GeneZip: Region-Aware Compression for Long Context DNA Modeling
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "GeneZip: Region-Aware Compression for Long Context DNA Modeling".
Marcus: Here's the gist of the paper. Long-context DNA models are limited by token-mixing cost and by how compression allocates representational budget across the genome GeneZip, a region-aware DNA compression framework that combines H-Net-style dynamic routing with a Region-Aware Ratio (RAR) objective and bounded routing GeneZip provides three main benefits First, it is effective: GeneZip variants achieve the best validation PPL among encoder-based compressors, with GeneZip-70M already operating at 137.6 BPT, and across four reproducible DNALongBench tasks—contact map prediction, eQTL prediction, enhancer–target gene prediction, and transcription-initiation signal prediction—GeneZip obtains the best average rank among compared sequence models Second, it is redundancy-aware: a post-hoc RepeatMasker/TRF analysis shows that, without repeat supervision, GeneZip assigns higher local BPT to TE-derived interspersed repeats and tandem repeats, two major classes of repetitive DNA sequence redundancy Third, it is efficient: by reducing the effective token-mixing length, GeneZip enables longer-context and larger-capacity pretraining, including 128K-context and 636M-parameter variants on a single A100 80GB GPU, and fine-tunes the eQTL task 50.4× faster than JanusDNA (50 vs.
Ines: 2520 minutes) Pruning redundant reasoning tokens GeneZip addresses the bottleneck where existing approaches operate close to base-pair resolution, apply fixed downsampling, or learn content-dependent chunks without an explicit genomic budget Motivated by this imbalance, we propose GeneZip, a region-adaptive compression module that learns non-uniform token allocation over ultra-long DNA Specifically, its Region-Aware (RAR) objective trains the router toward region-wise base-pairs-per-token (BPT) targets derived from static gene-structure annotations, and bounded routing enforces a global token budget for stable long-context training The region annotations are used only during compression training, so GeneZip compresses unseen raw DNA at inference time without region labels How the training signal is built The compression process involves three stages: encoding, token mixing, and decoding The encoder instantiates Encode(·) with a multi-stage hierarchical genomic compression mechanism inspired by H-Net The Dynamic Routing module induces adaptive boundaries by measuring the representation shift between adjacent positions in hˆ(s), computing boundary probabilities p(s) via cosine dissimilarity between adjacent projected representations Hierarchical dynamic chunking partitions the sequence into variable-length chunks, and bounded routing enforces a global token budget for stable training Region supervision via gene structure involves assigning a genomic region label r(t) to each base position t, which is used only to define supervision targets (RAR) The RAR loss enforces region-specific compression by aligning regionwise keep rates across S routing stages Where the gains come from GeneZip achieves effective compression and prediction, reaching 137.6 BPT with only a +0.31 perplexity increase, and obtains the best average rank across four DNALongBench tasks GeneZip is redundancy-aware because post-hoc RepeatMasker and TRF analyses show that it assigns stronger local compression to repetitive DNA sequence redundancy, including both TE-derived interspersed repeats and tandem-repeat (TR) interiors Furthermore, GeneZip is efficient, enabling 128K-context and 636M-parameter pretraining on a single A100 80GB GPU and fine-tunes the eQTL task 50.4× faster than JanusDNA Downstream performance and utility GeneZip demonstrates strong transferability, as its learned routing policy transfers to held-out human chromosomes and an unseen species (mouse) without inference-time annotations In the DNALongBench evaluation, GeneZip achieves the best average rank across all compared sequence models For example, in enhancer–target gene prediction (ETGP), GeneZip-70M attains the best AUPRC (0.446), outperforming JanusDNA by +0.008 and the Expert Model by +0.039 This supports the use of region-adaptive compression for cross-element regulatory association when the allocation prior matches the task We propose GeneZip, a region-aware DNA compressor that combines bounded dynamic routing with a region-aware compression objective to reallocate token budget across genomic compartments GeneZip achieves 137.6× compression with only a 0.31 perplexity increase, preserves downstream utility across long-context genomic tasks, and obtains the best aggregate rank across four reproducible DNALongBench tasks among all compared sequence models A post-hoc repeat analysis shows redundancy-aware behavior: without TE or TR supervision, GeneZip assigns higher local compression to repetitive DNA sequence redundancy than generic H-Net routing, with robust cross-species signals for LINE and tandem repeats and a particularly sharp human SINE effect GeneZip also supports 128K-context and 636M-parameter training within the reported single-A100 setup and substantially reduces wall-clock cost for downstream long-context fine-tuning Implementation Details The implementation involves a two-stage hierarchical dynamic chunking encoder with S = 2, where each stage is implemented with two Mamba2 layers The token-mixing backbone consists of 15 Mamba2 layers For bounded routing, we define a floor and ceiling around K(s) to stabilize training by projecting routing decisions onto a per-stage token-budget interval The Region-Aware Ratio (RAR) loss is defined by Eq.
