From FASTA to Regulatory Hypothesis: A New Method for Sequence-to-Network Inference

Author: CAIRN Institute

Published: August 25, 2026

Read time: 10–12 minutes

#Genomics #Bioinformatics #GeneRegulation #SystemsBiology #OpenScience #SequenceModels

This sequence-to-network prototype was built externally, but validated against atlas-linked data and benchmark edges.

Quick Summary

Today we are publishing a public research update on a new external prototype for sequence-to-network inference.

The motivating question is simple to state and difficult to solve:

• start from a DNA sequence in FASTA format
• infer which gene or promoter that sequence most likely belongs to
• rank the transcription factors most likely to regulate that target
• measure whether each added layer improves recovery of known regulatory edges

This prototype was built outside the main GRN Atlas repository, but validated against atlas-linked data and benchmark edges.

The current result is encouraging, but still narrow:

• the new wide sweep tested 501 atlas-backed regulatory edges across Arabidopsis (250), tomato (250), and petunia (1)
• in blind mode, weighted top-15 recovery improved from 303/501 = 0.605 with a motif baseline to 388/501 = 0.774 with the strongest current hybrid variant
• in target-aware mode, the baseline already reached 487/501 = 0.972, and the best hybrid variants reached 501/501 = 1.000
• the clearest gain came from Arabidopsis blind mode, where top-15 recovery on 250 edges improved from 56/250 = 0.224 to 141/250 = 0.564
• the main remaining bottleneck is not TF ranking by itself, but mapping sequence to the correct target context

This is not yet a system that can take any arbitrary sequence and infer a full gene regulatory network in a generic way.

It is, however, a real step toward that direction: a working pipeline that combines motif evidence, biological priors, promoter similarity, and sequence embeddings, with direct benchmarks showing what those layers add.

Why We Built It

Researchers often start with sequence before they start with a gene name.

A gene regulation workflow may begin with:

• a promoter fragment from a reporter assay
• a cis-regulatory interval highlighted by chromatin data
• a variant-bearing regulatory region
• a candidate conserved noncoding sequence
• a newly assembled region in a non-model plant
• an engineered promoter and a question about likely upstream control

In all of those cases, the biological question is broader than “what motif is present?”

given a sequence, what gene is it likely tied to, which regulators are plausible upstream controllers, and how much confidence should we assign to that claim?

That requires more than a motif scanner. It requires a layered inference problem: sequence-to-target mapping, target-to-regulator ranking, and confidence calibration across multiple weak sources of evidence.

What This Prototype Is

For readers who are not already familiar with it, GRN Atlas is the CAIRN Institute platform for multi-species gene regulatory network analysis.

It combines:

• curated and inferred TF-target relationships
• promoter and motif context
• plant expression layers
• cross-species transfer surfaces
• pathway and trait annotation
• perturbation-oriented downstream analysis

The prototype described here is not the atlas itself.

It is an external experimental branch built to answer a narrower question:

how much regulatory inference can we recover from sequence when we combine motif support, atlas-linked biological structure, promoter similarity, and lightweight sequence models?

The working system is best understood as sequence -> likely target -> ranked candidate regulators, not arbitrary DNA -> complete inferred regulatory network.

What We Built

The prototype was implemented in external grn_tools code and evolved in stages.

1. PWM baseline

The starting point is a promoter-level PWM scan linked to atlas TF annotations, with candidates ranked by motif strength, hit count, and basic network context.

2. Real FIMO integration

We installed and integrated a working local FIMO binary from MEME Suite 5.5.9. That matters for credibility, but it did not turn out to be the main source of benchmark gains.

3. PlantRegMap prior support

We added PlantRegMap prior ingestion into the same external pipeline so a candidate TF can gain support when there is independent plant regulatory prior evidence for the same TF-target relationship.

4. k-mer promoter similarity

We added a lightweight reranking branch based on TF-specific promoter k-mer centroids learned from high-confidence known targets.

Biologically, this asks:

does this candidate promoter resemble promoters already associated with this TF, even when no single motif hit explains the match cleanly?

5. Sequence embedding reranking

We added a working embedding path using the official InstaDeepAI/nucleotide-transformer-v2-100m-multi-species checkpoint. This expands the feature space beyond classical motif hits and simple k-mer counts.

6. Trained downstream reranking

We also implemented a fuller reranking branch over engineered features such as motif statistics, atlas support, k-mer similarity, and embedding similarity. This branch is promising, but still experimental.

How We Evaluated It

Blind mode

In blind mode, the system must work from sequence without being told the correct target gene. This is the more realistic and more difficult setting for an unknown FASTA.

Target-aware mode

In target-aware mode, the true target context is supplied or effectively known. This isolates the TF-ranking problem from the sequence-to-target mapping problem.

That distinction turned out to be one of the most informative parts of the benchmark.

The New Wide Benchmark

The most important result in this update is the new wide cross-species benchmark.

Using the currently available atlas-linked benchmark edges, the sweep covered:

• Arabidopsis: 250 usable edges
• Tomato: 250 usable edges
• Petunia: 1 usable edge

for a total of 501 tested edges.

This is a much stronger validation surface than the earlier pilot slice because it is large enough to separate genuine gains from anecdotal wins.

Main Result: Weighted Top-15 Recovery

The clearest single summary is weighted top-15 recovery across all 501 tested edges.

