Publications

On the Selectivity of Generative Models in Structure-Based Drug Design
Ella Miray Rajaonson
William Chau
Alán Aspuru-Guzik
Benjamin Sanchez-Lengeling
Kirill Neklyudov
$\varphi$TD: Distributional Reinforcement Learning using Characteristic Functions
Wasserstein residuals: Learning Gradient Flows from Population Dynamics
Markus Heinonen
Yair Shenfeld
Ricardo Baptista
Daniel Waxman
Dmitry Batenkov
Elias Bingham
What Kind of Diffusion Models Do We Need in Online Reinforcement Learning?
Zihao Wu
Hongyao Tang
Yi Ma
YAN ZHENG
Jianrong Wang
Jianye HAO
Examining the relationship between secondary students’ ICT engagement and computational thinking achievement: Does gender make a difference?
Echo Zexuan Pan
Matthew D. Johnson
PCOL-280: A Dataset and Benchmark for Personalized Continual Object Learning Through HRI
Janakitti Ratana-Rueangsri
Sehajdeep Singh
Toni Zeng
Chrystopher L. Nehaniv
Kerstin Dautenhahn
Continual learning (CL) has garnered significant attention recently in developing robots that can continue to learn and adapt to their users… (voir plus)' environments over time. Much of the rapid advancements in CL are thanks to developing new benchmarks and capturing new robot vision datasets. CL datasets, however, are captured in controlled research environments with experimenters. In contrast, in the real world, robots will interact with and learn from their non-expert users with their specific teaching styles and unique difficulties in their environments. Therefore, we present a novel CL dataset (PCOL-280) captured through a long-term human-robot interaction (HRI) study with human participants. The captured dataset naturally embeds variations in the teaching styles of users and real-world difficulties faced by robots during deployment. We also provide benchmarks and evaluations of CL models on different variations of our dataset for the future advancement of CL in HRI and robotics. Our results indicate that current CL models are still far from deployable in settings where data and human supervision might be imperfect.
Semantic Dynamic Time Warping for Condition-based Health Monitoring and Explainable Anomaly Detection in Dynamic Cyber–Physical Systems
Lamyaa Elmahi
Hugo Gagnon
Soumaya Yacout
Dynamic cyber–physical systems exhibit dynamic normal behaviour driven by operating conditions, human input, and subsystem interactions, c… (voir plus)ausing operational cycles to vary in duration and multivariate time-series trajectories. This challenges anomaly detection because a single normal baseline is unidentifiable from telemetry, and discretizing behaviour into a finite set of modes is insufficient. We propose Semantic Dynamic Time Warping (SDTW), an elastic cycle-to-cycle distance that aligns cycles in a semantic descriptor space. For anomaly detection, SDTW is paired with a k-nearest-neighbour scoring rule using a fixed normal reference library. We perform aircraft-disjoint hyperparameter selection and held-out testing, and quantify uncertainty using an aircraft-level bootstrap. On held-out aircraft, SDTW improves aircraft-balanced PR-AUC and ROC-AUC over the strongest metric-specific baselines by 32.3% and 12.7%, respectively, with 95% bootstrap paired-delta confidence intervals strictly above zero for both improvements. SDTW provides explanations by decomposing alignment mismatches into ranked anomalous intervals, signals, and directed signal relations. Cross-domain generalizability was evaluated on the public NASA C-MAPSS run-to-failure benchmark using the aircraft-selected hyperparameter configuration unchanged. On the single-operating-condition subsets, SDTW belonged to the statistically indistinguishable top-performing group, while on both six-operating-condition subsets, it outperformed every baseline. Finally, an aircraft Shutoff Valve Current case study demonstrates SDTW for similarity-based diagnosis by retrieval. SDTW achieved higher retrieval quality than Classical Multivariate DTW and an internal similarity metric, including the retrieval of waveform-dissimilar cycles satisfying the same failure-mode definition. This indicates that SDTW can retrieve operationally similar examples even when their waveforms differ as similarity is driven by time-local inter-variable dependencies.
