From Sequence to Shortlist — Peptide Development Field Guide

01 · Mental model

One system. Nine evidence gates.

A peptide platform does not “predict a drug.” It moves a large possibility space through increasingly expensive evidence gates until a scientist has a small, diverse, testable set.

Target biologyWhat must change?

Candidate libraryWhat can we try?

Structure prepWhat exactly are we docking?

Pose searchWhere could it sit?

InteractionsWhat contacts support it?

ShortlistWhich evidence is strongest?

SynthesisCan we make it cleanly?

AssaysDoes it bind and function?

FormulationCan it be loaded and delivered?

The scientist asks

Is the biological hypothesis meaningful, and what experiment would falsify it?

The software calculates

Sequences, structures, poses, interactions, warnings, and reproducible evidence.

The experiment decides

Whether a candidate truly binds, changes function, survives, and can be formulated.

The central product rule

Every screen must distinguish what was computed from what was measured. A convincing visualization cannot turn a docking score into experimental affinity.

02 · Peptide fundamentals

The biology vocabulary your API will encode.

At 2–7 residues, small chemical details dominate. A hidden terminal cap or incorrect protonation choice can change charge, solubility, pose, and the meaning of every downstream score.

Amino acid / residue: An amino acid is the building block; once linked into a peptide chain, it is usually called a residue.
Peptide bond: The covalent link joining residues. It creates the repeating backbone while side chains provide chemical identity.
Sequence: The ordered one-letter residue string, such as WKR. It does not define a unique 3D shape.
Conformer: One physically possible 3D shape of the free peptide. Flexible peptides can have many.
Binding pose: A candidate 3D arrangement of the peptide inside or near the protein binding site.
Protein pocket: The region selected for local docking, defined by a known ligand, residues, or coordinates.
Protonation and pH: Which atoms carry protons and charges at a given pH. This changes electrostatic interactions.
Terminal chemistry: Free, acetylated, amidated, or otherwise modified N/C termini. It must be explicit for short peptides.
Hydrophobicity: Tendency to avoid water and associate with nonpolar environments, including membrane interiors.
Solubility: How readily the peptide remains dispersed in the chosen solvent and formulation conditions.

Interactive peptide strip

Touch the chemistry.

Select a residue to see how its side chain changes the candidate’s behavior.

Select a residue.

Its class, likely contacts, and formulation implications will appear here.

Useful developer analogies

Scientific idea	Engineering analogy
Sequence	Source code: compact instructions, not the runtime state.
Conformer	One possible runtime state under a specific environment.
Docking pose	A proposed integration arrangement between two systems.
Docking score	A heuristic benchmark, not measured production performance.
Experimental assay	The integration test against physical reality.

Files that cross system boundaries

Format	What it carries
FASTA	Protein or peptide sequence text.
PDB / mmCIF	Macromolecular atom coordinates and structure metadata.
SDF / SMILES	Small-molecule-like chemical connectivity and coordinates.
CSV	Human-reviewable candidate and score tables.
JSON manifest	Inputs, versions, parameters, seeds, artifacts, and provenance.

03 · Length strategy

Two residues and seven residues are different products.

They differ in library size, training coverage, flexibility, contact surface, compute cost, and which docking assumptions are defensible. We therefore validate and rank by length strata.

Choose peptide length

400

possible canonical sequences · 20²

Do not compare raw scores across lengths.

Atom count, contacts, flexibility, and scoring bias change with length.

Dipeptide · 2 residues

Enumerate the entire space.

All 400 canonical dipeptides can be scored reproducibly before expensive structural work.

Library routeExhaustive enumeration

Docking experimentBenchmark Vina and ADCP

Structural complexityLow, but termini dominate

Expected computeLow for enumeration; moderate for replicated docking

Validation warningBundled training data contains only limited dipeptide examples. Target conditioning must be tested separately.

What the local data tells us

The bundled 10,000-row PepMLM training file contains 2,340 sequences of seven residues or fewer, but only 46 dipeptides. The model can technically emit two residues; that does not establish equal reliability at every length.

