Protein-conditioned nucleic-acid generation

Checking local runtime

Turn selection dynamics into design hypotheses.

Follow AptGEN from early-round SELEX counts through enrichment-aware learning, multi-scale protein conditioning, and candidate triage—entirely on your machine.

Explore the workflow

No cloud calls CSV stays in browser Evidence-labelled outputs

aptgen.local / design LOCAL

Target proteinMulti-scale context READY

01 AGGAAGAATGCAGATTCCCAATAGCG QC 91

49.7M

target-agnostic
pretraining sequences

429

protein targets in
the reported SFT corpus

2.14M

target-associated
SELEX sequences

3×

protein context scales:
global, residue, surface

One system · two routes

From selection history to new-target design

Historical SELEX teaches the model what selection looks like. At design time, a new target can enter through protein features alone—without same-target SELEX.

Model-building route New-target design route

01 Historical selection route

02 SELEX-free design route

01STEP

Historical input

Measure how every sequence changes across rounds.

AptGEN starts from a sequence × round count matrix. Counts are normalized per round before enrichment is interpreted, so deeper sequencing alone is not mistaken for selection.

Input: FASTQ-derived counts or CSV
Output: Normalized round trajectories
Guardrail: Counts are not affinity measurements

Interactive lab

Inspect real round dynamics. Explore real model outputs.

Local data boundaryYour file is parsed in this browser tab and is never uploaded.

DATA INPUT

Early SELEX rounds

Drop a CSV / TSV here or click to choose a local file

Accepted schema

sequence,R1,R3,R5,R7
ACGT...,12,41,190,830
TGCA...,31,28,18,7

At least two numeric round columns. count_R* is recognized automatically.

Enrichment view

Top trajectories Pseudocount

Awaiting a round-count matrix

Load the bundled BRK example or inspect your own experiment.

We will calculate per-round frequencies, diversity, fold change, and enrichment trajectories locally.

DATASET

BRK repository demo

Parsed locally

Unique sequences—valid nucleotide strings

Selection rounds——

Final-round reads—sum of supplied counts

Rising trajectories—positive log₂ frequency change

TOP SEQUENCES

Selection trajectories

POPULATION

Diversity across rounds

SELECTION SHIFT

Final abundance vs enrichment

RANKED OUTPUT

Enrichment candidates

Rank	Sequence	Start freq.	Final freq.	log₂ change	Trajectory

Interpretation boundary: frequencies are normalized within the supplied rows; the bundled demo is an enrichment-ranked excerpt, not a complete round. Enrichment is a selection-derived proxy, and PCR bias or non-binding effects can also change counts.

LOCAL GENERATION

Candidate controls

ACTIVE MODEVerified output snapshot

Versioned candidates; no model load required.

Target protein Number of candidates 12 Temperature 0.80 Nucleus top-p 0.90 Maximum length Random seed

Snapshot mode re-ranks stored local model outputs by transparent sequence QC. Start the local server for fresh sampling.

CANDIDATE SET

BRK / snapshot

Versioned local model output

Returned—sequences

Validity—ACGT alphabet

Unique—within this set

Mean GC—sequence property

SEQUENCE SPACE

Composition map

clean QC flagged

VERSIONED METRICS

Snapshot context

REPLAY

Model: local DPO snapshot
5-mer F1: —
6-mer F1: —
Mean edit distance: —

These quantify recovery of an enriched sequence distribution. They do not establish binding affinity.

RankCandidate sequenceLengthGCComplexityQC priorityFlags

Candidate, not conclusion. Generated sequences require orthogonal filtering and experimental validation. QC priority is not predicted affinity, K_d, or proof of binding.

Inside the model

Protein context at three biological scales

The active MSAG route combines protein identity, residue context, and surface geometry before autoregressive sequence generation.

MSAG DECODERgated hierarchical cross-attention

next-token distribution

ACGT

GLOBAL CONTEXT

What protein is this?

A mean-pooled protein language-model embedding supplies broad family and functional context, including for targets not present in the same-target training set.

Captures: identity · family · function
Does not prove: a physical binding interface

Read every output at the right level

A transparent evidence ladder

Computational confidence and distribution recovery are useful filters. Neither is a substitute for direct biochemical measurement.

LEVEL 01 · OBSERVED

SELEX selection dynamics

Directly observed sequencing counts and their normalized change across rounds.

round frequency
enrichment slope
population diversity

Data-derived

LEVEL 02 · MODELLED

Sequence-distribution recovery

How generated populations resemble held-out enriched sequence sets.

k-mer F1 / JSD
edit-distance recall
validity / uniqueness

Computational proxy

LEVEL 03 · HYPOTHESIS

Structural plausibility

Predicted co-folds and confidence metrics can prioritize structural hypotheses.

Boltz-2 confidence
surface complementarity
optional MD follow-up

Predicted structure

LEVEL 04 · VALIDATED

Experimental binding

Direct wet-lab assays are required to establish binding, affinity, and specificity.

SPR / BLI / MST
EMSA / pull-down
counter-target panel

Gold standard

A≠K_d

Core interpretation rule

Enrichment, model reward, and structure confidence are different quantities.

The website never relabels a surrogate score as measured affinity. Bundled evidence is strongest within represented protein families; broad unseen-family generalization is not established. Advance candidates only when the corresponding experiment exists.

No external service required

Open the story.
Or load the model.

The static site includes a real BRK round-count excerpt and versioned local model outputs. The bundled Python server adds live MSAG sampling from repository weights and protein tensors.

01 Works by opening index.html02 Standard-library local server03 Lazy checkpoint loading

LOCAL TERMINAL

# From the AptGEN repository root

$ python website/server.py

AptGEN offline website: http://127.0.0.1:8000

Live inference: enabled

# Lightweight static-only mode

$ python website/server.py --no-live-inference

LOCALHOST ONLYDATA POLICY: LOCAL