Release notes

Single-file .amino pool ingest

Streams a real phage-display NGS export (Count/Ratio/Length/AA_Seq, no manifest) into the database, deduping by sequence and keeping abundance. Chain-format-agnostic (VHH now; scFv / paired VH+VL reserved). Antigen-panel and known-positive (binding-Excel) loaders land the specificity design alongside.

10⁶-scale cheap sequence funnel

Stage 1 streams the whole pool with O(N) gates (length, structural Cys) + CDR3-bucket clustering, doing ~99% of the reduction before anything structural runs. Abundance is a weak tiebreak, never a gate — rare unique-CDR3 clones survive.

Epitope × specificity matrix

Scores each candidate against every antigen-panel construct, assigns an epitope bin, and enforces specificity via the counter-screen (e.g. DLL1 / DLL4). Reproduces the shape of the wet-lab binding table. Specificity the cheap scorers can't resolve is flagged, not silently passed.

Anchor-validated ranking

Composite rank (affinity / interface / on-target epitope / specificity / developability / weak abundance) with a tunable-threshold shortlist. Known positives validate the ranking (recall into top decile + specificity reproduction) — never train it, since there are no negatives yet.

Two-tier docking depth

Bulk triage uses cheap scoring; true docking poses run only on the final shortlist, gated behind an explicit toggle and hard-capped. Never fabricates a pose when no backend is configured.

Praxis platform branding

Platform is Praxis (Greek for 'putting knowledge into action'), owned by IASO Bio. Whitepaper, sidebar, tooltips, and tutorials reflect the new name; internal codename remains GenPrime in repo metadata.

Sidebar reorganized around the canonical 6-step workflow

Three sections: Workflow (Import → Epitope Picker → Pipeline → Run History → Interface → Mutation Resistance → Candidates → Compare → CAR Design), Specialized (VHH Design, Two-VHH CAR, Mini-Binder, Programmable Design), Utilities (Structure Viewer, Scoring Dashboard, Models). Agent / Guides / Settings collapsed into three single-entry destinations below the main groups.

Praxis platform whitepaper

Long-form (~6500 EN words + ZH translation) master documentation at /guides/whitepaper covering therapeutic mission, industry context, design philosophy, platform capabilities by 6-step workflow, differentiation, future roadmap, and an honest open-questions list. Sticky TOC, EN/中文 toggle, single-source markdown editable via PR.

Guides hub at /guides

Single-entry sidebar destination consolidating every reading surface: Whitepaper, Tutorials, Scoring Audit, Readout Guide, Knowledge Map, Use Cases, References. Replaces the previous expandable Learning group.

Bilingual hover hints across the sidebar

Every sidebar nav item and section label has a tooltip showing EN above and ZH below with a hairline separator. ~33-term in-product glossary via the new <InfoTip term="…" /> component; <TutorialLink slug="…" /> for module-page top-right tutorial CTAs.

Bilingual score tooltips with execution logic

Every score (44 entries — ipTM, pTM, pLDDT, PAE, iPAE, BSA, AbLang, TAP, ThermoMPNN, AggreScan, CDR geometry, HCDR3, glyco, MW/pI/FcRn, cross-species, selectivity, tonic risk, CAR composite, Pareto rank, composite score, etc.) now exposes bilingual range / meaning / calc / caveat, plus links to its Readout Guide entry and the Scoring Audit page.

Vendor-name scrub

Removed provider references from every user-facing surface (Overview Models card, /pipeline backend selector, /interface job-status copy, /mini-binder banner, /settings labels, /releases entries, and docs walkthroughs). Replaced them with capability labels (API / GPU / Local); backend config identifiers remain internal.

Mutation Resistance + Readout Guide + Wet-Lab Loop + Wave-1 Report

/mutation-resistance (escape-mutation panel re-scoring), /docs/readout-guide (per-score operating manual), /docs/wet-lab-loop (Waves 2–4 architecture), /docs/wave-1-report (Boltz-2 swap sanity check on 30 BCMA candidates) — all added in the same sprint.

