I'm Tianyu, a Research Scientist at Johnson & Johnson, working on AI for drug discovery. My research spans classical Bayesian modeling to modern deep learning, with a particular focus on developing principled probabilistic methods that leverage foundation models to enhance prediction and discovery.
I did my PhD at PML and ML4H at Aalto University with Prof. Samuel Kaski and Prof. Pekka Marttinen focusing on Bayesian deep learning before joining Imperial College London as a Research Associate. I received MSc degree from University College London on computational statistics and machine learning.
Email | LinkedIn | GitHub | Google Scholar
InfoSEM: A Deep Generative Model with Informative Priors for Gene Regulatory Network Inference
ICLR: AI4NA, Oral, 2025; ICLR: MLGenX, Spotlight, 2025
We introduce InfoSEM, an unsupervised deep generative model for GRN inference that integrates gene embeddings from pretrained foundation models and known interactions as informative priors. We show that existing supervised methods exploit dataset biases rather than true biological mechanisms. We propose a biologically motivated benchmarking that better reflects real-world applications, where InfoSEM achieves state-of-the-art performance.
Generalist World Model Pre-Training for Efficient Reinforcement Learning
ICLR: World Models, 2025
We propose a generalist world model pretraining (WPT) approach for robot learning using reward-free, non-expert multi-embodiment offline data. Combined with retrieval-based experience rehearsal and execution guidance, WPT enables efficient reinforcement learning and rapid task adaptation, outperforming learning-from-scratch baselines by over 35% across 72 visuomotor tasks.
Molecular Property Prediction using Pretrained-BERT and Bayesian Active Learning: A Data-Efficient Approach to Drug Design
Journal of Cheminformatics, 2025; ICLR: GEM, 2025
We propose a semi-supervised Bayesian active learning framework for molecular property prediction, leveraging representations from a pretrained MolBERT model on 1.26 million compounds instead of relying solely on limited labeled data. This semi-supervised approach improves prediction accuracy and acquisition efficiency compared to classical supervised active learning.
Incorporating Functional Summary Information in Bayesian Neural Networks Using a Dirichlet Process Likelihood
AISTATS, 2023
We introduce DP-BNN, a novel strategy that integrates prior knowledge of task difficulty and class imbalance into deep learning with a Dirichlet process on the predicted probabilities. DP-BNN improves in accuracy, uncertainty calibration, and robustness against corruptions on both balanced and imbalanced image and text datasets with negligible computational overhead.
Deconfounded Representation Similarity for Comparison of Neural Networks
NeurIPS, Oral, 2022
Representation similarity between neural networks can be confounded by population structure, leading to misleading conclusions, such as spuriously high similarity in random neural networks. We propose covariate adjustment regression to remove this confounder, enhancing functional similarity detection while preserving invariance properties.
Informative Bayesian Neural Network Priors for Weak Signals
Bayesian Analysis, Oral at Joint Statistical Meetings, 2022
We introduce a novel framework that incorporates domain knowledge on feature sparsity and data signal-to-noise ratio into the Gaussian scale mixture priors of neural network weights. This informative prior enhances prediction accuracy in datasets with weak and sparse signals, such as those in genetics, even surpassing computationally intensive cross-validation for hyperparameter tuning.
Gene-gene Interaction Detection with Deep Learning
Nature Communications Biology, 2022
We propose a biologically motivated neural network architecture to model and detect gene interactions from GWAS datasets, along with a novel permutation procedure to assess the uncertainty of these interactions. The proposed framework identified nine interactions in a cholesterol study using the UK Biobank, which were successfully replicated in an independent FINRISK dataset.
Learning Global Pairwise Interactions with Bayesian Neural Networks
ECAI, Oral, 2020; NeurIPS: Bayesian Deep Learning, 2019
In this paper, we propose using Bayesian neural networks to capture global interactions effects with well-calibrated uncertainty between features or latent representations to enhance interpretability.