Marcus: (8), which couples the realized keep rate F(s)r after projection with its expectation G(s)r before projection to match τ⋆(s)r Ablation Studies The Region-Ratio ablation study ablates the central control variable in GeneZip: the region multiplier µr in the Region-Aware Ratio (RAR) objective The analysis shows that task-aware region allocation remains competitive under a two-dimensional long-range task and also helps beyond local promoter or variant tasks For instance, TISP favors high resolution around promoter/TSS-proximal sequence, consistent with base-pair-resolution transcription initiation profiling Future Work Extending the training setup to broader multi-species resources with harmonized region annotations is an important direction for future work Incorporating richer function-oriented supervision beyond annotation labels, for example sequence-to-function objectives inspired by models such as Enformer and NT-family genomic predictors, is a promising next step GeneZip builds on the underlying H-Net routing architecture
10: , and its contribution is to introduce a region-aware compression objective and bounded budget control on top of this dynamic routing framework Comparison to Tokenization Methods GeneZip targets high-ratio, controllable length reduction for long-context inputs, distinguishing itself from BPE tokenization which does not enforce a per-sample BPT target or a region-wise allocation constraint GeneZip reaches about 170 bp/token on these windows, enabling the downstream speed and scale advantage compared to methods like MxDNA Data Details The pre-training corpora are constructed from the GRCh38 primary assembly together with the GENCODE v49 basic gene annotation The sampling procedure removes all genomic intervals used by DNALongBench validation/test sets from the sampling pool for train/valid and fully holds out the DNALongBench transcription-initiation held-out chromosomes chr8, chr9, and chr10 during sampling Comparison to Tokenization Methods GeneZip studies a different problem from tokenization.
Ines: Most DNA tokenization methods are designed to define useful sequence units or vocabularies GeneZip targets high-ratio, controllable length reduction for long-context inputs, distinguishing itself from BPE tokenization which does not enforce a per-sample BPT target, a bounded routing budget, or a region-wise allocation constraint such as Nbr ≈ N⋆r GeneZip reaches about 170 bp/token on these windows, enabling the downstream speed and scale advantage compared to methods like MxDNA Limitations The current GeneZip pre-training is derived on human genomic data The supervision is centered on static region annotations and the Region-Aware Ratio objective, and a promising next step is to incorporate richer function-oriented supervision beyond annotation labels The current GeneZip pre-training is derived on human genomic data Limitations The current GeneZip pre-training is derived on human genomic data The supervision is centered on static region annotations and the Region-Aware Ratio objective, and a promising next step is to incorporate richer function-oriented supervision beyond annotation labels Comparison to Tokenization Methods GeneZip studies a different problem from tokenization.
Paper summary: Ines: So we're looking at this paper called "GeneZip: Region-Aware Compression for Long Context DNA Modeling". Basically, the authors are tackling the problem that long context DNA models get bogged down by token mixing costs and how they decide to spend their representational budget across a whole genome.
Marcus: Right. They propose GeneZip as this region-aware compression framework that mixes H-Net style dynamic routing with a Region-Aware Ratio objective and bounded routing to solve that allocation problem.
Yuki: What's the big deal here for us? If we can get better control over where the model spends its tokens, does it help us understand sequence structure more effectively?
Ines: Well, they claim this framework gives three main benefits: first, it's effective because their GeneZip variants achieve the best validation perplexity among encoder-based compressors. They show that GeneZip at seventy million parameters already operates at one hundred thirty-seven point six BPT and only sees a small increase of zero point three one in perplexity >
Marcus: And they say this is across four reproducible DNALongBench tasks, which are contact map prediction, eQTL prediction, enhancer target gene prediction, and transcription initiation signal prediction >
Yuki: That’s good to hear. But what about the other things we care about? Are there any limitations you should be aware of right off the bat?
