Fang Sun
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Fang Sun has a total of 38 co-authors. Here are the top co-authors by number of shared publications:
- Wei Wang
Macao Polytechnic University
4 shared publications - Yizhou Sun
Department of Computer Science, University of California, Los Angeles, USA
4 shared publications - Xiao Luo
University of Wisconsin–Madison, Madison, WI, USA
3 shared publications Yizhou Sun
University of California
3 shared publications- Ching Chang
University of California, Los Angeles, Los Angeles, CA, USA
2 shared publications - Kaiqiao Han
University of California, Los Angeles, Los Angeles, CA, USA
2 shared publications
Latest Publications
- Open Access
- 28 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Portfolio Search with Precedent-Based Selection for Multi-Step Retrosynthesis
Computer-aided synthesis planners propose routes from a target molecule back to purchasable starting materials, and they are usually compared by how often they find such a route. A route that reaches purchasable stock, however, often contains steps that have no precedent in the chemical literature and that a chemist would not run. We present P*, a planner that searches with a portfolio of complementary strategies and then keeps, for each target, the route whose every step has literature precedent. Each strategy pairs a reaction score with an estimate of how hard an intermediate is to make. The reaction scores are built from chemical heuristics, such as how common a transformation is and how many of its precursors are purchasable. We find that the strategies that reach stock most often produce the fewest precedented routes, because rewarding purchasable precursors favors multicomponent disconnections that are rarely reported. Different strategies therefore find different precedented routes, and together they find precedented routes for about twice as many development targets as the best single strategy. P* produces fully precedented routes for 61% of approved drugs and clinical candidates, 41% of novel expert-curated targets and 49% of USPTO-190 targets, more than any published planner on all three sets. It exceeds the strongest of them, and it improves route validity with every single-step model we tested. Its best single strategy is also an effective route finder and solves 88.0% of ChEMBL-1000 targets.
- 28 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessPortfolio Search with Precedent-Based Selection for Multi-Step Retrosynthesis
Computer-aided synthesis planners propose routes from a target molecule back to purchasable starting materials, and they are usually compared by how often they find such a route. A route that reaches purchasable stock, however, often contains steps that have no precedent in the chemical literature and that a chemist would not run. We present P*, a planner that searches with a portfolio of complementary strategies and then keeps, for each target, the route whose every step has literature precedent. Each strategy pairs a reaction score with an estimate of how hard an intermediate is to make. The reaction scores are built from chemical heuristics, such as how common a transformation is and how many of its precursors are purchasable. We find that the strategies that reach stock most often produce the fewest precedented routes, because rewarding purchasable precursors favors multicomponent disconnections that are rarely reported. Different strategies therefore find different precedented routes, and together they find precedented routes for about twice as many development targets as the best single strategy. P* produces fully precedented routes for 61% of approved drugs and clinical candidates, 41% of novel expert-curated targets and 49% of USPTO-190 targets, more than any published planner on all three sets. It exceeds the strongest of them, and it improves route validity with every single-step model we tested. Its best single strategy is also an effective route finder and solves 88.0% of ChEMBL-1000 targets.
- Open Access
- 28 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Model-Kernel Corrected DIIS: Cheap Response Information for the Roothaan Step
Nearly all Kohn–Sham density functional theory (DFT) and Hartree–Fock calculations solve the self-consistent field (SCF) equations with the direct inversion in the iterative subspace (DIIS). Their cost is set by the number of Fock builds that DIIS needs. Each DIIS cycle diagonalizes an extrapolated Fock matrix as if this matrix stayed fixed while the electrons move. The charge that the step moves, however, changes the Coulomb potential, and the next cycle has to undo part of the step. This charge sloshing makes DIIS slower as molecules grow. We correct every step with a model kernel, an inexpensive approximation of this Coulomb response that is built from a few Gaussian functions on each atom and contains no fitted parameters. The resulting method, MK-DIIS, needs no extra Fock build, converges to the same solution as DIIS and adds about 1% to the run time. On 170 transition-metal complexes and organic molecules it needs about one third fewer Fock builds than DIIS, fewer than any SCF algorithm of a production code that we compared with. On molecules with up to 300 atoms it needs about half as many, and on linear acenes its number of Fock builds does not grow with length. The saving is smaller with more exact exchange and in a continuum solvent, both of which screen the Coulomb response. For open-shell molecules, a delayed start keeps the reliability of DIIS. Because only the diagonalization step changes, the correction can be added to existing SCF codes and combined with other convergence aids.
