Peptiter / DiscoveryLab
Explore workflow
Competitive position

DiscoveryLab is not claiming better weights. It is claiming a better discovery control plane.

ApexGO, PepMLM, PepTune, and RFpeptides are strong specialized generators. DiscoveryLab becomes better when a research team needs those model families to run as executable, auditable, test-covered lanes with controls, calibration gates, wet-lab handoff, and explicit fallback status.

Verdict

Better at accountable execution, not yet better at every model's core science.

The right comparison is not one model against another. The sharper claim is that DiscoveryLab turns the leading generator patterns into a controlled competition system where a candidate cannot hide its provenance, missing evidence, or calibration state.

We should not claim

Superior wet-lab performance over the published models.

ApexGO and RFpeptides report substantial experimental validation. DiscoveryLab's current roadmap lanes are local deterministic fallbacks until real artifacts, licenses, and calibration manifests are attached.

We can claim

A stronger operational layer for generator competition.

DiscoveryLab already makes the four lanes executable, auditable, and test-covered, while keeping controls, near misses, manifest checksums, caveats, and calibration gates visible.

Comparison matrix

The external systems are specialized generators. DiscoveryLab is the competition harness around them.

Each lane has a legitimate scientific niche. DiscoveryLab's advantage is strongest after generation: evidence accounting, controls, calibration, and handoff.

ApexGO
Template-preserving antimicrobial lead optimization.
Transformer VAE plus Bayesian optimization; synthesized 100 optimized peptides; reported strong in vitro hit rates and Gram-negative improvement.
Implements the ApexGO-style lane as parent-preserving optimization with similarity, edit-distance, parent controls, and edit-attribution evidence requirements.
ApexGO is ahead as a published wet-lab validated antimicrobial optimizer. DiscoveryLab is ahead when the job is unbiased batch export, audit receipts, and calibrated handoff across multiple generator families.
PepMLM
Target-sequence-conditioned linear peptide binder generation.
Fine-tunes ESM-2 with span masking so binder regions can be reconstructed from target sequence context, without requiring a target structure.
Implements the PepMLM-style lane as target-sequence-only generation with target sequence hashes and required co-folding, binding, or orthogonal follow-up before advancement.
PepMLM is ahead as a sequence-conditioned model. DiscoveryLab is stronger at making sequence-only proposals visibly provisional until structure, binding, and calibration evidence arrive.
PepTune
Multi-objective generation of therapeutic peptide SMILES.
Uses masked discrete diffusion and Monte Carlo Tree Guidance to balance binding, permeability, solubility, hemolysis, non-fouling, and other objectives.
Implements the PepTune-style lane as Pareto selection over activity, solubility, stability, permeability, low toxicity, and synthesis, with near-miss calibration controls retained.
PepTune is ahead on model-native multi-objective generation. DiscoveryLab is stronger where every Pareto claim must keep controls, objective vectors, and probability-language gates attached.
RFpeptides
Structure-first de novo macrocyclic peptide binders.
Uses denoising diffusion for macrocycles and reported high-affinity binders across diverse protein targets, with close agreement between design models and crystal structures.
Implements the RFpeptides-style lane as macrocycle generation with cyclization metadata, closure quality control, synthesis risk tags, and complex-prediction follow-up requirements.
RFpeptides is ahead as a macrocycle design engine. DiscoveryLab is stronger as the governance layer that refuses to treat a macrocycle as promotable without closure QC, synthesis accounting, and returned assay calibration.
Operating model

Why DiscoveryLab can become the better product even when another model is the better generator.

A research team does not only need candidate sequences. It needs to know what was generated, by which artifact, against which objective, under which caveats, with which controls, and whether returned assay data supports stronger language.

Executable lanes

All four roadmap families run through a typed generator competition harness.

Audit manifest

Parent controls, negative controls, near-miss calibration rows, Pareto assignments, and active-learning picks are kept in the same manifest.

Claim boundary

Scores stay rank-only unless a frozen calibration dashboard says probability language is allowed.

Fallback honesty

Current local generators are marked deterministic fallback and not model-backed.

Wet-lab handoff

Candidates carry required next evidence, caveats, lab metadata, and handoff-ready records.

Test coverage

Unit tests assert all four lanes, model-backed false status, manifest integrity, Pareto mapping, and calibration dashboard behavior.

Research anchors

Sources used for the comparison.

These links anchor the external-model side of the page. The DiscoveryLab side is grounded in the local generator competition harness, manifest rules, calibration dashboard, and unit tests.

ApexGO

Nature Machine Intelligence, published 13 May 2026

Open source
PepMLM

target sequence-conditioned peptide binders

Open source
PepTune

ICML 2025 multi-objective discrete diffusion

Open source
RFpeptides

macrocyclic protein-binding peptide design

Open source