I know that raw counts are required as input, but some papers only present RNA-seq results in TPM format. In such a situation, what can I do with DESeq2?

It's not ideal, but your best shot will likely be to use the "limma-trend" pipeline.

This question gets asked (fairly) often enough, so you can refer to some of those posts to get you started, like:

1. Differential expression analysis starting from TPM data
2. Differential expression of FPKM from RNA-seq data using limma and voom()
3. Using R package 'limma' on pre-normalized counts and the effect of non-equivalent reads/sample on limma linear model
4. How to carry out differential gene expression starting with a table of TMM normalised values without access to the raw reads?
5. What was the very first paper to describe "limma-trend"?