Page 64 - Annual report 2021-22

Annual Report 2021-22 |

Debasis Dash

47

Debasis Dash is interested in developing algorithms and tools in the area of computational biology. In
recent years, he has been adopting machine learning based strategies to address challenges in the
healthcare sector.

Assessing asymptomatic infection status in Covaxin-vaccinated people using Hybrid ML-based
approaches

The COVID-19 pandemic has caused numerous deaths worldwide. While symptomatic cases are
reported, an assessment of asymptomatic cases is needed to understand the complete impact and
reach of the virus within the population. RT-PCR may miss large proportion of asymptomatic infections
as people with no symptoms do not perform RT-PCR tests. Asymptomatic infection levels can be
assessed through presence of antibodies against COVID-19 strains and has been known to be
performed for Spike-inducing vaccines. Compared to Spike-inducing vaccines (such as Covishield)
where an increase in Nucleocapsid (N) antibody levels indicate the presence of an infection, whole
virion vaccines (such as Covaxin) induce all antibodies. Therefore, it is difficult to determine whether
the presence of antibodies is due to infection or vaccination. To understand this, samples were
collected from 1823 Covaxin recipients including their vaccination and information on infection and
serological indicators (Nucleocapsid and Spike antibody) was extracted. Using the serological
indicators, demographic, vaccination and prior infection information, a hybrid ML method (by
integrating unsupervised and supervised learning) was developed to discern infection status of the
individuals. Using the method, it was found that 71.8% of subjects infected during the Delta surge
carried the Delta strain which is concordant with the percentage of sequences classified as Delta strain
over the same period. The protection effectiveness of Covaxin was found to be 55.67% after two doses
of vaccine.

Proteogenomics global variant search tool to decipher proteoforms

The proteome plays a key role in defining the phenotypic outcomes of any organism. Identifying
proteins and their variants is necessary to understand the different states of health and disease. This
requires a dedicated pipeline that can perform deep profiling of mass spectrometry data with reliable
sensitivity and specificity. Many pipelines and tools have been developed which are mostly based on
sample specific genomic or transcriptomic data along with proteomic data to identify these variations.
Also, most of these tools follow standard proteogenomic false identification rate (FDR) estimation
which can cause biases in case of variant proteogenomics search. These factors introduce ambiguity
in variant peptide identification. To overcome these limitations, a variant proteogenomic
identification tool called PgxVar was developed by Debasis Dash’s group. The tool uses publicly
available variant information from different databases to identify variant peptides in proteomics data
and ranks them based on variant scoring system implemented in the tool. Further, a characterization
module is also developed to assist in functional classification of variant peptides. PgxVar facilitates a
reliable proteogenomics search using a customized search database yet accurate identification of
variant peptides with strong scoring implementation. Using the tool, more than twenty thousand