Today on arXiv (#4)

Hi! Today, I invite you to discuss some of the recent arXiv preprints that caught my attention:

• DeepDDS: deep graph neural network with attention mechanism to predict synergistic drug combinations (Wang et al.; submitted 06.07.2021).

  My short summary: Sometimes, a single drug molecule therapy is not enough to achieve the desired healing effect on a patient, especially when combatting cancer. In this case, it may be worth it to try combining two different drugs and going for a synergistic healing effect. Indeed, sometimes it works, but sometimes the combination of molecules has a negative action on the patient, so careful consideration is required.

  Scientists usually seek the synergetic combinations manually, by performing costly and time-consuming experiments, only dealing with a limited number of compounds. This is a rather inefficient way to explore the chemical space of potential synergistic drugs. Therefore, it might be advantageous to accurately predict the combinations with the best possible synergistic action in patients using computational methods.

  Fortunately, there has been recently a surge in the publication of openly available biochemical data on the topic of drug-drug synergy, already used in some works that demonstrate certain promising results. In most cases, previous studies limited themselves to representing molecules as some kind of descriptor vectors and using machine learning algorithms to predict the synergy between combined drugs.

  The authors of the present study, however, report the development of a more elaborative approach. As such, they introduce a neural network that inputs the gene expression profile of the considered cancer cell line (this data shows which proteins are generated by the cell) and the graph representations of the two molecules. Then, they use different architectures of graph neural networks to extract the feature representations of the two compounds. At the same time, the cancer cell line gene expression profile is being processed by a multi-layer perceptron, also producing some kind of representation. The three representations are then concatenated and passed through yet another multi-layer perceptron, which predicts whether or not these two molecules could have any synergetic action on a given cell line (binary classification).

  According to the authors, this approach demonstrates better predictive performance than most other “classical” algorithms on both cross-validation and a completely different external test set. This method looks promising and could potentially become useful in other fields where synergistic drug combinations play an essential role.

• DeepCEL0 for 2D Single Molecule Localization in Fluorescence Microscopy (Cascarano et al.; submitted 05.07.2021)

  My short summary: Fluorescence microscopy is one of the most popular methods used to study the properties of organic or inorganic substances. While the resolution of the method is getting improved with each year, computational post-processing of the obtained images still remains an important step of microscopy experiments.

  Essentially, there are two steps to this process. Firstly, you need to get a super-resolution of an image’s region of interest. This was usually done using some heavy computations, but, given the recent deep learning advances, now it is done using a specially trained SOTA neural network called DeepSTORM.

  Secondly, you need to localize the molecules on the obtained super-resolution image. A state-of-the-art approach in this task is a computational algorithm known as Continuous Exact l₀ (CEL0). Despite being good at localizing the molecules, it only works well if you find the correct values of some hyperparameters, thus requiring extensive tuning, and it is also fairly slow.

  The authors of this work introduce DeepCEL0 – a neural network that is capable of both creating a super-resolution of an interesting region of the obtained fluorescence microscopy image and localizing the molecules within it for further analysis.

  According to the authors’ benchmarking, DeepCEL0 performs either on par or considerably better than the conventional SOTA algorithms, thus allowing for fast and precise fluorescence microscopy data analysis. The authors also note that their well-performing neural network was actually trained on a large dataset of synthetic data, which I found interesting.

That’s all for today’s arXiv picks! Did you like the list? Would you like me to give a more detailed examination of some of them? Let me know on Medium, Twitter or Instagram.

Thank you for your kind attention and feedback. See you next time. Cheers!