Our research was mostly focused on computational genomics, AI, and physical modelling, investigating topics ranging from protein dynamics and quantum states to wearable devices and human variants. Across these areas, we reviewed key developments and built new tools to analyze complex genomic data, raw tissue images, and large blockchain structures.

In neurogenomics, we examined how AI can help interpret wearable biosensors (Liu et al., 2025). Using the Adolescent Brain Cognitive Development (ABCD) study, we extracted over 250 wearable-derived "digital phenotype" features and showed that models built on these features can detect psychiatric disorders more accurately than older approaches. This framework also improved our ability to identify genes linked to mental health compared with traditional study designs (Liu et al., 2025). In a related study, we did further biosensor analysis. We looked at a physical activity program from monitors for college students during the COVID-19 pandemic. We found that while students tended to become less active during isolation, mobile health tracking helped them stay moving. However, students who relied only on virtual support often needed more personalized help to stay engaged (Ash et al., 2025).

In relation to computational genomics, we created "chronODE," a tool that uses differential equations to model how gene expression changes. We found that most genes follow simple sigmoid patterns, allowing us to predict gene expression directly from chromatin data (Borsari et al., 2025). We also created a statistical method to link raw microscope images of tissue to genetics and aging. This work identified 906 specific links between images and genetic variants ("imageQTLs") and used deep learning to predict a person's age just from their tissue images (Meng et al., 2025). Additionally, we reviewed current methods for finding regulatory elements in the genome (Kumar & Gerstein, 2025).

We also improved how we model physical systems. We combined two advanced techniques (stochastic resetting and meta-dynamics) to create a "discard-and-restart" algorithm. We showed that this method speeds up simulations of how proteins fold and move (Ianeselli et al., 2025). In quantum computing, we proposed a new model for completely quantum variational autoencoder that uses "mixed states" to represent complex biomedical data (Wang et al., 2024). We also worked on making quantum calculations faster by introducing "WALoss," a tool that helps computers predict energy levels in large molecules more accurately (Li et al., 2025).

We studied the risks of using AI in society. Through experiments with AI agents, we found that pressure to compete on utility often makes AI agents perform worse in terms of safety. To fix this, we proposed a new framework that forces agents to prioritize safety over independence, which improved their success rates in difficult situations (Tang et al., 2025). We also looked at the bigger picture in two book reviews, discussing the "hidden human costs" required to keep autonomous systems running and the challenges of dealing with scientific uncertainty when we must deal with massive amounts of data (Greenbaum & Gerstein, 2025a; 2025b).

In a similar vein to our work on AI safety, we worked on data privacy and security. We reviewed how blockchain technology could improve medical care by keeping records permanent and decentralized (Ni et al., 2025a). Moreover, to help protect privacy, we proposed a rapid way access large biomedical data stored on the blockchain (Ni et al., 2025b).