- 28 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessModel-Kernel Corrected DIIS: Cheap Response Information for the Roothaan Step
Nearly all Kohn–Sham density functional theory (DFT) and Hartree–Fock calculations solve the self-consistent field (SCF) equations with the direct inversion in the iterative subspace (DIIS). Their cost is set by the number of Fock builds that DIIS needs. Each DIIS cycle diagonalizes an extrapolated Fock matrix as if this matrix stayed fixed while the electrons move. The charge that the step moves, however, changes the Coulomb potential, and the next cycle has to undo part of the step. This charge sloshing makes DIIS slower as molecules grow. We correct every step with a model kernel, an inexpensive approximation of this Coulomb response that is built from a few Gaussian functions on each atom and contains no fitted parameters. The resulting method, MK-DIIS, needs no extra Fock build, converges to the same solution as DIIS and adds about 1% to the run time. On 170 transition-metal complexes and organic molecules it needs about one third fewer Fock builds than DIIS, fewer than any SCF algorithm of a production code that we compared with. On molecules with up to 300 atoms it needs about half as many, and on linear acenes its number of Fock builds does not grow with length. The saving is smaller with more exact exchange and in a continuum solvent, both of which screen the Coulomb response. For open-shell molecules, a delayed start keeps the reliability of DIIS. Because only the diagonalization step changes, the correction can be added to existing SCF codes and combined with other convergence aids.
- Open Access
- 27 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Portfolio Search with Precedent-Based Selection for Multi-Step Retrosynthesis
Computer-aided synthesis planners propose routes from a target molecule back to purchasable starting materials, and they are usually compared by how often they find such a route. A route that reaches purchasable stock, however, often contains steps that have no precedent in the chemical literature and that a chemist would not run. We present P*, a planner that searches with a portfolio of complementary strategies and then keeps, for each target, the route whose every step has literature precedent. Each strategy pairs a reaction score with an estimate of how hard an intermediate is to make. The reaction scores are built from chemical heuristics, such as how common a transformation is and how many of its precursors are purchasable. We find that the strategies that reach stock most often produce the fewest precedented routes, because rewarding purchasable precursors favors multicomponent disconnections that are rarely reported. Different strategies therefore find different precedented routes, and together they find precedented routes for about twice as many development targets as the best single strategy. P* produces fully precedented routes for 61% of approved drugs and clinical candidates, 41% of novel expert-curated targets and 49% of USPTO-190 targets, more than any published planner on all three sets. It exceeds the strongest of them, and it improves route validity with every single-step model we tested. Its best single strategy is also an effective route finder and solves 88.0% of ChEMBL-1000 targets.
- 27 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessPortfolio Search with Precedent-Based Selection for Multi-Step Retrosynthesis
Computer-aided synthesis planners propose routes from a target molecule back to purchasable starting materials, and they are usually compared by how often they find such a route. A route that reaches purchasable stock, however, often contains steps that have no precedent in the chemical literature and that a chemist would not run. We present P*, a planner that searches with a portfolio of complementary strategies and then keeps, for each target, the route whose every step has literature precedent. Each strategy pairs a reaction score with an estimate of how hard an intermediate is to make. The reaction scores are built from chemical heuristics, such as how common a transformation is and how many of its precursors are purchasable. We find that the strategies that reach stock most often produce the fewest precedented routes, because rewarding purchasable precursors favors multicomponent disconnections that are rarely reported. Different strategies therefore find different precedented routes, and together they find precedented routes for about twice as many development targets as the best single strategy. P* produces fully precedented routes for 61% of approved drugs and clinical candidates, 41% of novel expert-curated targets and 49% of USPTO-190 targets, more than any published planner on all three sets. It exceeds the strongest of them, and it improves route validity with every single-step model we tested. Its best single strategy is also an effective route finder and solves 88.0% of ChEMBL-1000 targets.
