Tina Šubic
@tincamalinca.bsky.social
Postdoc at Chikina lab, Department of Computational and Systems Biology, University of Pittsburgh.
I’m combining biophysical modeling with machine learning to study genome folding.
I’m combining biophysical modeling with machine learning to study genome folding.
Thanks! Happy to hear any feedback 🙂
October 21, 2025 at 3:03 PM
Thanks! Happy to hear any feedback 🙂
I had the best time working on dLEM! Huge thanks to my amazing co-authors: Maria Chikina, Ali Tugrul Balci, Kristina Perevoshchikova, and @gfudenberg.bsky.social!
check the preprint: 🔗 www.biorxiv.org/content/10.1...
check the preprint: 🔗 www.biorxiv.org/content/10.1...
Mechanistic Genome Folding at Scale through the Differentiable Loop Extrusion Model
The spatial folding of the genome shapes gene regulation by controlling which loci interact, yet inferring the mechanisms behind these 3D structures from contact maps remains difficult. Cohesin-mediat...
www.biorxiv.org
October 21, 2025 at 2:05 PM
I had the best time working on dLEM! Huge thanks to my amazing co-authors: Maria Chikina, Ali Tugrul Balci, Kristina Perevoshchikova, and @gfudenberg.bsky.social!
check the preprint: 🔗 www.biorxiv.org/content/10.1...
check the preprint: 🔗 www.biorxiv.org/content/10.1...
Bottom line: dLEM shows we don't have to choose between mechanistic insight and predictive power.
By reformulating loop extrusion as a differentiable process, we get:
✅ Interpretability
✅ Scalability
✅ Perturbation prediction
✅ Integration with ML
10/10
By reformulating loop extrusion as a differentiable process, we get:
✅ Interpretability
✅ Scalability
✅ Perturbation prediction
✅ Integration with ML
10/10
October 21, 2025 at 2:04 PM
Bottom line: dLEM shows we don't have to choose between mechanistic insight and predictive power.
By reformulating loop extrusion as a differentiable process, we get:
✅ Interpretability
✅ Scalability
✅ Perturbation prediction
✅ Integration with ML
10/10
By reformulating loop extrusion as a differentiable process, we get:
✅ Interpretability
✅ Scalability
✅ Perturbation prediction
✅ Integration with ML
10/10
The parameter reduction compared to other deep learning models in deep dLEM is wild 🤯:
- 280× fewer than C.Origami
- 750× fewer than Orca
Yet we achieve competitive accuracy!
Why? Because we hard-coded the biophysics instead of making the model learn it.
9/10
- 280× fewer than C.Origami
- 750× fewer than Orca
Yet we achieve competitive accuracy!
Why? Because we hard-coded the biophysics instead of making the model learn it.
9/10
October 21, 2025 at 2:03 PM
The parameter reduction compared to other deep learning models in deep dLEM is wild 🤯:
- 280× fewer than C.Origami
- 750× fewer than Orca
Yet we achieve competitive accuracy!
Why? Because we hard-coded the biophysics instead of making the model learn it.
9/10
- 280× fewer than C.Origami
- 750× fewer than Orca
Yet we achieve competitive accuracy!
Why? Because we hard-coded the biophysics instead of making the model learn it.
9/10
We took dLEM to the next level: deep dLEM 🧠
We embedded dLEM as a biophysical layer in a neural network that learns from:
- DNA sequence
- Chromatin accessibility (DNase/ATAC)
Result: Sequence → structure prediction that stays mechanistically grounded
8/10
We embedded dLEM as a biophysical layer in a neural network that learns from:
- DNA sequence
- Chromatin accessibility (DNase/ATAC)
Result: Sequence → structure prediction that stays mechanistically grounded
8/10
October 21, 2025 at 2:01 PM
We took dLEM to the next level: deep dLEM 🧠
We embedded dLEM as a biophysical layer in a neural network that learns from:
- DNA sequence
- Chromatin accessibility (DNase/ATAC)
Result: Sequence → structure prediction that stays mechanistically grounded
8/10
We embedded dLEM as a biophysical layer in a neural network that learns from:
- DNA sequence
- Chromatin accessibility (DNase/ATAC)
Result: Sequence → structure prediction that stays mechanistically grounded
8/10
Example 2: CTCF depletion
→ We systematically reduced CTCF coefficients in our linear model
→ Optimal prediction at α=0.33, matching the 33% of CTCF peaks that remain after auxin treatment
The model reveals residual CTCF is still functionally active!
7/10
→ We systematically reduced CTCF coefficients in our linear model
→ Optimal prediction at α=0.33, matching the 33% of CTCF peaks that remain after auxin treatment
The model reveals residual CTCF is still functionally active!
7/10
October 21, 2025 at 2:01 PM
Example 2: CTCF depletion
→ We systematically reduced CTCF coefficients in our linear model
→ Optimal prediction at α=0.33, matching the 33% of CTCF peaks that remain after auxin treatment
The model reveals residual CTCF is still functionally active!
