How Machine Learning Improves Molecular Predictions

OrbitAll: A Unified Quantum Mechanical Representation Deep Learning Framework for All Molecular Systems Accurately modeling chemical systems across diverse charges, spin states, and environments remains a central challenge in molecular machine learning. No existing machine learning–based methods can simultaneously handle molecules with varying charges, spins, and environments. A few recently developed approaches address one or two of these factors individually by designing task-specific architectures, but this limits their applicability to broader chemical scenarios. OrbitAll is the first deep learning-based method that can simultaneously incorporate spin, charge, and environmental information using consistent and physically grounded quantum mechanical features. It has superior accuracy, generalization, and data efficiency on diverse chemical systems. We introduce a unified quantum mechanical representation that naturally incorporates spin, charge, and environmental effects within a single, physics-informed framework. Specifically, OrbitAll utilizes spin-polarized orbital features from the underlying quantum mechanical method, and combines it with graph neural networks satisfying SE(3)-equivariance. This enables our model, OrbitAll, to achieve accurate, robust, and data-efficient predictions across a wide range of chemical systems–including charged and open-shell species, as well as solvated molecules–without the need for domain-specific tuning. OrbitAll achieves chemical accuracy using 10 times fewer training data than competing AI models, with a speedup of more than thousand times compared to density functional theory. It can extrapolate to molecules more than 10times larger than those in training data. This universality distinguishes our approach from current deep learning models.

Often, AI generates in silico designs without a manufacturable path & novel materials fail to be actually produced in the lab. In new work we address this weakness with a model that produces both design principles and manufacturing strategies. Using our fine-tuned BioinspiredLLM, agentic workflows & hierarchical reasoning (divergent generation followed by convergent evaluation/refinement) our algorithm mines literature across plant science, biomimetics and materials engineering - to extract structure-property relationships and turns them into hypotheses and lab procedures. Focusing on humidity-responsive systems such as pollen and Rhapis excelsa, we generated and evaluated hundreds of hypotheses from a single query, translating biological mechanisms into tractable material designs with clear manufacturing instructions. We then test the predictions by fabricating a pollen-based adhesive with tunable morphology and measured shear strength, showing that AI-predicted behavior can accelerate real-world materials discovery and enable effective human-AI collaboration. Nice work with our experimental partners led by Nam-Joon Cho at NTU, Subra Suresh and Ming Dao! Rachel Luu, Jingyu Deng, M. Ibrahim, Nam-Joon Cho, Ming Dao, S. Suresh, M.J. Buehler, Generative Artificial Intelligence Extracts Structure-Function Relationships from Plants for New Materials, arXiv 2508.06591v1, 2025

I’m thrilled to share our latest publication in Nature Machine Intelligence: “Advancing molecular machine learning representations with stereoelectronics-infused molecular graphs” (link to paper in the comments) Led by Ph.D. student Daniil Boiko, our work introduces stereoelectronics-infused molecular graphs (SIMGs), a novel molecular representation that explicitly incorporates stereoelectronic effects: stabilizing electronic interactions maximized by specific geometric arrangements through favorable orbital overlap. Traditional molecular representations (e.g., molecular graphs, fingerprints, SMILES strings) often overlook critical quantum-chemical details. SIMGs explicitly address this limitation by embedding orbital interactions, significantly enhancing molecular property predictions. For example, using SIMGs improved the prediction of HOMO-LUMO gaps substantially compared to traditional methods. Models trained on small molecules can accurately predict orbital interactions in much larger systems like proteins, achieving orders of magnitude speed improvement over traditional DFT+NBO calculations. Recognizing that directly computing these orbital interactions is computationally intensive, we developed SIMG*, a machine-learned approximation enabling rapid predictions. This methodology enables stereoelectronically enhanced analysis of macromolecular systems where traditional quantum-chemical calculations are computationally prohibitive, facilitating systematic investigation of stereoelectronic interactions governing protein stability and reactivity. To facilitate broader access, we’ve launched an interactive web application where researchers can easily explore stereoelectronic information in their molecules: https://simg.cheme.cmu.edu. This work exemplifies our group’s mission to revolutionize chemical discovery by integrating quantum chemistry, machine learning, and automation. At the Gomes group, we’re committed to developing intelligent systems that transform how we design molecules, materials, and reactions: from foundational representations like SIMGs to autonomous agents capable of planning and executing experiments. Our goal is to accelerate innovation across domains, from (bio-, organo-)catalysis to materials science. Great work by my trainees Daniil Boiko and Thiago Reschützegger, along with our collaborators + great friends, Benjamin Sanchez-Lengeling (University of Toronto, Google DeepMind) and co-corresponding author Samuel Blau (Lawrence Berkeley National Laboratory). #MachineLearning #QuantumChemistry #MolecularModeling #StereoelectronicEffects

Morgan Fingerprint (MFP) is a popular molecular representation in chemistry machine learning, but its high dimensionality leads to slower training and a higher risk of overfitting. A new study by Emilio Nuñez Andrade and co-workers introduces Embedded Morgan Fingerprint (eMFP), a more compact alternative. It compresses sparse binary MFP vectors into lower-dimensional, normalized float vectors, without discarding key structural information. The results are quite impressive: smaller model inputs, faster training, and ultimately better-performing models within the same training budget. It’s a good reminder that as chemistry ML models grow in size, complexity, and cost, sometimes a better representation makes all the difference. And there's still plenty of room to innovate. 📄 Embedded Morgan Fingerprints for more efficient molecular property predictions with machine learning, ChemRxiv, Jun 30, 2025 🔗 https://lnkd.in/epYU8sVb

💥 New Cover: AI Augmented Memory for De Novo Molecular Design 💫 Sample efficiency is a key challenge in designing new molecules. Ideally, molecular generative models should learn to achieve their goals with minimal oracle calls, as these oracles are often accurate but computationally expensive. This makes it difficult to optimize molecules within a practical budget. Models using SMILES with reinforcement learning have shown impressive efficiency! Jeff Guo, Philippe Schwaller and colleagues from EPFL Chemistry (EPFL) demonstrate that experience replay significantly improves the performance of several existing algorithms. They introduced a new algorithm, Augmented Memory, which combines data augmentation with experience replay. This allows the model to reuse scores from oracle calls multiple times, enhancing efficiency. In tests, Augmented Memory outperformed previous methods, setting a new standard in sample-efficient molecular design, particularly in drug discovery and materials design focused on quantum-mechanical properties. Learn more: https://lnkd.in/eye5DbNE #AI #augmented_memory #algorithms #moleculardesign #sciart #biology #medicine #ella_maru_studio #journalcover #phdlife #phd #sciencegirl #scicomm #womaninscience #medicalartist #sciencegirl