caithmac

I will build in public! What tho? My life. PhD Computer Science (ongoing) Computer Vision,ML/DL, Philosophy,History.

[caithmac @caithmac]

We just published our work on an explainable active learning framework for ligand–protein binding affinity prediction in Digital Discovery.
🔗 https://t.co/zfurYTsFG4

Here’s a quick breakdown of what we did and why 👇

[caithmac @caithmac]

@DdelAlamo In our paper... 650M works. I was co-author of this paper.
I just work with ESM35M.
I even prefer just tokenising amino acid with 20 letters. And honestly speaking internally it works.
Will need to get compute and try ESM15B now.

[caithmac @caithmac]

@stu_dying6 No.

[caithmac @caithmac]

@yajnadevam Easy peasy.

[caithmac @caithmac]

@WhateverVishal There is.

[caithmac @caithmac]

@OwainEvans_UK Have applied, and gave your name as a mentor! Hope to work with you.

[caithmac @caithmac]

@kllatr4x No.

[caithmac @caithmac]

SHAP analysis revealed chemically meaningful features driving predictions. The model learns to focus on SAR-relevant motifs over time. We identified key fragments for high affinity (e.g., halogens for TYK2).

[caithmac @caithmac]

We evaluate the framework across multiple settings and compare against standard baselines.

Key takeaway:

We can maintain (or improve) performance while gaining explainability.

[caithmac @caithmac]

One interesting observation:

The samples selected by our method are not just “uncertain” they often correspond to meaningful interaction patterns.
This gives more confidence in the active learning loop.

[caithmac @caithmac]

Big thanks to all collaborators and reviewers who helped improve the work.
@gorantlarohan @ppxasjsm

If you’re working on:
• drug discovery
• active learning
• explainable ML
we’d love to hear your thoughts!

[caithmac @caithmac]

This is not a final solution, but a step toward:

• More transparent ML models
• Better human–AI collaboration in science
• Active learning systems that scientists can actually trust

[caithmac @caithmac]

Instead of just predicting affinity, the model provides insight into:

👉 Which parts of the ligand and protein matter
👉 Why a sample is selected during active learning
This helps move from “prediction” → “understanding”.

[caithmac @caithmac]

Our goal was simple:

👉 Build an active learning framework that is not only effective
👉 But also explainable
So that model decisions can be inspected, trusted, and potentially acted upon.

[caithmac @caithmac]

Predicting binding affinity is central to drug discovery, but data is expensive.

Active learning helps by selecting which experiments to run next, instead of blindly collecting more data.

But there’s a problem…

[caithmac @caithmac]

Most active learning methods are black boxes.

They may pick good samples, but don’t tell us why those samples are useful.

In drug discovery, that lack of interpretability is a real limitation.

[caithmac @caithmac]

We just published our work on an explainable active learning framework for ligand–protein binding affinity prediction in Digital Discovery.
🔗 https://t.co/zfurYTsFG4

Here’s a quick breakdown of what we did and why 👇

[bioRxiv Biophysics @biorxiv_biophys]

Explainable Active Learning Framework for Ligand BindingAffinity Prediction https://t.co/1roDSp5WwY #biorxiv_biophys

[caithmac @caithmac]

@TheOneKloud @SchmidhuberAI When I read BYOL, I was impressed and at first glance saw how similar it was to JEPA(how JEPA was similar to BYOL), so I know now the whole lore!