7/10
→ We systematically reduced CTCF coefficients in our linear model
→ Optimal prediction at α=0.33, matching the 33% of CTCF peaks that remain after auxin treatment
The model reveals residual CTCF is still functionally active!
7/10
Because dLEM parameters are physically interpretable, we can do something deep learning can't: predict perturbation effects!
Example: WAPL depletion
→ We modified just the detachment rate
→ Successfully predicted emergence of new loops & extended stripes
Example: WAPL depletion
→ We modified just the detachment rate
→ Successfully predicted emergence of new loops & extended stripes
October 21, 2025 at 2:00 PM
Because dLEM parameters are physically interpretable, we can do something deep learning can't: predict perturbation effects!
Example: WAPL depletion
→ We modified just the detachment rate
→ Successfully predicted emergence of new loops & extended stripes
Example: WAPL depletion
→ We modified just the detachment rate
→ Successfully predicted emergence of new loops & extended stripes
What do we learn from these velocity profiles?
😌 CTCF shows expected directional asymmetry (blocks L or R depending on orientation)
😲 BUT: Transcription machinery (H3K4me3, H3K36me3, H3K4me1,...) ALSO modulates cohesin dynamics!
CTCF isn't the whole story 👀
5/10
😌 CTCF shows expected directional asymmetry (blocks L or R depending on orientation)
😲 BUT: Transcription machinery (H3K4me3, H3K36me3, H3K4me1,...) ALSO modulates cohesin dynamics!
CTCF isn't the whole story 👀
5/10
October 21, 2025 at 1:58 PM
What do we learn from these velocity profiles?
😌 CTCF shows expected directional asymmetry (blocks L or R depending on orientation)
😲 BUT: Transcription machinery (H3K4me3, H3K36me3, H3K4me1,...) ALSO modulates cohesin dynamics!
CTCF isn't the whole story 👀
5/10
😌 CTCF shows expected directional asymmetry (blocks L or R depending on orientation)
😲 BUT: Transcription machinery (H3K4me3, H3K36me3, H3K4me1,...) ALSO modulates cohesin dynamics!
CTCF isn't the whole story 👀
5/10
dLEM acts as an encoder-decoder.
📥 ENCODER: Compress complex 2D Hi-C maps into simple 1D velocity profiles
📤 DECODER: Reconstruct contact maps from these profiles
It's dimensionality reduction, but with built-in biophysics!
4/10
📥 ENCODER: Compress complex 2D Hi-C maps into simple 1D velocity profiles
📤 DECODER: Reconstruct contact maps from these profiles
It's dimensionality reduction, but with built-in biophysics!
4/10
October 21, 2025 at 1:56 PM
dLEM acts as an encoder-decoder.
📥 ENCODER: Compress complex 2D Hi-C maps into simple 1D velocity profiles
📤 DECODER: Reconstruct contact maps from these profiles
It's dimensionality reduction, but with built-in biophysics!
4/10
📥 ENCODER: Compress complex 2D Hi-C maps into simple 1D velocity profiles
📤 DECODER: Reconstruct contact maps from these profiles
It's dimensionality reduction, but with built-in biophysics!
4/10
💡The key insight: While contact maps are 2D, loop extrusion is inherently 1D—cohesin walks along chromatin!
dLEM captures this with two velocity profiles:
→ L: leftward cohesin speed
→ R: rightward cohesin speed
Obstacles (like CTCF) = dips in velocity 📉
3/10
dLEM captures this with two velocity profiles:
→ L: leftward cohesin speed
→ R: rightward cohesin speed
Obstacles (like CTCF) = dips in velocity 📉
3/10
October 21, 2025 at 1:56 PM
💡The key insight: While contact maps are 2D, loop extrusion is inherently 1D—cohesin walks along chromatin!
dLEM captures this with two velocity profiles:
→ L: leftward cohesin speed
→ R: rightward cohesin speed
Obstacles (like CTCF) = dips in velocity 📉
3/10
dLEM captures this with two velocity profiles:
→ L: leftward cohesin speed
→ R: rightward cohesin speed
Obstacles (like CTCF) = dips in velocity 📉
3/10
The challenge: Current methods face an impossible tradeoff.
❌ Polymer simulations = mechanistic, but not scalable
❌ Deep learning = predictive, but black-box
❌ Summary metrics = compressed, but mechanism-agnostic
We needed something that's mechanistic, interpretable, AND scalable.
✅ Enter dLEM!
2/10
❌ Polymer simulations = mechanistic, but not scalable
❌ Deep learning = predictive, but black-box
❌ Summary metrics = compressed, but mechanism-agnostic
We needed something that's mechanistic, interpretable, AND scalable.
✅ Enter dLEM!
2/10
October 21, 2025 at 1:54 PM
The challenge: Current methods face an impossible tradeoff.
❌ Polymer simulations = mechanistic, but not scalable
❌ Deep learning = predictive, but black-box
❌ Summary metrics = compressed, but mechanism-agnostic
We needed something that's mechanistic, interpretable, AND scalable.