Ines: The authors point out that GeneZip is redundancy-aware because a post hoc RepeatMasker and TRF analysis shows that without repeat supervision, it assigns higher local BPT to TE derived interspersed repeats and tandem repeats, which are major classes of repetitive DNA sequence redundancy >
Marcus: That's interesting. It means the compression isn't just happening randomly; it's specifically targeting these highly redundant areas >
Yuki: So they’re doing something more intelligent than generic routing when it comes to repetitive elements. What does this mean for how we look at population genetics or species evolution?
Ines: It suggests that the model is learning where the biological complexity—the high redundancy—is concentrated, and it allocates more tokens there without us needing to explicitly tell it where those repeats are during inference >
Paper summary: Marcus: And on the efficiency side, they say by reducing the effective token mixing length, GeneZip enables longer context and larger capacity pretraining. They're talking about one hundred twenty-eight thousand context and six hundred thirty six million parameter variants running on a single A100 eighty gigabyte GPU >
Yuki: That sounds like a significant computational saving for training models on huge genomic datasets. Does this efficiency translate to better generalization when we move to unseen data, like held-out human chromosomes or even an unseen mouse species?
Ines: They do. The paper states that GeneZip's region-aware allocation generalizes from training sequences to held out human chromosomes and an unseen species like mouse without needing inference time annotations >
Marcus: And when we look at downstream performance, they found it achieves the best average rank across all compared sequence models on the DNALongBench tasks >
Yuki: If it performs better than other models on those specific prediction tasks, what does that imply about the underlying biological signal? Is it just better compression or is it actually capturing a more useful representation of gene regulation?
Ines: It supports the idea that using region-adaptive compression can work well for cross element regulatory association when the allocation prior matches the specific task you are trying to perform >
Marcus: They also showed that in enhancer target gene prediction, GeneZip at seventy million parameters gets an AUPRC of zero point four four six, which is better than JanusDNA by zero point zero eight and better than the Expert Model by zero point zero three nine >
Yuki: That comparison against other models shows it's competitive in a very specific biological prediction task. Does this mean we can trust its predictions for transcription initiation signals when we use it?
Ines: It suggests that the region-adaptive compression is aligned with the task, so the allocation prior matches what you need for those predictions >
Paper summary: Marcus: But I do want to bring up one thing: their method relies on static gene structure annotations during compression training. They say these annotations are only used for defining supervision targets, and GeneZip compresses unseen raw DNA at inference time without region labels >
Yuki: So the framework itself is robust enough to handle new sequences, even if we don't have those specific annotation labels when we use it >
Ines: Exactly. That’s a key distinction from other methods that might require region-specific training data for every new sequence you want to process >
Marcus: And they also have some interesting details about the training setup, like using two stage hierarchical dynamic chunking with S equal to two and a token mixing backbone of fifteen Mamba two layers >
Yuki: It sounds like a very specific engineering detail. Does that setup really capture the complexity of ultra long DNA sequences?