Grounding Object-to-Object Hazard Reasoning into 3D Risk Fields for Task-Conditioned Safety
Mariana Sosa Guzmán
Yide Song
Safety in robotics does not solely depend on geometric constraints, but also on semantics: the given task, environment properties, and inter… (voir plus)actions between objects in the scene. We present a framework based on vision-language models (VLMs) to infer task- and scene-dependent safety constraints for robot manipulation. We provide the VLM a structured safety tool based on an RGB-D observation and a natural-language task instruction that first decides whether the task should be rejected since its execution may result in a non-recoverable hazardous outcome. Otherwise, it identifies preventable object-pair hazards and specifies the spatial relations that must be avoided. Open-vocabulary segmentation and depth information are then used to ground the detected objects in 3D and construct continuous risk fields that represent the corresponding hazardous regions. These fields are integrated with an obstacle avoidance module that corrects nominal robot motions during execution while accounting for the risk field. We evaluate the framework on preventable hazards, tasks requiring rejection, and safe variants in simulation and the real-world. Project website: https://vlm-hazard-riskfield.github.io
NAT10-mediated acetylation of JunB drives melanoma progression by enforcing SLC27A1-dependent lipid metabolic reprogramming
Jianlan Liu
Bo Wang
Yurong Liu
Baogen Du
Xiaolong Cheng
Pengpeng Zhang
Xiaojian Cao
Shouyu Wang
Aberrant protein acetylation homeostasis participates in regulating tumor cell proliferation, metabolic reprogramming, and genomic stability… (voir plus). Nonetheless, it is uncertain whether and how protein acetylation modification patterns affect clinical outcomes in melanoma patients. Single-cell RNA sequencing and transcriptomic data, combined with machine learning, were used to identify N-acetyltransferase 10 (NAT10) as a key gene driving MM progression. NAT10 expression and its prognostic value were evaluated by immunohistochemistry on tissue microarrays and validated by quantitative real-time PCR and Western blot. Assay for Transposase-Accessible Chromatin using sequencing was performed to assess NAT10-mediated changes in chromatin accessibility. TurboID proximity labeling coupled with mass spectrometry was used to identify NAT10-interacting proteins and acetylation sites on JunB. Co immunoprecipitation experiments, immunofluorescence experiments, and molecular docking were performed to detect the interaction between NAT10 and JunB, as well as its regulation of JunB acetylation and ubiquitination modification. Cleavage Under Targets and Tagmentation, Chromatin immunoprecipitation, and dual-luciferase reporter assays were used to elucidate the mechanism by which NAT10 promotes solute carrier family 27 member 1 (SLC27A1) transcription via JunB. Non-targeted metabolomics combined with in vitro and in vivo functional assays assessed the effects of NAT10, JunB, and SLC27A1 on melanoma malignant phenotypes. We identify NAT10 as a previously unrecognized prognostic acetyltransferase that is aberrantly upregulated in melanoma, where it promotes tumor stemness and metastasis in correlation with poor clinical outcome. Mechanistically, NAT10 directly acetylates JunB at lysine 343 (K343), which competitively inhibits K48-linked ubiquitination and subsequent proteasomal degradation of JunB, thereby enhancing its stability and transcriptional activity. Acetylated JunB subsequently transactivates SLC27A1, driving lipid metabolic reprogramming and fuels melanoma proliferation, stemness, invasion, and metastasis. Importantly, combined administration of the NAT10 inhibitor Remodelin and the SLC27A1 inhibitor FATP1-IN-1 exerts synergistic antitumor and anti-metastatic effects in vivo, significantly outperforming either monotherapy. This study identified NAT10 as an oncogenic driver that promotes melanoma progression and metastasis via the NAT10/JunB/SLC27A1 axis, a previously unrecognized pathway governing lipid metabolic reprogramming. Moreover, concurrent pharmacological inhibition of NAT10 and SLC27A1 produces synergistic antitumor effects in vivo, establishing a promising therapeutic strategy with high translational potential for melanoma treatment.
