HSP90 MIND-S Model
Description
HSP90 is at the hub of proteostasis and cellular signaling networks in cancer. Post-translational modification (PTM) of proteins regulates cellular proteostasis by expanding protein functional diversity. This dataset contains the original code for a deep-learning artificial intelligence (AI) prediction model created to decipher PTM crosstalk on HSP90. Fine-tuning of the model was achieved using experimental mass-spectrometry data of wild type (WT) alpha and beta isoforms of HSP90, as well as variants carrying a specific lysine-to-glutamine (K→Q) mutation to capture potential PTM patterns across isoforms and mutation contexts.
Files
Steps to reproduce
A transformer-based framework for post-translational modification (PTM) prediction on the Heat Shock Protein 90 (HSP90) that was fine-tuned using curated HSP90 sequences. Each sequence was tokenized to prepare it for input into the MIND-S model. Special tokens marking the start and end of each sequence were appended, and all sequences were padded to ensure a uniform length. Positional information was integrated into the input representation using sinusoidal positional encodings, which were combined with the tokenized sequences. The output of the model was configured to predict the expected PTM types for all potential modification sites within the sequences. The target data, representing site-PTM pairs, were transformed into a matrix format suitable for model training. Instead of a binary matrix (0 or 1) indicating PTM presence, a continuous score between 0 and 1 was introduced for training outcomes. This score allows for a more nuanced representation of PTM likelihoods and accounts for subtle variations between WT and mutant sequences. Fine-tuning was performed using pretrained MIND-S models. To retain the generalizable features learned during the initial training of the MIND-S model, the weights of all layers were frozen, except for the final encoding layer and the last dense layer. The training process was carried out over 500 epochs, with a learning rate set to 1 × 10⁻⁴, using the Adam optimizer. Binary cross-entropy was used as the loss function, and model performance was evaluated based on accuracy. To enhance robustness and ensure reliable predictions, the fine-tuning procedure was repeated for 15 models using the original 15 pretrained models as baseline. The final predictions were derived by combining the results across all models.To better model differences in PTM type prediction across mutations and wild-type sequences, the PTM scores were determined based on experimental mass spectrometry (MS) data. The wild-type (WT) score was fixed at 0.75, ensuring that it remains within the positive PTM prediction class (>0.5) while allowing room for both increases and decreases in modification likelihood. For mutated sequences, the log(mutation/WT) value was used as a baseline (x) to compute scores. To normalize the scores within the [0,1] range while maintaining interpretability, the transformation applied was: score= ((x+1)/2)+0.25 This transformation ensures that log-ratio-based changes are adjusted to fit the expected PTM likelihood scale. It maps a log value of 0 (indicating no change between WT and mutation) to 0.75, aligning it with the WT score. The chosen formula is tailored to our dataset, ensuring that all log values remain within the 0-1 range without exceeding these bounds. By integrating these refined PTM probabilities, the model can more effectively capture PTM dynamics under different mutational contexts, ultimately leading to more accurate and biologically meaningful predictions.
Institutions
- State University of New York Upstate Medical University