Ines: The region-aware ratio loss is what guides the budget via region targets, aligning the realized keep rate after projection with its expectation before projection to match those targets >
Marcus: So, in summary, GeneZip introduces a region-aware compression module that combines dynamic routing and a ratio objective to reallocate token budget across genomic compartments >
Yuki: It achieves high compression at one hundred thirty seven point six BPT while keeping the perplexity increase small, and it maintains strong utility across long context tasks and shows redundancy awareness in repetitive regions >
Ines: And for the end of it, it provides a way to get one hundred twenty-eight thousand context training on a single A100 eighty gigabyte GPU and speeds up fine tuning of eQTL prediction by fifty point four times compared to JanusDNA >
Marcus: It seems like they've managed to solve the problem of controlling token budget allocation in long context models by tying it directly to known biological structures like gene annotations, which is a smart way to add control >
Yuki: It opens up possibilities for using these large models on much longer genomic sequences, which could help us look at regulatory interactions across entire chromosomes with more fidelity >
Conclusion: Ines: So, to wrap up this, GeneZip is about taking long context DNA models and giving them smarter ways to compress their information by looking at where things are on the genome >
Marcus: Right. It’s about using a region-aware compression framework that mixes dynamic routing with an objective designed to allocate tokens better across the whole sequence >
Ines: The authors say this method achieves high compression, around one hundred thirty seven point six BPT, while keeping the perplexity increase pretty small >
Marcus: That’s a solid number. And they show it works across four different long context tasks, which is important because those tasks actually test how well the model understands things like enhancer target genes and transcription initiation signals >
Yuki: So, what does this mean for us in population genetics? If the compression is better at capturing regulatory associations in these long sequences, does that mean we can pull out more meaningful biological insights from vast genomic regions >
Ines: It means the model isn't just throwing tokens randomly around; it’s learning to concentrate its budget where it matters for those specific predictions >
Marcus: They also found this method is redundant awareness. They showed that without special supervision for repeats, GeneZip naturally compresses more heavily in areas with lots of repetitive DNA, like those transposable elements >
Yuki: That's interesting. It suggests the model is picking up on inherent biological structure even when we don't explicitly label every repeat region >
Ines: And they also showed that this framework helps the model generalize its learning to new species and chromosomes without needing specific annotations for those areas during inference >
Marcus: So, it seems like GeneZip is making a lot of headway in giving these models more reliable tools for working with truly massive genomic data sets >
Yuki: It definitely sounds like a step toward using these large models on much longer sequences, which could help us look at regulatory interactions across entire chromosomes with more fidelity >
Jianan Zhao, Xixian Liu, Zhihao Zhan, Xinyu Yuan, Hongyu Guo, Jian Tang
Mila - Québec AI Institute University de Montréal · National Research Council of Canada University of Ottawa Department of Human Embryology and Cell Biology HEC Montreal
q-bio.GN, cs.AI, cs.LG
Submitted: 2026-02-19
Updated: 2026-11-09
Comments: NeurIPS 2026. Project page: https://andyjzhao.github.io/GeneZip/
Code: https://github.com/AndyJZhao/GeneZip
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: Long-context DNA models are limited by token-mixing cost and by how compression allocates representational budget across the genome GeneZip…
Key concepts
- Region-Aware Ratio (RAR) objective
- This loss function trains the compression router to meet specific base-pairs-per-token targets for different genomic regions. It uses static gene structure annotations to guide the router, ensuring that tokens are allocated more densely or sparsely based on whether they fall within a promoter region or a repetitive sequence.
- Bounded Routing
- This mechanism enforces a global token budget across the entire DNA sequence during training. By setting floor and ceiling limits around routing decisions, it stabilizes long-context training, preventing the model from allocating an unbounded number of tokens to any single section.
- Dynamic Routing (H-Net inspired)
- The encoder uses a multi-stage hierarchical mechanism inspired by H-Net to determine token boundaries adaptively. It measures representation shifts between adjacent positions to compute boundary probabilities, allowing the model to create variable-length chunks tailored to local sequence features.
Terminology
Summary