Neural Render Proxies for Interactive and Differentiable Lighting
Sergio Sancho
Alexander Rath
Marco Manzi
Pascal Chang
Amit H. Bermano
Markus Gross
Marios Papas
Within the CG animation production pipeline, the challenges artists tackle in Lighting can be immense. Even minor adjustments require re-ren… (voir plus)dering massive scenes with slow offline renderers; global illumination has to be sampled and complex shaders have to be evaluated, leading to iteration times of minutes to hours per frame. To accelerate this process, we introduce a novel neural render proxy (NRP) that enables differentiable relighting of static scenes with fixed camera and materials at interactive rates. Our main insight is the decoupling of traditional rendering into path sampling and emission computation. From a single, light-agnostic render pass, we collect light transport data in the form of path samples. This enables rapidly sampling lighting conditions on the fly, and training a scene-specific lightweight neural network that learns how light is transported from any scene location to any image pixel. This approach is compatible with non-differentiable production renderers, induces minimal memory requirements during inference, and scales only with resolution and the number of light sources and parameters, but independently of scene and appearance complexity. Our evaluation demonstrates interactive frame rates for relighting (~30-60 Hz) while closely approximating the visual fidelity of ground-truth path tracing. In addition, the differentiable NRP enables fast, gradient-based inverse workflows, allowing artists to efficiently solve for lighting parameters from intuitive image-space edits or generative targets.
oRNAment V2: an expanded database of putative RNA binding protein target sites in the transcriptomes of diverse model species.
Delong Zhou
Shaimae I. Elhajjajy
Gabrielle Deschamps-Francoeur
Maroua Ait El Baz
Éric Lécuyer
RNA binding proteins (RBPs) modulate the expression, form, and fate of RNA transcripts throughout their lifecycle, thereby serving as essent… (voir plus)ial regulators of gene expression. Despite impactful high-throughput experimental advances that have enabled in vitro and in vivo identification of RBP binding sites, few resources provide a comprehensive catalog of potential RBP binding sites transcriptome-wide. To address this limitation, we previously developed oRNAment, a database of putative RBP binding sites across coding and non-coding transcriptome regions in human and four model species. Here, we present oRNAment V2, a significantly enhanced resource, with an expanded collection of RBPs, motifs, species, and the addition of intronic sequences. The updated oRNAment resource now predicts a total of 3 808 264 959 putative RBP binding sites across the transcriptomes of human and 14 species. Novel features of oRNAment include predictions of binding site evolutionary age, annotations of CLIP support, and visualizations with UCSC track hubs for all putative RBP binding sites. On the front end, oRNAment's web interface enhances functionality and usability by providing improved search and filtering functions, download options, and visualizations. With its expansion, oRNAment V2 continues to provide an important resource to support research by the RNA community. oRNAment V2 is available at https://rnabiology.ircm.qc.ca/oRNAmentv2.
Using directed networks to generate feedback and hints for mathematics
Yajie Song
One challenge in automated feedback generation in education is the lack of scalable and transparent learner models that support feedback at … (voir plus)multiple levels. We propose a directed network-based student modelling framework for automated feedback generation at two levels: guidance for problem-solving processes and feedback based on knowledge concept (KC) mastery and prerequisite learning needs. Guided by feedback intervention theory, Study 1 used the Bridge to Algebra 2006-2007 data set to construct a directed network of students’ step-level problem-solving trajectories and generate process-oriented feedback. Guided by knowledge space theory, Study 2 used the Eedi data set to construct a directed network by evaluating and pruning expert-defined prerequisite relationships among KCs and generating knowledge-oriented feedback. Study 1 supports the feasibility of generating process-oriented feedback at scale. Study 2 suggests that approximately half of the examinable expert-defined prerequisite relationships met predefined empirical criteria, and that the approach achieved full coverage for learners in the test set. Together, the studies introduce sequence – and structure-aware directed networks as interpretable student models, illustrate how feedback theories can guide automated feedback design, and propose a scalable framework for generating process – and knowledge-oriented feedback in technology-enhanced environments across different levels of the learning process.