Blind mode

Target-aware mode

Three things stand out. PlantRegMap alone did not improve blind recovery in this wide sweep. The meaningful blind improvement came from the hybrid promoter-similarity branch. And once the correct target context is supplied, the downstream TF-ranking problem becomes dramatically easier.

What The Benchmark Says Biologically

The aggregate result is useful, but the species-level pattern matters more.

1. Arabidopsis is the hard case

Arabidopsis is where the blind-ranking problem is real and where the method genuinely improves.

At 250 edges and top-15 regulators:

• PWM only: 56/250 = 0.224
• PWM + PlantRegMap: 56/250 = 0.224
• PWM + PlantRegMap + kmer: 141/250 = 0.564

In target-aware mode on the same 250-edge slice:

• PWM only: 236/250 = 0.944
• PWM + PlantRegMap: 250/250 = 1.000
• PWM + PlantRegMap + kmer: 250/250 = 1.000

This tells us that the main bottleneck is not simply “find a better TF ranker.” It is identifying the correct target context from sequence strongly enough that downstream regulator ranking can operate in the right neighborhood.

2. Tomato is probably easier than it looks here

Tomato is close to ceiling on the currently available benchmark slice.

At 250 edges and top-15 regulators:

• blind mode: 247/250 = 0.988 for all three variants
• target-aware mode: 250/250 = 1.000 for all three variants

That is operationally excellent, but scientifically it should make us cautious. It likely means the current tomato slice is unusually easy, motif-dense, or otherwise biased toward highly recoverable edges.

3. Petunia remains too sparse for strong conclusions

Petunia currently contributes only one usable edge in the filtered wide benchmark slice:

• blind mode: 0/1
• target-aware mode: 1/1

That is not enough to support strong claims in either direction.

The Deeper Arabidopsis Benchmark Still Matters

The wide sweep tells us about breadth. A narrower advanced Arabidopsis benchmark still tells us something useful about depth.

On a harder top-10 Arabidopsis blind benchmark, richer modeling layers behaved as follows:

This deeper slice reinforces the same design lesson. The first major gains did not come from swapping one motif scanner for another. They came from combining motif evidence, plant regulatory priors, promoter similarity, and sequence embeddings.

What The Current System Is Good At

1. Recovering plausible regulators when the target is known

This is the strongest current result. Once the sequence is mapped to the correct promoter or transcript neighborhood, TF ranking is already strong enough to be useful as a research aid on this benchmark surface.

2. Improving blind ranking in difficult Arabidopsis cases

The jump from 0.224 to 0.564 at Arabidopsis top-15 across 250 edges is not trivial. It means the hybrid approach is doing more than reordering a few easy examples.

3. Showing that regulatory inference benefits from layered evidence

The strongest gains came from integrating weak signals rather than relying on any single one of them. That is biologically plausible and methodologically useful.

What The Current System Is Still Bad At

1. It does not yet infer a broad de novo regulatory network from arbitrary FASTA

The current system is best understood as sequence -> likely target -> ranked candidate regulators, not arbitrary unknown DNA -> full correct regulatory network. That broader goal remains future work.

2. The benchmark is still uneven

The new sweep is much broader than the pilot, but it is still dominated by Arabidopsis and tomato. Petunia remains too sparse to support strong cross-species conclusions.

3. Some benchmark slices are probably too easy

Tomato is near ceiling even in blind mode, which is useful operationally but not sufficiently stressful as a validation challenge.

4. Target mapping remains the central unsolved problem

The gap between blind and target-aware performance is the most important limitation in the current prototype.

The Main Takeaway

The best result from this entire project is not “we found the right model.”

It is this:

the strongest gains came from treating FASTA-to-network inference as a layered biological inference problem rather than a pure motif-scanning problem.

The best-performing branches were the ones that combined:

• motif support
• target mapping structure
• known regulatory priors
• promoter-level sequence similarity
• embedding-derived sequence context

That is a useful biological and engineering result. It suggests that broader FASTA-to-network inference, if it becomes practical, will probably come from a structured system that integrates multiple weak signals and learns how to weight them.

Recommended Next Steps

Based on the August 25, 2026 rerun, the next priorities are clearer now.

1. Expand the wide benchmark with harder and more diverse held-out slices, especially beyond the easy tomato regime.
2. Improve sequence-to-target mapping, because the target-aware benchmark is already much stronger than the blind benchmark.
3. Keep the trained reranker branch, but treat it as experimental until it is validated on larger held-out slices.
4. Add more species with enough usable benchmark edges to support fair cross-species comparisons.
5. Continue testing alternate embedding backbones only after target-mapping and held-out evaluation are in better shape.

Final View

This project now has a working external prototype for sequence-to-network inference with measurable gains from added biological context.

The evidence today supports a narrower but real claim:

• for known or correctly mapped plant regulatory regions, the current hybrid pipeline can recover plausible upstream regulators substantially better than a motif-only baseline, especially in difficult Arabidopsis blind settings

The evidence does not yet support the stronger generic claim:

• that we can take any arbitrary FASTA and robustly infer its full gene regulatory network in a species-agnostic way

That distinction matters. It is what turns this from a story about hype into a story about progress.

© 2026 CAIRN Institute | Advancing AI for the Public Good