Six Models, One Question: Can Deep Learning Predict Gene Regulation?

Author: CAIRN Institute

Published: October 25, 2026

Read time: 10–12 minutes

#Genomics #Bioinformatics #GeneRegulation #DeepLearning #StructuralBiology #OpenScience

Quick Summary

We tested six deep learning approaches for predicting whether a transcription factor binds its target gene’s promoter DNA, validated at scale with n = 50+50 controls per condition across multiple species.

Results:

• JASPAR motif scanning is the strongest plant binding discriminator: arabidopsis AUC 0.795 (59.4% hit rate in real targets vs 0.2% in random), petunia AUC 0.715
• Enformer shows weak overall signal at scale (AUC 0.518, mean separation +0.84) — TF-specific and noisy, useful only for select human TFs with strong ChIP-seq training data
• Boltz-2 confirmed uninformative across both arabidopsis (AUC 0.420) and human (AUC 0.520)
• Evo2 (forward pass, SAE features, and linear probe) encodes no binding information in any linearly extractable form
• DeepBind model files are no longer accessible

For the GRN Atlas, this means: JASPAR for plant edges, Enformer selectively for human edges, and perturbation data — not sequence models — for regulatory sign.

Repository: github.com/cairninstitute/grn-atlas

Boundary statement: “All calibration AUCs measure model discrimination between atlas-curated positive controls and random negative pairings. They do not measure in vivo binding probability. Motif presence, chromatin accessibility, and cellular context all intervene between a sequence match and a regulatory outcome.”

The Problem: 162,000 Edges with No Physical Evidence

The GRN Atlas is a cross-species gene regulatory network built from eight organisms — from arabidopsis to human. It holds 162,000 candidate regulatory edges: predicted relationships where a transcription factor (TF) may regulate a target gene. These candidates come from expression data, perturbation experiments, and curated databases. But for the vast majority of them, we have no direct evidence that the TF protein physically binds the target gene’s promoter DNA.

Physical binding is the gold standard. A TF can only regulate a gene if it can reach the promoter (or an enhancer) and interact with the DNA. Techniques like ChIP-seq and DAP-seq measure this directly, but they are expensive, species-specific, and cover only a fraction of known TFs. For plant species especially, binding data is sparse.

We wanted a computational shortcut: given a TF and a target gene, can a deep learning model predict whether binding is plausible? We tested every credible approach we could find — protein-DNA structure prediction, DNA language models, chromatin models, sparse autoencoders, and classical motif scanning. The results were humbling.

The Contenders

Each approach attacks the binding question from a different angle. We calibrated all of them the same way: score known positive controls (TF-target pairs with experimentally confirmed binding) against negative controls (random TF-gene pairings with no known relationship), and measure whether the model separates the two.

Of these seven, six returned results. One — Enformer — showed real binding signal. The rest did not.

Enformer: The One That Worked

Enformer is Google DeepMind’s sequence-to-function model. It takes 393,216 bp of DNA context and predicts 5,313 genomic tracks — including ChIP-seq signal for hundreds of human transcription factors. The idea is simple: if Enformer predicts high ChIP-seq signal for TF X at target gene Y’s promoter, that is evidence of binding.

We ran Enformer on our RTX 5070 Ti in two rounds. A pilot with 20+20 controls produced the per-TF results shown below. A large-scale validation with 50+50 unique-target controls (73 scored after Ensembl REST errors) told a different story: overall AUC 0.518, mean separation +0.84. The pilot’s +2.61 separation was inflated by a duplicate-target bug and small-sample effects. The TF-specific pattern, however, persisted.

The pattern is revealing. TFs like MYC and E2F1, which bind specific promoter motifs in a context-dependent way, show strong separation: Enformer correctly predicts high signal at their real targets and low signal at random genes. But ubiquitous binders like SP1 and YY1 — which bind broadly across the genome — show no separation or even reversed signal, because they score high everywhere.

In the large-scale validation, most TFs had only one or two edges per class — too few for individual AUC estimates. The TFs with enough data showed RELB (AUC 0.833, 2+3 edges) and SP1 (AUC 0.667, 3+2 edges) with positive signal, while others like ATF7 (−5.99 separation) were strongly reversed. The overall AUC of 0.518 confirms that Enformer’s signal is real but too noisy and TF-dependent for use as a global binding discriminator.