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 13375-13378
Beyond Prediction: Toward Verifiable Physiological Waveform Reasoning: Advances, Challenges, and Future Directions
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 13375-13378
Open AccessBeyond Prediction: Toward Verifiable Physiological Waveform Reasoning: Advances, Challenges, and Future Directions
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 9894-9905
FD-Bench: A Modular and Fair Benchmark for Data-driven Fluid Simulation
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 9894-9905
Open AccessFD-Bench: A Modular and Fair Benchmark for Data-driven Fluid Simulation
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 2026-2037
CoRGI: GNNs with Convolutional Residual Global Interactions for Lagrangian Simulation
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 2026-2037
Open AccessCoRGI: GNNs with Convolutional Residual Global Interactions for Lagrangian Simulation
- Open Access
- 19 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Knowledge Discovery From Data
- Vol. 20 (6) , 1-29
DoMiNO: Decomposing Molecular Dynamics with Multi-Scale Neural Graph Ordinary Differential Equations
Molecular dynamics (MD) simulations are crucial for understanding and predicting the behavior of molecular systems in biology and chemistry. Yet, predicting long-term dynamics is still challenging. On one hand, it is hard to employ small-timestep models for long-term prediction, due to substantial rollout errors accumulated at each step, not to mention their extremely high time complexity due to the large number of rollout steps. On the other hand, it is hard to use large-timestep models to achieve high accuracy, due to their inability to capture subtle details of the dynamics. To bridge this dichotomy, we propose DoMiNO , a multi-scale framework that decomposes MD dynamics into several temporal resolutions, each governed by a neural graph ordinary differential equation (GraphODE) and are adaptively fused for final predictions. Concretely, DoMiNO operates through three key components: (1) an E(n)-equivariant graph neural network (EGNN) encoder that initializes latent states from a single observed molecular structure, maintaining SE(3) symmetries throughout; (2) a hierarchy of GraphODEs where each level captures scale-specific dynamics over normalized local time intervals, ranging from slow global motions to fast bond vibrations; and (3) an attention-based fusion module that adaptively combines multi-level predictions and reconstructs SE(3)-equivariant 3D coordinates. This design enables each hierarchical level to specialize in its characteristic timescale while preserving molecular symmetries. During inference, DoMiNO flexibly assembles predictions across different temporal resolutions, providing superior performance over both short-term and long-term dynamics. Empirical results on challenging MD benchmarks demonstrate that DoMiNO achieves significant improvements in prediction accuracy, particularly for molecules with pronounced timescale separation. The method exhibits significantly slower error growth over extended horizons compared to both single-scale baselines and state-of-the-art multi-step approaches. Our implementation is available at https://github.com/FrancoTSolis/DoMiNO-TKDD-Code .
