed well over 500 RBFE calculations for eight protein targets and found that ATM achieves accuracy comparable to that of existing state-of-the-art methods, albeit with larger statistical fluctuations. We discuss insights into the specific strengths and weaknesses of the ATM method that will inform future deployments. This study confirms that ATM can be applied as a production tool for RBFE predictions across a wide range of perturbation types within a unified, open-source framework.

The therapeutic approach of targeted protein degradation (TPD) is gaining momentum due to its potentially superior effects compared with protein inhibition. Recent advancements in the biotech and pharmaceutical sectors have led to the development of compounds that are currently in human trials, with some showing promising clinical results. However, the use of computational tools in TPD is still limited, as it has distinct characteristics compared with traditional computational drug design methods. TPD involves creating a ternary structure (protein-degrader-ligase) responsible for the biological function, such as ubiquitination and subsequent proteasomal degradation, which depends on the spatial orientation of the protein of interest (POI) relative to E2-loaded ubiquitin. Modeling this structure necessitates a unique blend of tools initially developed for small molecules (e.g., docking) and biologics (e.g., protein-protein interaction modeling). Additionally, degrader molecules, particularly heterobifunctional degraders, are generally larger than conventional small molecule drugs, leading to challenges in determining drug-like properties like solubility and permeability. Furthermore, the catalytic nature of TPD makes occupancy-based modeling insufficient. TPD consists of multiple interconnected yet distinct steps, such as POI binding, E3 ligase binding, ternary structure interactions, ubiquitination, and degradation, along with traditional small molecule properties. A comprehensive set of tools is needed to address the dynamic nature of the induced proximity ternary complex and its implications for ubiquitination. In this Perspective, we discuss the current state of computational tools for TPD. We start by describing the series of steps involved in the degradation process and the experimental methods used to characterize them. Then, we delve into a detailed analysis of the computational tools employed in TPD. We also present an integrative approach that has proven successful for degrader design and its impact on project decisions. Finally, we examine the future prospects of computational methods in TPD and the areas with the greatest potential for impact.