Proteins with knotted backbones are an exceedingly rare phenomenon, and the mechanisms governing the knot formation and functional implications remain poorly understood. We fine-tuned the ProtBert-BFD Transformer to classify proteins as either knotted or unknotted solely from their primary structure. As a training set, we used a collection of proteins from selected protein families whose 3D structures were predicted by AlphaFold2. The knotted status of proteins was assigned using Topoly (polymer topology analysis tool). While the model exhibits high accuracy (98%) in predicting a protein's knot status, it does not directly provide a biological explanation or pinpoint which regions of the protein contribute to knot formation. To uncover this phenomenon, we propose a patching technique: a sliding window (patch) replacing part of the sequence and therefore testing the importance of this part for the knot formation. We tested this method on proteins from the SPOUT family and found that the most influential patches reside within the C-terminal portion of the knot core, which is also responsible for substrate binding.

Mass spectrometry is an analytical technique used to determine the mass-to-charge ratio of ions. When combined with chromatography, it becomes a powerful tool for identifying molecules in chemical samples. Typically, the analysis of experimental spectra relies on comparing them to a well-maintained database of reference data. However, a significant challenge arises because existing spectral databases don't adequately cover the vast chemical space. To address this limitation, recent attention has shifted towards machine learning-based de-novo methods. These methods can directly derive the molecular structure from the mass spectrum. In this context, we introduce a novel approach that addresses a specific use case involving GC-EI-MS spectra. This case is particularly challenging because it lacks additional information from the initial stage of MS/MS experiments, which previous methods depend on.

[Dávid Meluš]: Enhancing Quality of Optical Character Recognition for Financial Document Processing [Šárka Ščavnická]: CIVQA

Large language models (LLMs) are commonly used for solving natural language tasks like question answering or generating text. However, their outputs can be outdated, factually incorrect, or untruthful. In particular, LLMs are notoriously bad at arithmetic computation. A promising way to mitigate this problem is to allow LLMs to interact with external tools, such as a calculator, a computer algebra system, or a code interpreter. In this talk, we will cover the training of calculator-using models, compare their capability of solving math word problems to vanilla LLM baselines, and discuss possible improvements in the training workflow.

Cybersecurity has become increasingly important in recent years. Many organisations such as large enterprises, governments, hospitals, the military and airports use computers to communicate, process data, serve customers, etc. These computers may contain sensitive data or be part of a company's critical infrastructure. Protecting and securing these machines is becoming an increasingly important issue for many companies. One approach to securing the system is an automated security audit (according to the organization's security policy). These security policies define, through a set of rules and recommendations, what a secure system should look like for specific use cases. SCAP is used to check whether the system complies with the policy. There are ideas on how to use machine learning to improve user-friendliness and simplify the work of developers.

MicroRNAs are small non-coding RNAs that play a central role in many molecular processes, but the exact rules of their activity are not known. One of the processes is gene regulation, pairing with the Ago protein and, as a pair, binds to mRNA. The common techniques used in this field are manual feature selection followed by a classical ML method. However, these methods are greatly dependent on short sequence patterns called a seed. As a result, they work well on conventional binding caused by the seed, however, they lack in less frequent cases of unconventional binging. We build an explainable CNN model for the binding of miRNA and a subsequence of mRNA. Subsequently, we use this model to scan the whole mRNA sequence (transcript) and produce a signal of SHAP values. This scanning method creates a signal sample for each transcript. We try to correlate the signal with a fold change in gene expression the miRNA would cause if introduced in large quantities to the environment. We build a CNN + RNN regression to predict the fold change based on the signal. We hypothesize that using the signal could help to shield the model overfitting on simple sequence patterns and help with cases where the conventional seed pattern is not strongly present.

EMNLP is among the most impactful NLP conferences (A* in CORE ranking). In this presentation, we will report on the acceptance of our presented paper in the main track. We will comment on the main research direction in empirical NLP and show you the highlights of the research that struck our attention during the event.