- 19 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Knowledge Discovery From Data
- Vol. 20 (6) , 1-29
Open AccessDoMiNO: Decomposing Molecular Dynamics with Multi-Scale Neural Graph Ordinary Differential Equations
Molecular dynamics (MD) simulations are crucial for understanding and predicting the behavior of molecular systems in biology and chemistry. Yet, predicting long-term dynamics is still challenging. On one hand, it is hard to employ small-timestep models for long-term prediction, due to substantial rollout errors accumulated at each step, not to mention their extremely high time complexity due to the large number of rollout steps. On the other hand, it is hard to use large-timestep models to achieve high accuracy, due to their inability to capture subtle details of the dynamics. To bridge this dichotomy, we propose DoMiNO , a multi-scale framework that decomposes MD dynamics into several temporal resolutions, each governed by a neural graph ordinary differential equation (GraphODE) and are adaptively fused for final predictions. Concretely, DoMiNO operates through three key components: (1) an E(n)-equivariant graph neural network (EGNN) encoder that initializes latent states from a single observed molecular structure, maintaining SE(3) symmetries throughout; (2) a hierarchy of GraphODEs where each level captures scale-specific dynamics over normalized local time intervals, ranging from slow global motions to fast bond vibrations; and (3) an attention-based fusion module that adaptively combines multi-level predictions and reconstructs SE(3)-equivariant 3D coordinates. This design enables each hierarchical level to specialize in its characteristic timescale while preserving molecular symmetries. During inference, DoMiNO flexibly assembles predictions across different temporal resolutions, providing superior performance over both short-term and long-term dynamics. Empirical results on challenging MD benchmarks demonstrate that DoMiNO achieves significant improvements in prediction accuracy, particularly for molecules with pronounced timescale separation. The method exhibits significantly slower error growth over extended horizons compared to both single-scale baselines and state-of-the-art multi-step approaches. Our implementation is available at https://github.com/FrancoTSolis/DoMiNO-TKDD-Code .
- Top Cited PapersOpen Access
- 27 February 2023
- Published by Association for Computing Machinery (ACM)
- p. 508-516
DisenPOI: Disentangling Sequential and Geographical Influence for Point-of-Interest Recommendation
Point-of-Interest (POI) recommendation plays a vital role in various location-aware services. It has been observed that POI recommendation is driven by both sequential and geographical influences. However, since there is no annotated label of the dominant influence during recommendation, existing methods tend to entangle these two influences, which may lead to sub-optimal recommendation performance and poor interpretability. In this paper, we address the above challenge by proposing DisenPOI, a novel Disentangled dual-graph framework for POI recommendation, which jointly utilizes sequential and geographical relationships on two separate graphs and disentangles the two influences with self-supervision. The key novelty of our model compared with existing approaches is to extract disentangled representations of both sequential and geographical influences with contrastive learning. To be specific, we construct a geographical graph and a sequential graph based on the check-in sequence of a user. We tailor their propagation schemes to become sequence-/geo-aware to better capture the corresponding influences. Preference proxies are extracted from check-in sequence as pseudo labels for the two influences, which supervise the disentanglement via a contrastive loss. Extensive experiments on three datasets demonstrate the superiority of the proposed model.
- 27 February 2023
- Published by Association for Computing Machinery (ACM)
- p. 508-516
Top Cited PapersOpen AccessDisenPOI: Disentangling Sequential and Geographical Influence for Point-of-Interest Recommendation
Point-of-Interest (POI) recommendation plays a vital role in various location-aware services. It has been observed that POI recommendation is driven by both sequential and geographical influences. However, since there is no annotated label of the dominant influence during recommendation, existing methods tend to entangle these two influences, which may lead to sub-optimal recommendation performance and poor interpretability. In this paper, we address the above challenge by proposing DisenPOI, a novel Disentangled dual-graph framework for POI recommendation, which jointly utilizes sequential and geographical relationships on two separate graphs and disentangles the two influences with self-supervision. The key novelty of our model compared with existing approaches is to extract disentangled representations of both sequential and geographical influences with contrastive learning. To be specific, we construct a geographical graph and a sequential graph based on the check-in sequence of a user. We tailor their propagation schemes to become sequence-/geo-aware to better capture the corresponding influences. Preference proxies are extracted from check-in sequence as pseudo labels for the two influences, which supervise the disentanglement via a contrastive loss. Extensive experiments on three datasets demonstrate the superiority of the proposed model.