04 · Liposome intent

Small is a formulation hypothesis, not a delivery result.

Liposomes can carry peptides in different physical locations. Charge, hydrophobicity, termini, lipid composition, pH, preparation, and release conditions determine what is practical.

Aqueous-core encapsulationMost intuitive for a soluble, hydrophilic peptide. Encapsulation and leakage must be measured.

Bilayer associationPossible for a sufficiently hydrophobic peptide, but aggregation and release behavior become critical.

Surface attachmentRequires an anchor or linker and becomes a different chemical entity with different biological behavior.

What software can track

Molecular weight, pH-dependent charge, pI, hydrophobicity, solubility and aggregation warnings, terminal chemistry, and intended loading mode.

What only experiments can establish: encapsulation efficiency, particle size/PDI, zeta potential, leakage, release, storage or serum stability, cellular delivery, and retained activity.

05 · Discovery funnel

Every stage earns permission for the next.

Scroll through the workflow. The left panel translates each scientific decision into a software responsibility, durable artifact, known failure mode, and release gate.

Define target biology and mechanism

Docking cannot rescue a weak biological hypothesis. Start with the functional change and the assay that would detect it.

Obtain a suitable protein structure

Experimental structures are preferred when relevant; predicted structures require confidence and conformation warnings.

Define the binding site

The MVP assumes a known site. Unknown-site discovery is a separate, more expensive product.

Establish positive and negative controls

Controls are not optional decoration; they are the bridge from plausible computation to meaningful evaluation.

Generate or enumerate candidates

Use exhaustive search when it is affordable and generative modeling when the sequence space becomes enormous.

Apply physicochemical filters

Filters reduce compute and synthesis waste, but they should not pretend to predict liposome performance.

Prepare protein and peptide structures

Structure preparation is part of the scientific method, not an invisible file conversion.

Run fast docking

Fast docking is a triage stage. Its purpose is to reduce the library, not make an affinity claim.

Run thorough docking on finalists

ADCP recommends many independent replicas; this stage belongs in isolated workers, not a web request.

Refine and cluster poses

Repeated convergence toward a pose family is stronger evidence than a single extreme score.

Calculate residue interactions

Hydrogen bonds, salt bridges, hydrophobic and aromatic contacts explain why a pose may be plausible.

Produce a transparent shortlist

The shortlist is a decision aid, not a claim that the top candidate binds best.

Synthesize and assay candidates

Identity, purity, solubility, stability, binding, and functional assays are separate evidence layers.

Test liposome formulation

Formulation is an experimental program connected to discovery—not a checkbox computed from sequence.

Feed experimental results back

The platform becomes valuable when each experiment makes the next one more informed and reproducible.

06 · Model and tool guide

Use a funnel, not a model collection.

The best architecture assigns one narrow responsibility to each tool and validates it on our target class. More models do not automatically create more truth.

