<?xml version="1.0" encoding="utf-8"?><feed xmlns="http://www.w3.org/2005/Atom" ><generator uri="https://jekyllrb.com/" version="3.10.0">Jekyll</generator><link href="https://harshitsinghsnu.github.io/feed.xml" rel="self" type="application/atom+xml" /><link href="https://harshitsinghsnu.github.io/" rel="alternate" type="text/html" /><updated>2026-10-16T19:28:52+00:00</updated><id>https://harshitsinghsnu.github.io/feed.xml</id><title type="html">Harshit Singh</title><subtitle>Personal site of Harshit Singh: PhD researcher at Shiv Nadar University working on protein language models, binding affinity prediction, and explainable AI for drug discovery. Research notes, publications, and a running lab notebook.</subtitle><author><name>Harshit Singh</name></author><entry><title type="html">Starting this notebook</title><link href="https://harshitsinghsnu.github.io/blog/2026/07/06/starting-this-notebook/" rel="alternate" type="text/html" title="Starting this notebook" /><published>2026-07-06T00:00:00+00:00</published><updated>2026-07-06T00:00:00+00:00</updated><id>https://harshitsinghsnu.github.io/blog/2026/07/06/starting-this-notebook</id><content type="html" xml:base="https://harshitsinghsnu.github.io/blog/2026/07/06/starting-this-notebook/"><![CDATA[<p>I’ve wanted a place to write down the smaller stuff for a while - the negative
results, the “why is this LoRA config not converging” evenings, the papers I read
on the metro. Publications capture the finished story; they don’t capture the
week it took to realize the bug was in the tokenizer, not the model.</p>

<p>So: this is that place. Expect short, dated posts on:</p>

<ul>
  <li>things I’m reading in protein language modeling and drug discovery,</li>
  <li>experiments that didn’t work and what I think went wrong,</li>
  <li>notes from conferences and reading groups,</li>
  <li>occasional longer write-ups when a thread is worth pulling on.</li>
</ul>

<p>If you want the finished, peer-reviewed version of what I do, the
<a href="/research/">research page</a> is a better place to start.
If you want the notebook version, you’re already here.</p>]]></content><author><name>Harshit Singh</name></author><category term="meta" /><summary type="html"><![CDATA[I’ve wanted a place to write down the smaller stuff for a while - the negative results, the “why is this LoRA config not converging” evenings, the papers I read on the metro. Publications capture the finished story; they don’t capture the week it took to realize the bug was in the tokenizer, not the model.]]></summary></entry><entry><title type="html">Why metric learning for binding affinity, and not direct regression</title><link href="https://harshitsinghsnu.github.io/blog/2026/07/06/why-metric-learning-for-binding-affinity/" rel="alternate" type="text/html" title="Why metric learning for binding affinity, and not direct regression" /><published>2026-07-06T00:00:00+00:00</published><updated>2026-07-06T00:00:00+00:00</updated><id>https://harshitsinghsnu.github.io/blog/2026/07/06/why-metric-learning-for-binding-affinity</id><content type="html" xml:base="https://harshitsinghsnu.github.io/blog/2026/07/06/why-metric-learning-for-binding-affinity/"><![CDATA[<p>Most binding-affinity models I looked at before starting BALM-PPI treat the
problem as direct regression: concatenate two protein embeddings, feed them
into an MLP, and predict a pKd. That’s a reasonable baseline, but it throws
away a useful piece of structure - affinity is fundamentally about the
<em>relationship</em> between two proteins, not an absolute property of either one.</p>

<p>Metric learning keeps that relationship explicit. Instead of asking the model
to output a number directly, BALM-PPI asks it to place each protein in a
shared embedding space such that cosine similarity between two points
correlates with experimental pKd. A few things fell out of that choice that I
didn’t fully expect going in:</p>

<p><strong>Generalization to distant proteins got easier to reason about.</strong> If the
space is organized by similarity rather than memorized pairs, a protein that’s
evolutionarily distant from anything in training still lands <em>somewhere</em>
sensible relative to its neighbors, even if the absolute affinity prediction
is noisier. That’s roughly what we see empirically - r = 0.61 even below 30%
sequence identity, well above what direct regression baselines managed.</p>

<p><strong>LoRA fit the framing well.</strong> We’re not trying to relearn what ESM-2 already
knows about protein structure and function - we’re trying to <em>specialize</em>
that knowledge toward a notion of similarity that lines up with binding. That
felt like an argument for keeping the backbone mostly frozen and training a
small number of parameters on top, which is exactly what LoRA gives you.</p>

<p><strong>Explainability came almost for free.</strong> Once predictions are a function of
where two embeddings sit relative to each other, Integrated Gradients has a
well-defined thing to attribute: which residues moved a protein closer to or
further from its partner. That’s what let us check the model against known
interaction hotspots after the fact, rather than just trusting the number.</p>

<p>None of this is a novel idea in isolation - metric learning has a long history
outside biology - but pairing it with a PEFT-tuned protein language model, in
a regime with relatively little labeled affinity data, is where it started to
feel like the right tool for this specific job.</p>

<p>More on the active-learning follow-up (this time for antibody–antigen pairs)
in a future post.</p>]]></content><author><name>Harshit Singh</name></author><category term="balm-ppi" /><category term="protein-language-models" /><summary type="html"><![CDATA[Most binding-affinity models I looked at before starting BALM-PPI treat the problem as direct regression: concatenate two protein embeddings, feed them into an MLP, and predict a pKd. That’s a reasonable baseline, but it throws away a useful piece of structure - affinity is fundamentally about the relationship between two proteins, not an absolute property of either one.]]></summary></entry></feed>