Protein-DNA binding prediction and binder design from sequence, aligning protein and DNA language model embeddings instead of co-folding a complex.
No providers recorded yet. Browse all providers
Protein-DNA recognition underlies transcriptional regulation and every programmable genome-editing scaffold, but the strongest computational models of it are structural. Co-folding systems such as AlphaFold 3 and RoseTTAFold All-Atom can place a protein on a DNA duplex, yet they are far too slow for library-scale screening and are least reliable exactly where they are most needed: transcription factors, whose DNA-binding regions are often conformationally unstable in isolation.
DPAC — DNA-Protein binding Alignment via Contrastive learning — from the Chatterjee lab at Duke University, sidesteps structure entirely. It is a two-tower CLIP-style model that takes frozen embeddings from a protein language model and a DNA language model, projects both into a shared latent space through small MLP heads, and trains that projection with a symmetric InfoNCE loss over observed protein-DNA binding pairs. Cosine similarity in the aligned space becomes a binding score for any protein-DNA pair, so ranking a library reduces to a matrix multiplication.
Because the score is a differentiable function of sequence, it also doubles as a design objective. DPAC pairs the scorer with simulated annealing that proposes single amino acid substitutions sampled from a protein language model rather than uniformly over the 20 residues, biasing exploration toward naturalistic sequences while optimizing predicted affinity.
The protein tower uses ESM-2 650M with max-pooled token embeddings (1,280-dimensional); the DNA tower uses Nucleotide Transformer v2 500M multi-species with average-pooled embeddings (1,024-dimensional). Both encoders stay frozen. Each branch is a linear projection, GELU, a second projection to a 1,280-dimensional joint space, and layer normalization. The learnable InfoNCE temperature was swept from 0.01 to 0.3 and fixed at 0.085. Training data comes from BioLiP2: 23,817 entries filtered to remove duplicates, DNA motifs shorter than 3-mers, and proteins over 500 residues, then clustered separately for protein and DNA at 0.7 identity with MMseqs2. This yields a strict low-identity set of 99 complexes (54 unique DNA motifs, 58 protein sequences) plus 1,000 randomly sampled high-identity pairs, for 11,419 training entries total. The model trains for 1,000 epochs with AdamW on a single A100.
On the low-identity set DPAC reaches AUC 0.591, producing substantially higher enrichment factors than RoseTTAFold All-Atom and RFNA while matching them on other retrieval metrics. AUC shows no systematic decay as pairwise sequence identity to the training set drops. In a design test on the Zif268 zinc finger (PDB 1AAY), sequences were corrupted at 10%, 30%, and 50% of positions and then re-optimized; at the 30% and 50% corruption levels, which disrupt the third alpha helix contacting the DNA minor groove, annealing recovered roughly 20% of the lost binding confidence as measured by AlphaFold 3 ipTM. Against LigandMPNN redesign of RoseTTAFold-folded complexes, DPAC-guided optimization achieved comparable binding scores, particularly under a 10% mutation budget.
The immediate use is virtual screening: given a target DNA motif, rank a large protein library for likely binders, or given a candidate protein, rank motifs it may recognize. That covers transcription factor specificity profiling and the search for scaffolds for zinc-finger nucleases and other programmable DNA-targeting proteins. The annealing module supports the reverse task — starting from a scaffold and optimizing its sequence toward a chosen target site — which is the core loop in engineering sequence-specific DNA binders for genome editing and synthetic transcriptional control.
DPAC is a preprint awaiting peer review, and its validation is entirely in silico: binding recovery is scored by AlphaFold 3 confidence metrics and affinity correlation by regression against deposited dissociation constants, with no wet-lab characterization of designed sequences. The training set is also small — 99 low-identity complexes anchor the strict evaluation — and the reported AUC of 0.591 reflects a genuinely hard retrieval task rather than a solved one. The contribution is the demonstration that a lightweight contrastive projection over frozen language model embeddings can outperform structure-based co-folding on protein-DNA retrieval while running roughly five orders of magnitude faster. Code and the trained alignment checkpoint are published in the repository, whose README states an MIT license.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.