Tool	Product role	Input → output	Best fit	Compute / license	Scientific limit
PepMLM Generate	Target-conditioned candidate generation and pseudo-perplexity.	Target sequence + length → peptide sequences	Sample 4–7-mers; score enumerated short candidates.	GPU preferred; existing model.	Sequence plausibility is not binding or affinity.
Enumeration Generate	Complete canonical library for tiny sequence spaces.	Length/alphabet → every unique sequence	400 dipeptides; optional 8,000 tripeptides.	CPU; deterministic.	Combinatorics grow rapidly after length 3.
AutoDock Vina Dock	Experimental ligand-style route for fragment-sized peptides.	Prepared receptor + ligand → poses and scores	Benchmark on 2–3-mers.	CPU; Apache 2.0.	Not peptide-specific; must be benchmarked against controls.
ADCP Dock	Flexible peptide folding and docking.	Receptor + peptide sequence/site → pose ensemble	Primary 4–7-mer technical spike.	Many CPU replicas; open source.	Can output local clashes; requires minimization and convergence analysis.
OpenMM Refine	Energy minimization and later finalist simulation.	Complex + force field → minimized/simulated structures	Top poses after docking.	CPU/GPU; MIT/LGPL components.	Results depend on preparation, force field, solvent, and protocol.
PLIP / ProLIF Analyze	Residue contacts and interaction fingerprints.	Complex pose → typed interactions	Every retained pose and cluster representative.	CPU; open source.	Rule-based contacts describe geometry, not binding strength.
Protenix Cross-check	Independent all-atom complex-structure prediction.	Biomolecular inputs → candidate complex structures	Cross-check top finalists.	GPU; Apache 2.0.	Must be validated specifically on ultra-short peptides and our targets.
Boltz-2 Cross-check	Alternative complex prediction and experimental affinity layer.	Complex inputs → structures / affinity estimates	Research comparison on finalists.	GPU; MIT.	Affinity validation is strongest in small-molecule contexts; do not assume peptide accuracy.
PBCNet2.0 Analyze	Experimental pairwise relative-affinity score.	Reference/query complexes → affinity delta	Optional 2–3-mer experiment with aligned reference.	CPU/GPU; current isolated runtime.	Protein–small-molecule training domain; not a docking engine.

Recommended MVP stack

PepMLM/enumeration → Vina experiment for 2–3 → ADCP for 4–7 → OpenMM minimization → PLIP/ProLIF interactions → transparent ranking → optional Protenix cross-check.

Why not integrate every open model?

Each model adds packaging, compute, schema, monitoring, and validation work. Until a baseline pipeline recovers known poses and ranks controls, adding more models mainly adds disagreement.

Where do DiffPepBuilder and PepEDiff fit?

They are promising future generator experiments. They should enter only after a frozen benchmark can show whether they add diverse, experimentally useful candidates in the 2–7 range.

Why is Rosetta optional?

FlexPepDock is scientifically relevant, but deployment and commercial licensing must be resolved. The adapter architecture keeps it available without making it an MVP dependency.

07 · Connected architecture

The science pipeline and software pipeline are the same graph.

Select a diagram node to see responsibility, inputs, outputs, failure isolation, and scaling behavior.

Why docking cannot remain inside FastAPI

Thorough peptide docking may require many independent replicas and tens of CPU-hours per candidate. The web request should create durable work and return immediately; isolated workers should own computation, timeout, retry, cancellation, and logs.

08 · Data and provenance

A result is an artifact graph, not a CSV row.

Every score must trace back to exact structures, chemistry, software, parameters, seeds, and container versions. That trace is what makes the work reviewable and repeatable.

ExperimentTarget, intent, owner, scope, status.

Target + pocketOriginal/prepared structures and site definition.

CandidateSequence, length, termini, source, descriptors.

Docking runEngine, config, seed, status, logs.

Pose + clusterRaw/refined structures and convergence.

EvidenceInteractions, rank profile, assays, formulation.

Immutable identity

Inputs and outputs receive content hashes. A rerun produces new IDs rather than replacing an old result. Ranking policies are versioned, so a new heuristic cannot silently rewrite history.

Minimum reproducibility envelope

Original and prepared structures, terminal state, pH, model/engine version, container digest, resolved parameters, random seed, runtime, warnings, and artifact checksums.

Why store raw and refined poses?

Refinement changes coordinates. Keeping both separates what the docking engine produced from what a later physics-based step repaired.

Example candidate manifest

{
  "experiment_id": "exp_enpp1_2026_001",
  "candidate_id": "cand_WKRF_0042",
  "sequence": "WKRF",
  "length": 4,
  "termini": { "n": "free", "c": "amide" },
  "formulation_ph": 7.4,
  "source": {
    "type": "pepmlm",
    "run_id": "pepmlm_run_20260714_103122",
    "pseudo_perplexity": 3.84
  },
  "receptor": {
    "prepared_sha256": "a839…c72",
    "pocket_id": "pocket_active_site_v2"
  },
  "docking": {
    "engine": "adcp",
    "image_digest": "sha256:91d…0b4",
    "seeds": [11, 23, 47, 89],
    "pose_clusters": 3
  },
  "evidence": {
    "tier": "B",
    "required_contacts": ["A:205", "A:547"],
    "warnings": ["cross_length_rank_not_calibrated"]
  }
}

09 · Ranking

Evidence enters a funnel. Truth does not exit as one number.