✅ Enter dLEM!
2/10
❌ Polymer simulations = mechanistic, but not scalable
❌ Deep learning = predictive, but black-box
❌ Summary metrics = compressed, but mechanism-agnostic
We needed something that's mechanistic, interpretable, AND scalable.
✅ Enter dLEM!
2/10
8/8 I had so much fun working on this paper with Ivo @mosaicgroup.bsky.social group at @mpicbg.bsky.social! Thanks for all the support, patience, and freedom to explore! # CompBio #Biophysics #PhDLife #Science #Research
December 20, 2024 at 1:05 PM
8/8 I had so much fun working on this paper with Ivo @mosaicgroup.bsky.social group at @mpicbg.bsky.social! Thanks for all the support, patience, and freedom to explore! # CompBio #Biophysics #PhDLife #Science #Research
7/8 There's more to explore, especially for 2D systems (like membranes), but this changes how we use these models to study cellular processes!
December 20, 2024 at 1:04 PM
7/8 There's more to explore, especially for 2D systems (like membranes), but this changes how we use these models to study cellular processes!
6/8 When these models "lose" reactions, they're actually capturing a real physical transition to the diffusion-limited regime. This opens new possibilities for studying complex molecular systems.
December 20, 2024 at 1:03 PM
6/8 When these models "lose" reactions, they're actually capturing a real physical transition to the diffusion-limited regime. This opens new possibilities for studying complex molecular systems.
5/8 On top of that, we find that different diffusion models (like GRDME vs traditional RDME) correspond to different molecular sizes on the same grid. So even with this interpretation of RDME, we can model systems with molecules of different sizes!
December 20, 2024 at 1:03 PM
5/8 On top of that, we find that different diffusion models (like GRDME vs traditional RDME) correspond to different molecular sizes on the same grid. So even with this interpretation of RDME, we can model systems with molecules of different sizes!
4/8 Our key finding is that grid size sets molecular size! Smaller boxes —> smaller molecules —> need to get closer to react—> fewer reactions. This perfectly matches the theory for diffusion-limited systems.
December 20, 2024 at 1:02 PM
4/8 Our key finding is that grid size sets molecular size! Smaller boxes —> smaller molecules —> need to get closer to react—> fewer reactions. This perfectly matches the theory for diffusion-limited systems.
3/8 But what if the grid size itself represents the reaction radius - the distance at which molecules can react? Then fewer reactions when boxes are smaller would be exactly what should happen physically! 💡
December 20, 2024 at 1:01 PM
3/8 But what if the grid size itself represents the reaction radius - the distance at which molecules can react? Then fewer reactions when boxes are smaller would be exactly what should happen physically! 💡
2/8 In grid-based models (RDME/GRDME), molecules jump between boxes and react when in the same box. As we make boxes smaller, reactions become less frequent - for years, this was seen as a flaw! 🤔
December 20, 2024 at 1:01 PM
2/8 In grid-based models (RDME/GRDME), molecules jump between boxes and react when in the same box. As we make boxes smaller, reactions become less frequent - for years, this was seen as a flaw! 🤔
7/7 This work led to some surprising insights about how these models actually represent physical reality, particularly the 'reaction loss' phenomenon... and that's a story of the just published paper! 😉 #CompBio #Biophysics #PhDLife #Science #Research
December 19, 2024 at 9:18 PM
7/7 This work led to some surprising insights about how these models actually represent physical reality, particularly the 'reaction loss' phenomenon... and that's a story of the just published paper! 😉 #CompBio #Biophysics #PhDLife #Science #Research
6/7 Here's where it gets interesting: RDME "loses" bimolecular reactions at very fine grid resolutions. Well, so does GRDME! We derived a rule of thumb to predict this limit - turns out GRDME's limit is higher by the same factor as its speedup.
December 19, 2024 at 9:18 PM
6/7 Here's where it gets interesting: RDME "loses" bimolecular reactions at very fine grid resolutions. Well, so does GRDME! We derived a rule of thumb to predict this limit - turns out GRDME's limit is higher by the same factor as its speedup.
5/7 Both RDME and GRDME show second-order convergence of the diffusion error with grid resolution. While GRDME has a slightly larger error, it runs up to 6x faster in 3D (or even more if you increase the smoothing length)! There's no free lunch in computing, but this is a pretty good deal!
December 19, 2024 at 9:18 PM
5/7 Both RDME and GRDME show second-order convergence of the diffusion error with grid resolution. While GRDME has a slightly larger error, it runs up to 6x faster in 3D (or even more if you increase the smoothing length)! There's no free lunch in computing, but this is a pretty good deal!
4/7 PSE naturally comes with a 'smoothing length' parameter (the spread of the Gaussian). By repurposing it for GRDME, we get a built-in way to tune the trade-off between speed and accuracy!
December 19, 2024 at 9:17 PM
4/7 PSE naturally comes with a 'smoothing length' parameter (the spread of the Gaussian). By repurposing it for GRDME, we get a built-in way to tune the trade-off between speed and accuracy!