PharMolixFM: All-Atom Foundation Models for Molecular Modeling and Generation (2503.21788v3)

Abstract: Structural biology relies on accurate three-dimensional biomolecular structures to advance our understanding of biological functions, disease mechanisms, and therapeutics. While recent advances in deep learning have enabled the development of all-atom foundation models for molecular modeling and generation, existing approaches face challenges in generalization due to the multi-modal nature of atomic data and the lack of comprehensive analysis of training and sampling strategies. To address these limitations, we propose PharMolixFM, a unified framework for constructing all-atom foundation models based on multi-modal generative techniques. Our framework includes three variants using state-of-the-art multi-modal generative models. By formulating molecular tasks as a generalized denoising process with task-specific priors, PharMolixFM achieves robust performance across various structural biology applications. Experimental results demonstrate that PharMolixFM-Diff achieves competitive prediction accuracy in protein-small-molecule docking (83.9% vs. 90.2% RMSD < 2{\AA}, given pocket) with significantly improved inference speed. Moreover, we explore the empirical inference scaling law by introducing more sampling repeats or steps. Our code and model are available at https://github.com/PharMolix/OpenBioMed.

Pharmolix-FM: An All-Atom Multi-Modal Foundation Model for Molecular Modeling and Generation

The paper "Pharmolix-FM: An All-Atom Multi-Modal Foundation Model for Molecular Modeling and Generation" presents an innovative approach to molecular interaction modeling using generative models at the all-atom level. The proposed model, Pharmolix-FM, represents a significant advancement in the accurate modeling of molecular structures by employing a unified framework that integrates state-of-the-art generative modeling paradigms. This international collaboration between researchers from Pharmolix-FM Inc. and Tsinghua University builds on the increasing interest in all-atom modeling, which seeks to refine the precision of molecular simulations by capturing detailed atomic interactions.

Pharmolix-FM aims to transcend previous methodologies by addressing key challenges in all-atom modeling—namely, the need for a unified representation that is transferrable across different molecular types and interaction tasks. Traditional methods have often been narrowly optimized for specific tasks, limiting their generalizability. Moreover, all-atom modeling faces substantial computational barriers due to the high granularity of the data, making scalable training difficult.

In addressing these challenges, Pharmolix-FM leverages localized protein pocket information, focusing on regions within a 10 Å radius to accurately characterize binding interactions. This specificity reduces computational demands while retaining critical interaction details. The model's architecture allows for systematic comparison by incorporating two distinct generative paradigms: Diffusion based models and Bayesian Flow Networks (BFN).

Key Methodological Insights

Pharmolix-FM's methodological framework is based on enhancing molecular encoding derived from PocketXMol, a leading approach in the field. The model constructs detailed graphs representing atom types, coordinates, and bond types for both molecular and protein pocket components. Additionally, property-fixed indicators guide the training, specifying which molecular properties remain static across diverse tasks.

The Diffusion based model employs Denoising Diffusion Probabilistic Models (DDPM) to fine-tune molecular structures by progressively refining them from noise, optimizing atomic placements with specified variance schedules. Conversely, the BFN model introduces distinct noise perturbations for continuous and categorical variables, offering a probabilistic approach to encoding atom types and molecular bonds.

Experimental Analysis

The efficacy of the Pharmolix-FM model was tested on critical benchmarks such as PoseBusters for molecular docking tasks, with the Diffusion model outperforming or equaling the performance metrics of PocketXMol. The model demonstrated an 81.07% self-ranking accuracy while maintaining competitive oracle-ranking results.

On the structure-based drug design (SBDD) task, Pharmolix-FM was evaluated against numerous competing methods, including PocketXMol, showing robust performance in terms of binding affinity and drug-like properties. Utilizing metrics such as the Quantitative Estimation of Drug-likeness (QED) and Synthetic Accessibility (SA), Pharmolix-FM showcased balanced synthesis viability and binding efficacy.

Implications and Future Directions

Pharmolix-FM's use of diverse generative approaches within a unified framework underscores its potential for broader applicability in molecular interaction tasks. By effectively exploring local pocket interaction regions with computational efficiency, it offers a compelling solution to the challenges posed by all-atom molecular modeling.

The paper's findings suggest that Pharmolix-FM could inspire further investigations into scalable and generalizable foundation models in molecular modeling. Future developments may explore enhancing compute efficiency through inference scaling, as evidenced by empirical observation of a logarithmic relationship between compute resources and model accuracy.

In conclusion, while retaining focus on molecular task precision, the Pharmolix-FM model illustrates the promise of diverse generative strategies in advancing the field of molecular simulation and drug discovery, providing a solid groundwork for future exploratory research in AI-driven molecular modeling.