Enformer’s limitation for the GRN Atlas is fundamental: it is human-only. Its 5,313 tracks come from human ChIP-seq, ATAC-seq, and DNase experiments. There are no plant TF tracks. Until someone trains an Enformer-equivalent on plant chromatin data, this approach cannot help with our arabidopsis, rice, tomato, or potato edges — which make up the majority of the atlas.

For the human edges it does cover, Enformer requires TF-specific thresholds rather than a global cutoff. A binding score that is meaningful for MYC is meaningless for SP1. We use Enformer selectively: for human TFs with strong ChIP-seq training data, checked against a curated track list.

The Evo2 Family: Three Attempts, Zero Binding Signal

Evo2 is a 7-billion-parameter (and 40-billion-parameter) DNA language model from the Arc Institute, trained on 128,000 genomes spanning the tree of life. We tried three different ways to extract binding information from it.

Approach 1: Forward pass logits. We fed 200 bp promoter sequences into Evo2-40B via NVIDIA’s NIM API and compared the mean logit scores for known binding targets against random gene pairings. The mean separation was −0.035 — indistinguishable from zero. Positive and negative controls clustered together.

Approach 2: Sparse autoencoder features. Goodfire trained a sparse autoencoder (SAE) on Evo2-7B’s layer-26 hidden states, decomposing each sequence into 32,768 interpretable features. We ran L1 logistic regression to classify binding vs non-binding. AUC: 0.50. No signal whatsoever.

Approach 3: Linear probe on raw hidden states. We extracted 8,192-dimensional mean-pooled hidden states from Evo2-40B’s layer 26 for 152 positive and 250 negative arabidopsis edges, and trained an L1 logistic regression probe. AUC: 0.50 across all regularization strengths. At weak regularization, every coefficient was zero — the probe found nothing to grab onto.

Three approaches. Three ways of asking the same model the same question. The answer was consistent: Evo2 does not encode TF-DNA binding specificity in a way that any linear method can extract.

But the SAE told us something unexpected. While it cannot predict binding, it is a powerful genomic feature extractor. When we tested it on region classification instead of binding, it separated promoter DNA from coding DNA with AUC 0.889 in arabidopsis, and distinguished plant from animal sequences with AUC 1.000. It reads genomes phylogenetically — kingdom, clade, species — with a discrimination gradient that tracks real evolutionary distance. We wrote about those findings in detail in our companion post.

The lesson: a DNA language model trained on raw genomes learns genome structure and evolutionary signatures. It does not learn the specific, context-dependent question of whether protein X binds DNA sequence Y. That question requires information about the protein, not just the DNA.

Boltz-2: When Structure Prediction Meets Its Limits

The intuition behind Boltz-2 was compelling. AlphaFold revolutionized protein structure prediction; Boltz-2 extends that to protein-DNA complexes. If we predict the 3D structure of a TF bound to a promoter, the model’s confidence score should be higher for real binding partners than for random pairings.

We tested this with NVIDIA’s Boltz-2 NIM API, predicting structures for arabidopsis TF-promoter pairs at two DNA window sizes (200 bp and 500 bp) and with an MSA-assisted variant.

The confidence scores clustered between 0.28 and 0.44 regardless of whether the TF-DNA pair was a known interaction or a random pairing. Extending the DNA window to 500 bp made no difference. Adding multiple sequence alignment made no difference either — and ColabFold’s MSA search returned zero results for every plant TF we queried, so the MSA variant was identical to the plain version.

Why does structure prediction fail here? Boltz-2 predicts whether a protein can fold against a DNA sequence in a physically plausible way. Most TFs have DNA-binding domains that can dock onto most B-form DNA — the structural geometry is generic. What makes binding specific is the sequence of the motif, the chromatin context, and the cofactor environment. A confidence score of 0.33 means the model is uncertain about everything, which is the honest answer: the structural prediction problem is underdetermined with just a protein and a short DNA sequence.

We confirmed this at scale: 50+50 controls across both arabidopsis (AUC 0.420) and human (AUC 0.520) yielded the same null result. Boltz-2’s confidence scores do not discriminate known binding partners from random pairings in either plant or animal genomes.