The default sort can help a scientist navigate, but every component remains visible and comparable only within its validated length, target, engine, and protocol group.

Sequence contextPepMLM + diversity

Physicochemical eligibilitycharge · solubility · termini

Valid geometryclashes · bonds · pocket

Docking within stratumsame engine + length

Replicate convergenceindependent seeds

Pose-cluster supportrecurring pose family

Mechanistic contactscritical residues + fingerprint

Scientist reviewdiverse experimental panel

Tier A

Converged

Multiple seeds, valid refined geometry, supported cluster, and required contacts.

Tier B

Plausible

Valid pose with partial convergence or incomplete contact support.

Tier C

Uncertain

Weak convergence, warnings, or missing evidence. Keep only for diversity or learning.

PBCNet remains a separate experiment

It compares relative affinity for protein–ligand complex pairs and was trained in a small-molecule context. It cannot become the default peptide rank until a representative peptide benchmark shows incremental value.

10 · Validation

Software correctness is not scientific usefulness.

A job can complete perfectly and still produce irrelevant science. We therefore validate infrastructure, pose recovery, ranking signal, biological activity, and formulation as separate layers.

Pose recovery

Remove a peptide from a known complex, re-dock it, and measure whether near-native poses appear in the top 1, 5, and 10.

Ranking signal

Compare known binders, inactive or scrambled controls, and generated candidates within each length and engine stratum.

Physical evidence

Confirm identity, stability, binding, function, formulation, and activity after release under documented conditions.

Computational metrics

Ligand RMSD (LRMSD): How far the peptide moved from the known pose after aligning the receptor.
Interface RMSD (iRMSD): Deviation of residues at the binding interface.
Fraction of native contacts: How many known residue contacts the predicted pose recovers.
Cluster convergence: Whether independent searches repeatedly produce a similar pose family.

Sampling failure vs ranking failure

If no near-native pose was generated, the search failed. If a near-native pose exists but ranks poorly, the scoring/ranking failed. The product must report both so we improve the correct subsystem.

Required strata

Report dipeptides, tripeptides, and 4–7-mers separately. A pooled average can hide a method that works only for one group.

What the platform can claim	What requires experiments
“PepMLM produced this target-conditioned sequence.”	“The peptide binds the target.”
“These docking replicas support this pose cluster.”	“This is the true bound pose.”
“This pose contains the following geometric contacts.”	“These contacts cause biological activity.”
“This candidate ranks highly within this validated protocol group.”	“It has the best affinity or IC50.”
“Its properties suggest this formulation route is worth testing.”	“It will encapsulate, release, reach cells, or work in vivo.”

11 · Delivery plan

Build the evidence spine first.

The technical MVP is achievable in weeks. A scientifically validated discovery workflow takes months because controls, benchmarks, synthesis, assays, and formulation cannot be compressed into frontend work.

Feasibility spike2–3 weeks

One target · known pocket · controls · Vina/ADCP comparison · measured runtime

Internal MVP8–12 weeks

Async workers · experiment model · poses · interactions · transparent ranking · exports

Validated platform4–6 months

Frozen benchmark · calibrated strata · reliability · security · operations

Experimental programRuns in parallel

Synthesis · binding/function assays · stability · liposome formulation · feedback

Parallel engineering workstreams

Scientific benchmark and controls
Target, site, known binders, negatives, frozen evaluation.
Docking-engine evaluation
Preparation, Vina/ADCP adapters, runtime, pose recovery.
Platform infrastructure
PostgreSQL, S3, queue, workers, idempotency, monitoring.
Analysis and interface
Clusters, interactions, evidence tiers, Mol* review, exports.