Long-context DNA models are limited by token-mixing cost and by how compression allocates representational budget across the genome GeneZip, a region-aware DNA compression framework that combines H-Net-style dynamic routing with a Region-Aware Ratio (RAR) objective and bounded routing GeneZip provides three main benefits First, it is effective: GeneZip variants achieve the best validation PPL among encoder-based compressors, with GeneZip-70M already operating at 137. p. 06 BPT, and across four reproducible DNALongBench tasks—contact map prediction, eQTL prediction, enhancer–target gene prediction, and transcription-initiation signal prediction—GeneZip obtains the best average rank among compared sequence models Second, it is redundancy-aware: a post-hoc RepeatMasker/TRF analysis shows that, without repeat supervision, GeneZip assigns higher local BPT to TE-derived interspersed repeats and tandem repeats, two major classes of repetitive DNA sequence redundancy Third, it is efficient: by reducing the effective token-mixing length, GeneZip enables longer-context and larger-capacity pretraining, including 128K-context and 636M-parameter variants on a single A100 80GB GPU, and fine-tunes the eQTL task 50. p. 04× faster than JanusDNA (50 vs. p. 0 2520 minutes)
Pruning redundant reasoning tokens
GeneZip addresses the bottleneck where existing approaches operate close to base-pair resolution, apply fixed downsampling, or learn content-dependent chunks without an explicit genomic budget Motivated by this imbalance, we propose GeneZip, a region-adaptive compression module that learns non-uniform token allocation over ultra-long DNA Specifically, its Region-Aware (RAR) objective trains the router toward region-wise base-pairs-per-token (BPT) targets derived from static gene-structure annotations, and bounded routing enforces a global token budget for stable long-context training The region annotations are used only during compression training, so GeneZip compresses unseen raw DNA at inference time without region labels
How the training signal is built
The compression process involves three stages: encoding, token mixing, and decoding The encoder instantiates Encode(·) with a multi-stage hierarchical genomic compression mechanism inspired by H-Net The Dynamic Routing module induces adaptive boundaries by measuring the representation shift between adjacent positions in hˆ(s), computing boundary probabilities p(s) via cosine dissimilarity between adjacent projected representations Hierarchical dynamic chunking partitions the sequence into variable-length chunks, and bounded routing enforces a global token budget for stable training Region supervision via gene structure involves assigning a genomic region label r(t) to each base position t, which is used only to define supervision targets (RAR) The RAR loss enforces region-specific compression by aligning regionwise keep rates across S routing stages
Where the gains come from
GeneZip achieves effective compression and prediction, reaching 137.6 BPT with only a +0. p. 031 perplexity increase, and obtains the best average rank across four DNALongBench tasks GeneZip is redundancy-aware because post-hoc RepeatMasker and TRF analyses show that it assigns stronger local compression to repetitive DNA sequence redundancy, including both TE-derived interspersed repeats and tandem-repeat (TR) interiors Furthermore, GeneZip is efficient, enabling 128K-context and 636M-parameter pretraining on a single A100 80GB GPU and fine-tunes the eQTL task 50. p. 14× faster than JanusDNA
Downstream performance and utility
GeneZip demonstrates strong transferability, as its learned routing policy transfers to held-out human chromosomes and an unseen species (mouse) without inference-time annotations In the DNALongBench evaluation, GeneZip achieves the best average rank across all compared sequence models For example, in enhancer–target gene prediction (ETGP), GeneZip-70M attains the best AUPRC (0.446), outperforming JanusDNA by +0. p. 0008 and the Expert Model by +0. p. 0039 This supports the use of region-adaptive compression for cross-element regulatory association when the allocation prior matches the task
Conclusion
We propose GeneZip, a region-aware DNA compressor that combines bounded dynamic routing with a region-aware compression objective to reallocate token budget across genomic compartments GeneZip achieves 137.6× compression with only a 0. p. 031 perplexity increase, preserves downstream utility across long-context genomic tasks, and obtains the best aggregate rank across four reproducible DNALongBench tasks among all compared sequence models A post-hoc repeat analysis shows redundancy-aware behavior: without TE or TR supervision, GeneZip assigns higher local compression to repetitive DNA sequence redundancy than generic H-Net routing, with robust cross-species signals for LINE and tandem repeats and a particularly sharp human SINE effect GeneZip also supports 128K-context and 636M-parameter training within the reported single-A100 setup and substantially reduces wall-clock cost for downstream long-context fine-tuning
Implementation Details
The implementation involves a two-stage hierarchical dynamic chunking encoder with S = 2, where each stage is implemented with two Mamba2 layers The token-mixing backbone consists of 15 Mamba2 layers For bounded routing, we define a floor and ceiling around K(s) to stabilize training by projecting routing decisions onto a per-stage token-budget interval The Region-Aware Ratio (RAR) loss is defined by Eq. (8), which couples the realized keep rate F(s)r after projection with its expectation G(s)r before projection to match τ⋆(s)r
Ablation Studies
The Region-Ratio ablation study ablates the central control variable in GeneZip: the region multiplier µr in the Region-Aware Ratio (RAR) objective The analysis shows that task-aware region allocation remains competitive under a two-dimensional long-range task and also helps beyond local promoter or variant tasks For instance, TISP favors high resolution around promoter/TSS-proximal sequence, consistent with base-pair-resolution transcription initiation profiling