JASPAR: The Old Approach That Still Wins

JASPAR is an open-access database of transcription factor binding profiles — position weight matrices (PWMs) derived from decades of experimental binding assays. The approach is straightforward: scan a promoter sequence for subsequences that match a TF’s known binding motif, and score each match by how well it fits the matrix.

JASPAR has none of the ambition of the deep learning approaches. It does not predict 3D structure, model chromatin accessibility, or learn latent representations of genome evolution. It looks for a short pattern in a string. And for plant species, it is the best tool we have.

The reason is coverage. JASPAR 2024 contains binding profiles for hundreds of plant TFs, derived from DAP-seq, ChIP-seq, and SELEX experiments. When a TF’s binding motif is known, scanning for it in a target promoter is a direct, interpretable test. The result is not a confidence score from a black-box model — it is the presence or absence of a specific recognized sequence, at a specific position, with a quantifiable match score.

Motif scanning has clear limitations. It cannot predict binding for TFs whose motifs are not in the database. It says nothing about whether the motif is accessible in the chromatin context of a given cell type. It cannot distinguish a motif that is actively bound from one that happens to occur in the sequence by chance — short motifs (6–10 bp) appear frequently in any genome. And it cannot predict regulatory sign: the same motif bound by the same TF can activate or repress depending on cofactors and chromatin state.

But within those limits, motif scanning does something none of the deep learning models could do for plants: it connects a specific TF to a specific DNA sequence through a known, experimentally validated binding preference. For the GRN Atlas, that directness matters more than sophistication.

We ran a formal calibration against the GRN Atlas: for each TF with a JASPAR motif, we tested whether motif hits appear more often in known regulatory targets than in random genes.

The arabidopsis hit rate separation — 59.4% in real targets versus 0.2% in random genes — is dramatic. Petunia, where most motifs come from cross-species transfer, still achieves AUC 0.715. Tomato’s zero-hit result reflects incomplete motif scanning coverage for the SL4.0 assembly, not a failure of the approach.

The Scorecard

The honest summary: predicting TF-DNA binding from sequence alone is a harder problem than current deep learning models can solve, except in the narrow case where a model was trained on direct experimental binding data for that specific TF (Enformer’s ChIP-seq tracks) or where the binding motif is already known (JASPAR). General-purpose DNA language models like Evo2 learn rich representations of genome structure and evolution, but those representations do not encode the specifics of which protein binds which sequence.

For the GRN Atlas, this means we use a layered strategy: JASPAR motif scanning for plant edges, Enformer for human edges with good TF coverage, and perturbation data — not sequence models — for regulatory sign. The deep learning revolution has not yet arrived for cross-species gene regulation prediction. But calibrating where each tool works and where it fails is itself a contribution — it saves the next group from repeating the same experiments.

What This Shows — and Does Not Show

Demonstrates:

• JASPAR motif scanning discriminates known plant TF-target pairs from random pairings (AUC 0.795 in arabidopsis, 0.715 in petunia)
• Enformer provides weak but real TF-specific signal for human edges with strong ChIP-seq training data (overall AUC 0.518)
• Current protein-DNA structure prediction (Boltz-2) and DNA language models (Evo2) cannot solve general binding prediction
• No tested sequence-based approach predicts regulatory sign (activation vs repression)

Does not demonstrate:

• That JASPAR motif presence implies actual binding in vivo — a motif match is necessary but not sufficient evidence
• That Enformer’s AUC 0.518 is sufficient for unsupervised binding classification
• That these approaches have been tested on all relevant TF families or genomic contexts
• That better-performing deep learning approaches do not exist — only that the most prominent current models were tested

Physical binding is one link in the chain from transcription factor to gene regulation. These calibrations measure the first link. The full chain remains a harder problem.

References

• Avsec et al. (2021), Effective gene expression prediction from sequence by integrating long-range interactions — doi:10.1038/s41592-021-01252-x
• Nguyen et al. (2024), Sequence modeling and design from molecular to genome scale with Evo — Arc Institute; arcinstitute.org/tools/evo
• Goodfire (2024), Sparse Autoencoders for Evo2 — goodfire.ai
• Abramson et al. (2024), Accurate structure prediction of biomolecular interactions with AlphaFold 3 — Boltz-2 via NVIDIA NIM
• Castro-Mondragon et al. (2022), JASPAR 2024: open-access database of transcription factor binding profiles — jaspar.elixir.no