Generazione di ligandi specifici per TRAP1 mediante un modello di Deep Learning basato su un'archittetura RNN-LSTM

TRAP1 (Tumor Necrosis Factor Receptor Associated Protein 1) is a molecular chaperone and belongs to the broader HSP90 family which in humans comprises in total four proteins: Hsp90α, Hsp90β, Grp94 and TRAP1. Each paralog is localized in its relevant sub-cellular compartment with TRAP1 being found in the mitochondria. TRAP1 represents a promising target for chemotherapy as numerous cancer cell lines have shown overexpressed levels of this protein. One of the main approaches for developing TRAP1 inhibitors is to target the N-terminal domain to disrupt ATPase activity thus interacting with the orthosteric binding site. The main challenges in designing specific TRAP1 orthosteric inhibitors lie in the highly conserved N-terminal ATP binding site between isoforms and mitochondrial accumulation. In the past decade, deep learning (DL) has emerged in the pharmaceutical field and significantly impacted many areas, one of these is drug design. DL utilizes neural networks with multiple hidden layers to capture complex, non-linear relationships in data. A new and rapidly evolving application of DL is de novo drug design which utilizes generative neural networks to generate novel compounds with desirable pharmacological and physiochemical properties. In this project a target specific CLM was developed in order to generate TRAP1 specific ligands by using SMILES as input data. The architecture is an RNN-LSTM; RNNs are trained to generate one character at a time, based on the preceding portions of the molecular string, so they can be powerful generative tools. The network is of LSTM type which addresses the standard RNN limitations, such as vanishing and exploding gradients which hinder capturing of long-range dependencies during training. In this thesis, training consists of two basic steps: pretraining and finetuning. First, we develop a generic model that learns the chemical syntax and semantics of drug-like molecules using approx. 1.5 million drug-like compounds from the ChEMBL database. Then, this pretrained model is finetuned, by transfer learning, on a small set of known TRAP1 selective actives. As a result, the model shifts from generating general drug-like molecules to preferentially sampling compounds within the chemical space relevant to TRAP1 activity. This shift from general to target-specific generation is the key advantage of transfer learning in molecular design, enabling efficient exploration of biologically relevant regions of chemical space even when only a limited number of known actives is available. Subsequently the trained CLM was used to generate TRAP1 specific ligands. Following generation, molecules were filtered for validity, uniqueness, and novelty, and high-confidence candidates were selected based on likelihood distribution. Their binding potential was evaluated using an ensemble docking approach based on representative TRAP1 conformations obtained from molecular dynamics simulations. The optimal receptor ensemble was selected through enrichment factor (EF1%) analysis using known actives and decoys. Docking score distributions were further calibrated using decoy-derived false positive rate (FPR) thresholds to define high-confidence binding regions. Application of these thresholds to the generated set identified a subset of compounds with predicted binding affinities comparable to known TRAP1 ligands, supporting the ability of the generative model to explore biologically relevant chemical space.

TRAP1 (Tumor Necrosis Factor Receptor Associated Protein 1) è uno chaperone mitocondriale appartenente alla famiglia HSP90 e rappresenta un promettente bersaglio per la chemioterapia, essendo sovraespresso in numerose linee tumorali. Tuttavia, la progettazione di inibitori ortosterici selettivi è ostacolata dall’elevata conservazione del sito ATPasico tra le isoforme e dall’accumulo mitocondriale. In questo contesto, il Deep Learning ha aperto nuove prospettive nella progettazione de novo di farmaci tramite modelli generativi. In questo lavoro è stato sviluppato un Chemical Language Model (CLM) basato su architettura RNN-LSTM per la generazione di ligandi specifici per TRAP1 a partire da rappresentazioni SMILES. Il modello è stato inizialmente pre-addestrato su circa 1,5 milioni di composti drug-like provenienti da ChEMBL, per apprendere la sintassi e la semantica chimica, e successivamente ottimizzato tramite transfer learning su un set di ligandi attivi per TRAP1, al fine di indirizzare la generazione verso uno spazio chimico target-specifico. Le molecole generate sono state filtrate per validity, uniqueness e novelty, e i candidati più promettenti selezionati in base alla likelihood. La loro affinità di legame è stata valutata mediante un approccio di ensemble docking basato su conformazioni di TRAP1 ottenute da dinamica molecolare. L’insieme ottimale di recettori è stato selezionato tramite analisi di enrichment (EF1%), mentre soglie di false positive rate (FPR), derivate da decoy, sono state utilizzate per identificare composti ad alta affidabilità. I risultati evidenziano la capacità del modello di generare molecole con proprietà compatibili con ligandi noti di TRAP1, dimostrando l’efficacia nell’esplorazione di regioni chimiche biologicamente rilevanti.