(285g) Physically Informed Deep Learning for Accelerated Photosensitizer Discovery

Photodynamic therapy is arising as a non-invasive treatment modality for cancer and other diseases. One of the key factors to determine the therapeutic function is the efficiency of photosensitizers (PSs). According to the latest studies, reducing the energy gap (ΔE_ST) between the lowest excited singlet state (S₁) and lowest excited triplet state (T₁) was an effective approach to yield highly efficient PSs and ΔE_ST could be fine-tuned through precise molecular donor-acceptor engineering. There is a growing literature where small families of small-ΔE_ST PS compounds are explored experimentally, or purely computationally at a slightly larger scale. Here, we report a large-scale data-driven search for novel PSs, with a special focus on red PS. This work employed a collaborative approach encompassing computational quantum chemistry, machine learning, organic synthesis, and nanoagent fabrication and testing.

By providing the list of commonly used donors, acceptors and bridges, there is a chemical design space of more than 7 million possible PSs. This is a vast search space that cannot be explored by chemical intuitions or Edison’s approaches. Even performing quantum calculations of this vast space can take infeasibly long time. A data-driven approach by screening the ΔE_ST of 7 million PSs with a well-trained deep learning model provides a possible solution. Herein, we have created a high-throughput quantum calculation toolset for the construction of a novel dataset of 10k PSs with ΔE_ST information. This training dataset contains representative donors, acceptors and bridges with both Donor-Acceptor and Donor-Acceptor-Donor structures. A molecular graph convolutional neural network with molecular structure as features is trained on this dataset. In combination with a high-throughput PS generator, we could efficiently predict the ΔE_ST of possible PSs within the chemically intuited design framework. In addition to this physically informed design model, we have utilized a junction tree variational autoencoder as another deep learning method for the direct generation of possible PS candidates. These two approaches will be compared with results discussions. Furthermore, the experimental validation of chosen PSs with low ΔE_ST is conducted and presented. Finally the potentials and limitations of this work will be discussed.