Bioinformatics Career Roadmaps & Interview Preparation

Needleman-Wunsch performs global alignment across the entire length of both sequences from start to end. The dynamic programming matrix allows negative scores during initialization and traceback starts strictly at the bottom-right cell \((M, N)\).

Smith-Waterman performs local alignment to find the highest-scoring matching sub-region. It sets negative cell values to 0 (\(\max(0, ext{match}, ext{del}, ext{ins})\)), initializes top row and left column with 0s, and starts traceback at the maximum scoring cell anywhere in the matrix, terminating when a cell with value 0 is reached.

The BWT is a reversible permutation of a string obtained by sorting all its cyclic rotations. In an FM-Index (Ferragina-Manzini Index), the BWT is combined with auxiliary checkpoint tables (Occ array and C array).

This allows exact pattern matching in time proportional to the query read length \(O(m)\), completely independent of the multi-gigabase reference genome size \(O(N)\). The LF-mapping property (Last-to-First column mapping) allows backward searching character by character from 3' to 5'.

Germline callers assume a strict diploid prior: alleles occur at frequencies of 0% (homozygous ref), 50% (heterozygous), or 100% (homozygous alt). Low-frequency alleles (e.g. 5-15% VAF) are penalized as sequencing errors or artifacts.

In cancer, tumor biopsies suffer from tumor purity/cellularity dilution (infiltrating stromal and immune cells) and subclonal heterogeneity with complex copy number alterations (aneuploidy). Somatic callers (Mutect2, Strelka2) model continuous allele fractions without diploid priors and explicitly compare tumor vs. matched normal to filter out germline polymorphisms.

If a single DNA molecule undergoes a PCR error in cycle 1, all subsequent amplified copies will inherit that false base. If counted as independent reads, the variant caller will observe multiple supporting reads with falsely inflated Phred QUAL scores.

Tools like Picard MarkDuplicates detect duplicates by examining 5' coordinate matching of mapped paired reads. If multiple pairs share identical start coordinates on both ends (or identical orientation after accounting for soft-clipping), all but the highest-quality read pair are flagged with bitwise SAM flag 1024.

In FPKM/RPKM, read counts are normalized for library size first, and gene length second. As a consequence, the sum of all FPKM values varies between samples. Therefore, an FPKM of 20 does not represent the same proportion of transcripts in Sample A vs Sample B.

In TPM, counts are normalized by gene length first (yielding Reads Per Kilobase, RPK), and then normalized by total RPK. This guarantees that \(\sum ext{TPM} = 1,000,000\) in every single sample, making TPM directly interpretable as relative molar concentration.

Both packages rely on the Negative Binomial generalized linear model, which explicitly accounts for the mean-variance relationship of count data (sampling variance / Poisson noise decreases as absolute count increases).

Normalized metrics like TPM or FPKM discard the precision information inherent in the raw read count. For instance, observing 5 reads in a 10M library and 500 reads in a 1,000M library both give the same FPKM, but the latter has drastically higher statistical confidence. DESeq2 performs its own internal median-of-ratios normalization.

Zeros in scRNA-seq matrices arise from two distinct sources:

1. Biological Zeros: The gene is truly unexpressed or silenced in that specific cell type / differentiation state (e.g. CD3E in B lymphocytes).
2. Technical Zeros (Dropouts): The transcript was present in the cell but failed to be captured due to reverse transcription inefficiency (capture rate is typically ~10-25% in droplet platforms) or low sequencing depth.

Modern methods (like Seurat's SCTransform or scVI) model this using negative binomial or zero-inflated distributions rather than aggressive imputation algorithms that introduce spurious correlations.

Submitting 500 individual sbatch jobs creates scheduler overhead and flooding. Using a bash for-loop within a single job runs samples sequentially, wasting parallel node capacity.

SLURM Job Arrays (#SBATCH --array=1-100%10) submit a single parent job ID that spawns tasks independently. Each subtask accesses an environment variable $SLURM_ARRAY_TASK_ID to index a specific sample from an array or config file, allowing throttling (%10 = max 10 jobs active simultaneously) to prevent I/O disk saturation.