Future Work
Extending the training setup to broader multi-species resources with harmonized region annotations is an important direction for future work Incorporating richer function-oriented supervision beyond annotation labels, for example sequence-to-function objectives inspired by models such as Enformer and NT-family genomic predictors, is a promising next step GeneZip builds on the underlying H-Net routing architecture [10], and its contribution is to introduce a region-aware compression objective and bounded budget control on top of this dynamic routing framework
Comparison to Tokenization Methods
GeneZip targets high-ratio, controllable length reduction for long-context inputs, distinguishing itself from BPE tokenization which does not enforce a per-sample BPT target or a region-wise allocation constraint GeneZip reaches about 170 bp/token on these windows, enabling the downstream speed and scale advantage compared to methods like MxDNA
Data Details
The pre-training corpora are constructed from the GRCh38 primary assembly together with the GENCODE v49 basic gene annotation The sampling procedure removes all genomic intervals used by DNALongBench validation/test sets from the sampling pool for train/valid and fully holds out the DNALongBench transcription-initiation held-out chromosomes chr8, chr9, and chr10 during sampling
GeneZip studies a different problem from tokenization. Most DNA tokenization methods are designed to define useful sequence units or vocabularies GeneZip targets high-ratio, controllable length reduction for long-context inputs, distinguishing itself from BPE tokenization which does not enforce a per-sample BPT target, a bounded routing budget, or a region-wise allocation constraint such as Nbr ≈ N⋆r GeneZip reaches about 170 bp/token on these windows, enabling the downstream speed and scale advantage compared to methods like MxDNA
Limitations
The current GeneZip pre-training is derived on human genomic data The supervision is centered on static region annotations and the Region-Aware Ratio objective, and a promising next step is to incorporate richer function-oriented supervision beyond annotation labels The current GeneZip pre-training is derived on human genomic data
The current GeneZip pre-training is derived on human genomic data The supervision is centered on static region annotations and the Region-Aware Ratio objective, and a promising next step is to incorporate richer function-oriented supervision beyond annotation labels
GeneZip studies a different problem from tokenization. Most DNA tokenization methods are designed to define useful sequence units or vocabularies GeneZip targets high-ratio, controllable length reduction for long-context inputs, distinguishing itself from BPE tokenization which does not enforce a per-sample BPT target, a bounded routing budget, or a region-wise allocation constraint such as Nbr ≈ N⋆rGeneZip reaches about 170 bp/token on these windows, enabling the downstream speed and scale advantage compared to methods like MxDNA
GeneZip studies a different problem from tokenization. Most DNA tokenization methods are designed to define useful sequence units or vocabulariesGeneZip targets high-ratio, controllable length reduction for long-context inputs, distinguishing itself from BPE tokenization which does not enforce a per-sample BPT target, a bounded routing budget, or a region-wise allocation constraint such as Nbr ≈ N⋆rGeneZip reaches about 170 bp/token on these windows, enabling the downstream speed and scale advantage compared to methods like MxDNA
GeneZip studies a different problem from tokenization.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Better long-context DNA modeling by enabling 128K context and 636M-parameter variants on a single A100 80GB GPU, which makes large-scale long-context DNA pretraining more practical.
-
Improved efficiency for downstream fine-tuning, achieving eQTL task fine-tuning 50.4× faster than JanusDNA (50 vs. 2520 minutes) and reducing inference latency across all tested contexts to remain consistently fast even up to 1M bp input lengths.
-
Redundancy-aware compression for DNA, where GeneZip
assigns higher local BPT to TE-derived interspersed repeats and tandem repeats
without requiring explicit repeat supervision during training, leading to stronger local compression on predictable repeat subsets than generic routing methods. -
Region-aware allocation for downstream tasks, allowing the system to select task-specific priors such as
cis-regulatory focused
for eQTL orpromoter-distal regulatory
for TISP, which results in thebest average rank among compared sequence models.
-
Controllable compression interface by using region multipliers to specify where the token budget is spent, allowing researchers to adjust resolution based on biological relevance, such as setting promoter or intergenic resolution higher for tasks like transcription initiation signal prediction.
Related papers
- SeekRBP: Leveraging Sequence-Structure Integration with Reinforcement Learning for Receptor-Binding Protein Identification
- Genome-Factory: A Library for Tuning, Deploying, and Interpreting Genomic Foundation Models
- AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing
- C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation
- Leveraging genomic deep learning models for the prediction of non-coding variant effects
- Adaptive Allele-Haplotype Incidence Indexing for Pangenome Queries: Representation, Reconstruction, and Benchmarking