Deep tutorial sub-pages

/docs/tutorials/[slug] dynamic route renders an 8-section walkthrough per feature with prev/next nav. All 17 slugs covered: pipeline, epitope-picker, programmable-design, mini-binder, vhh-design, two-vhh, candidates, scoring, compare, viewer, interface, import, car-design, rare-diseases, runs, models, agent.

Single source for tutorials

Index card + deep page both render from web/src/app/docs/tutorials/_data.ts — adding or revising a tutorial is one edit.

Standard SaaS narrative voice

Stripped all internal-team attributions from user-facing UI/docs/code. References to specific individuals replaced with platform-logic framing. Team subsection removed from architecture walkthrough.

Sidebar declutter

Roadmap entry removed from Learning. Architecture, Release Notes, and Research moved into Settings as sub-pages — reduces top-level nav weight.

Epitope Picker (v0)

Interactive 3D Mol* viewer with curated presets for BCMA / CD20 / HER2 / GPRC5D. Produces canonical chain:residue token strings for every downstream design surface.

Partial-diffusion HCDR3 redesign

Backbone mode that redesigns only HCDR3 against a parent VHH framework. The supported path for warm-start affinity-maturation workflows.

Cross-reactivity counter-screen stage

Pipeline stage that scores designed binders against a configurable panel of off-target antigens; flags candidates with elevated cross-reactivity risk.

Scoring depth — BSA, CDR geometry, AggreScan3D, tonic-clustering

Buried surface area on every complex; ANARCI/IMGT CDR-length validation; AggreScan3D aggregation propensity; tonic-clustering risk proxy (composite of surface hydrophobicity + net charge + Fv-Fv docking). All visible in /candidates and the Scoring Audit doc.

Agent — DAG-planning + 11-tool surface

Agent (Claude Opus 4.6 via Anthropic API) plans workflows as DAGs and renders them inline; 11 platform tools (run_design, query_candidates, score_candidate, chroma_generate, export_candidates, propose_dag, …) invocable from natural-language prompts.

Architecture walkthrough (v4) + bilingual toggle

English-first comprehensive walkthrough covering current state, 6-stage pipeline, MCP architecture, frontend inventory, and ~4-week roadmap. CN version maintained as translation.

Three CAR Architectures

In vivo LNP delivery (VHH-targeted LNP + single-VHH CAR), bispecific tandem (dual-VHH in one construct with linker), and hybrid (LNP targeting + bispecific CAR). Each architecture has distinct molecular topology and validation requirements.

Expert Review System

9 structured scientific questions covering LNP display methods, component options, target libraries, and design strategy. Answers dynamically configure the pipeline.

Construct Component Library

20 curated CAR domain components: signal peptides (3), hinges (4), transmembrane (2), costimulatory (4), signaling (2), linkers (5). All with real amino acid sequences from UniProt.

Construct Assembly & Validation

Combinatorial assembly engine producing all construct variants (orientations × VHH combinations). Four-level validation: sequence (MW, pI, glyco), mRNA optimization (CAI, GC%), LNP compatibility scoring, and functional heuristics (signaling, persistence, AICD risk, fratricide).

Target Preset Libraries

6 T-cell markers (CD3ε, CD4, CD5, CD7, CD8α, NKG2D) and 8 disease targets (CD19, CD20, CD22, BCMA, GPC3, HER2, EGFR, Mesothelin) with PDB references, epitope suggestions, and clinical notes.

Structure Validation (Level 2)

Full construct structure prediction via AlphaFold-Multimer for top-K constructs. VHH domain independence, CDR accessibility, and inter-domain interaction analysis.

Four VHH Design Modes

De novo design from target structure, germline-guided with camelid IGHV3 frameworks, template-based CDR optimization, and humanization with FR2 hallmark preservation. Each mode has distinct scientific logic and tool recommendations.