Minimum working team

Backend / ML engineer
Orchestration, model adapters, artifacts, scientific pipeline.
Frontend / full-stack engineer
Experiment workflow, status, 3D review, comparisons.
Computational chemistry scientist
Protocol, target, preparation, benchmark, interpretation.
Wet-lab and formulation collaborators
Synthesis, assays, liposomes, measured feedback.

12 · First experiment

Two weeks to replace assumptions with evidence.

Before building the full platform, run one narrow scientific spike that measures tool compatibility, pose quality, runtime, and the usefulness of the proposed outputs.

Choose one target structure with a known pocket.
Obtain one known 2–7-mer binder and one negative or scrambled control, if available.
Lock terminal chemistry, formulation pH, and intended liposome loading mode.
Enumerate and score all 400 dipeptides; optionally score all 8,000 tripeptides.
Sample diverse PepMLM candidates at lengths 4–7.
Select a small balanced panel from each length stratum.
Compare Vina and ADCP for 2–3-mers; test ADCP for 4–7-mers.
Minimize, cluster, calculate interactions, and inspect the pose evidence.
Review with the scientist before designing cross-length ranking.
Use measured runtime and recovery to lock worker size, presets, and MVP scope.

Exit gate

We proceed only when:

One known complex can be prepared, docked, parsed, clustered, analyzed, and explained without manual file surgery—and the scientist finds the output useful enough to guide a decision.

What this spike will not prove

It will not prove that generated peptides bind or that liposome delivery works. It proves whether the computational workflow is worth productizing.

13 · Learning path

What you need to learn—and what you can leave to specialists.

Your job is not to become a medicinal chemist overnight. It is to understand the scientific contracts well enough to encode assumptions, prevent false claims, preserve provenance, and ask the right questions.

Learn deeply enough to build

Protein and peptide sequence/structure basics.
PDB/mmCIF/SDF semantics and chain/residue identity.
pH, protonation, charge, and terminal chemistry.
Docking sampling versus scoring.
RMSD, clustering, interactions, controls, and benchmark leakage.
Reproducibility, immutable artifacts, and model/data provenance.

Keep a scientist in the loop

Target mechanism and biological relevance.
Structure and binding-site selection.
Preparation protocol and force-field assumptions.
What counts as a useful docking benchmark.
Assay design and interpretation.
Liposome composition, manufacturing, and experimental acceptance criteria.

Searchable glossary

Translate the field.

Filter terms

18 terms

Affinity: Strength of binding under defined conditions. It is measured or carefully modeled—not identical to docking score.
Binding pocket: The receptor region searched during local docking.
CAPRI metrics: Community-standard complex-model metrics including LRMSD, iRMSD, and native contacts.
Conformer: One possible 3D shape of a molecule or peptide.
Docking: Computational search and scoring of candidate binding poses.
Evidence tier: A product label summarizing convergence and evidence completeness without claiming affinity.
Interaction fingerprint: A comparable representation of contact types between peptide and target residues.
Ligand: A molecule that binds a receptor; here, the short peptide is the candidate ligand.
Liposome: A lipid-bilayer vesicle that can carry payloads in its core, membrane, or surface.
LRMSD: Peptide displacement from a reference pose after receptor alignment.
pH / protonation: Environmental acidity and the resulting distribution of proton/charge states.
Pose: A candidate position, orientation, and conformation of the peptide relative to the receptor.
Pseudo-perplexity: PepMLM sequence plausibility score conditioned on the target; lower is not measured binding.
Receptor: The target protein structure used in the docking calculation.
Replica / seed: An independent search run used to test whether results converge rather than occur by chance.
Residue: An amino-acid unit within a peptide or protein chain.
Terminal cap: A chemical modification at the N or C end that changes charge and stability.
Zeta potential: A measured electrokinetic property used when characterizing colloidal systems such as liposomes.

Primary references and official documentation