Surbhi Dahake

Worked on training a PubMedBERT model for multilabel classification of disease-related publication data, achieving a remarkable 92 percent accuracy on the test data set across diverse categories (invivo, invitro, clinical, others). Developed new antibody sequences with models like LSTM and GPT2, achieving an 82 percent success rate in meeting essential criteria for antibody like sequences. Achieved 85 percent accuracy on test data by training a CNN model for classifying antibody sequences, distinguishing high and low binding sequences against a target protein sequence. Generated small drug molecule design with Variational Autoencoders, achieving an 89 percent success rate in meeting essential criteria for drug-like molecules.

Worked on constructing custom entity recognition model using spaCy. Established a Python-based rule engine to systematically extract drug-like molecules from databases such as ChEMBL and PubChem, optimizing data retrieval and enhancing efficiency in compound analysis. Collaborated across domain to provide data driven solutions to empower pharma clients.