VHH Hallmark Validation Gate

Four-pillar VHH identity verification: FR2 hallmark residues (IMGT 42/49/50/52), CDR3 integrity, canonical disulfide (C23-C104), and ANARCI annotation. Three stringency levels: strict, relaxed, off.

Germline Database Integration

5 camelid IGHV3 + 5 human IGHV3 germline references with germline assignment, framework extraction, and closest human germline finder for humanization.

VHH-Specific Scoring Presets

Four weight presets (VHH Default, Humanization, CAR Compatible, Custom) with stability and aggregation upweighted for single-domain biology. Mode-specific auto-selection.

Internal API for Two-VHH CAR

VHHDesignRequest/Response schemas and VHHDesignEngine orchestrator with well-defined interface for the Two-VHH CAR Architect module to call programmatically.

Streaming Chat with Claude Opus 4.6

Real-time token-by-token streaming via Server-Sent Events. Markdown rendering with tables, code blocks, and formatted lists. Persistent conversation threads stored in SQLite.

11 Platform Tools

Agent can: run_design (launch workflows), query_candidates (filter/sort), chroma_generate (call Chroma directly), query_database (flexible SQL), score_candidate, export_candidates, explain_platform (platform docs), create_plan/update_plan_step (multi-step planning).

Platform Guidance

Agent serves as interactive documentation — answers questions about pages (Pipeline, Programmable Design, Candidates, Scoring), metrics (ipTM, pLDDT, Pareto ranking), tool selection, and workflows. Knowledge priority: local data > platform features > domain knowledge.

Multi-Step Planning

For complex research tasks (3+ steps), the Agent creates structured plans with goal and steps, executes each step with tool calls, and summarizes findings. Plans are persisted in the database.

Chat UI with Thread Management

Thread sidebar (create/delete/rename), auto-scrolling messages, file attachment support (.pdb, .fasta, .csv), tool call badges, loading animation. 18 suggested questions for new users.

Chroma Modal GPU Deployment

Chroma deployed on Modal A100-40GB with 4 pretrained models (GraphBackbone, GraphDesign, GraphClassifier, ProteinCaption). Persistent volume for weight caching. Cold start ~22s, generation ~25-27s for 100-residue proteins.

Chroma MCP Server (6 tools)

MCP server exposing: chroma_sample, chroma_sample_conditioned, chroma_design, chroma_pack, chroma_score, chroma_redesign. Supports all 6 conditioner types + composed conditioners.

Programmable Design Page

Dedicated workflow page (/programmable-design) for Chroma's conditioner system. Radio card UI for constraint type selection, dynamic parameter forms, generation progress tracking. Designed for Dr. Zhang's demo.

Pipeline Integration

Chroma added to Pipeline page tool dropdown. When selected, shows conditioner-specific configuration UI. Chroma added to scout_tools in pipeline_config.yaml.

Candidate Detail — Chroma Metrics

Candidate detail page now shows source tool badge, conditioner type badge, and constraint satisfaction metrics (constraint_rmsd, symmetry_quality) for Chroma-generated candidates.

Release Notes Page

This page — platform documentation with feature descriptions, smoke test reports, and verification records for each release.

MCP Architecture

Adopted ProteinMCP-style orchestration. Each design tool is an MCP server. Agent picks tools dynamically based on task context.

mBER VHH Design

mBER deployed on Modal A100-80GB. AF-Multimer hallucination + ESM2/AbLang2 sequence priors. 45% target-level success rate. VHH-only.

Hosted Inference API

10+ tools via REST: TAP, ImmuneBuilder, AlphaFold-Multimer, ThermoMPNN, BioPhi, Scoring system, Antibody Evolution.

Pipeline Harness

6-stage DAG orchestrator: Epitope → Backbone → Sequence → Validation → Scoring → Optimization.

Web Platform

Next.js 15 + React 19 frontend with Pipeline builder, Candidates explorer, Scoring dashboard, Mol* 3D viewer.