Praxis — The Antibody Discovery Platform from IASO Bio

Putting structure-based generative biology into wet-lab-validated practice.

A platform whitepaper. May 2026.

0 · Executive summary

Praxis is an end-to-end platform for structure-based generative antibody and protein-binder discovery, built and operated by IASO Bio. It composes the current generation of open-source models — RFantibody, BindCraft, Chroma, ProteinMPNN, Boltz-2, AlphaFold-Multimer, ImmuneBuilder, TAP, ThermoMPNN, AggreScan3D, and a dozen more — into a single, reproducible six-stage workflow that translates a target structure and an epitope choice into wet-lab-ready candidate sequences in hours rather than months.

Three things distinguish the platform from the rapidly growing field of AI antibody startups:

A canonical workflow as the product, not a single model. Praxis is organized around the same six steps every structure-based binder design program follows — antigen → epitope → backbone → sequence → in-silico filtering → experimental validation. Every tool and page sits on this spine. Users do not assemble a pipeline; they walk one.
Honest, auditable scoring. Every metric the platform exposes has a documented range, biological meaning, calculation, reliability boundary, and ranking usage. Limitations are surfaced where they apply (the field's most-cited cutoffs are often miscalibrated for antibody–antigen interfaces). Scoring is treated as a research artifact, not a black box.
A wet-lab feedback loop, not a benchmark. Praxis runs against IASO Bio's clinical-stage BCMA program, which gave the world its first fully-human BCMA CAR-T (Fucaso / equecabtagene autoleucel, NMPA-approved 2023). Wet-lab affinity, developability, and on-cell expression data flow back into the platform's calibration layer.

The first production target is BCMA affinity maturation for in-vivo CAR-T. The platform's longer arc covers solid-tumor and autoimmune in-vivo CAR-T targets, the rare-disease antigen pipeline, and a self-hosted training stack that ingests IASO's internal sequence library.

1 · The therapeutic mission

1.1 Why in-vivo CAR-T

Ex-vivo CAR-T transformed relapsed/refractory hematologic oncology. Six approved CD19 and BCMA CAR-T products have produced overall response rates in the 80–97 % range and durable, sometimes curative responses. IASO Bio's own Fucaso, the first fully-human anti-BCMA CAR-T approved in China, sits in that lineage.

The clinical wall is not efficacy. It is manufacturing economics. Apheresis, centralized cell engineering, lymphodepletion, and a four-to-six-week vein-to-vein window keep ex-vivo CAR-T priced at several hundred thousand US dollars per patient and limited to a small number of qualified centers worldwide. The total addressable population reached today is a fraction of those who could benefit.

The frontier is in-vivo CAR-T: deliver the CAR construct as mRNA inside a lipid nanoparticle, and let the patient's own T cells transiently express it. Eliminate apheresis. Eliminate centralized cell engineering. Eliminate lymphodepletion in many indications. Convert a bespoke autologous therapy into an off-the-shelf infusion.

The field has moved from concept to clinic in under two years. Capstan Therapeutics dosed the first patient with CPTX2309 (an anti-CD19 in-vivo CAR-T for autoimmune disease) in mid-2025 and was acquired by AbbVie for USD 2.1 billion within weeks. EsoBiotec dosed the first multiple-myeloma patient with an in-vivo BCMA CAR-T (ESO-T01) in 2025 and was acquired by AstraZeneca. Umoja Biopharma's UB-VV111 entered first-in-human trials in late 2024 under an AbbVie codevelopment deal. Interius BioTherapeutics was acquired by Kite/Gilead. The field is real, well-funded, and consolidating fast.

1.2 Why in-vivo flips the binder design problem

A binder that works in ex-vivo CAR-T does not automatically work in vivo. Four constraints become first-class:

Size. The CAR cassette must fit in mRNA that an LNP can carry — roughly 335 amino acids, ~1.5 kb. Of that, the binder gets ~120 aa. An scFv binder (~250 aa) consumes budget that single-domain binders do not.
Folding without QC. Ex-vivo CAR-T has a release assay before infusion. In vivo, whatever the cell expresses lands on the surface. scFv misfolding in the CAR context is a well-documented failure mode; VHH (single-domain) and de novo mini-binders avoid the issue.
Tonic signaling. Antigen-independent CAR clustering exhausts T cells over time. The dual-VHH architecture that drove Carvykti's efficacy story also drove its FDA boxed warning for late-onset Parkinsonism and Guillain–Barré syndrome — likely through supra-physiological clustering. Higher affinity is not always better; the binder must be tuned, not maximized.
Narrow off-tumor margin. With no screening window and no patient-specific cell product, even minor cross-reactivity becomes a serious adverse event. Counter-screen specificity is non-negotiable.

These constraints make the binder design problem structurally biased toward small, well-folded, modestly-affine, surgically-specific recognition modules. Praxis is built around that bias. The default antibody scaffold is the camelid VHH (~120 aa, no light chain). The runner-up is the de novo mini-binder (60–240 aa). scFv is supported but de-emphasized.

1.3 Why BCMA is the first target

A wet-lab-validated anti-BCMA series exists from the Fucaso program. Ground-truth affinity data on real binders is the precondition for calibrating any in-silico scoring stack against reality.
Public structural context is excellent. Co-crystal structures (PDB 4ZFO, 6J7W) and a competitor patent disclosure (CN112062851B, covering the 026 scFv that became Fucaso) give the platform a clean comparison baseline.
A known competitive bar. Three approved BCMA CAR-Ts — Abecma, Carvykti, and Fucaso — span the design space: humanized mouse scFv, dual-VHH, and fully-human single-domain. Their efficacy and toxicity profiles are public. The platform knows what "good" and "too toxic" look like.
A live commercial pull, not a benchmark exercise. The next-generation in-vivo BCMA CAR-T is a credible product on a credible timeline.

Secondary targets — CD20, GPRC5D, HER2 — are seeded in the curated epitope library. A rare-disease antigen pipeline is tracked in the Research surface for the year-2-and-beyond horizon.

2 · Industry context

2.1 Three generations of antibody discovery

Praxis sits in the third generation of antibody discovery, which is in the middle of compressing decades of bench work into months of computation.

Generation 1 — Animal immunization (1975–present). Immunize a mouse, fuse a B cell with a myeloma cell, screen hybridoma supernatants. Humanize the winner if it has therapeutic potential. One to two years per target. The diversity ceiling is set by the mouse repertoire; humanization is a multi-round chore.
Generation 2 — In vitro display (1990s–present). Phage, yeast, mammalian, and ribosome display libraries break the animal-repertoire ceiling and dramatically widen the searched space. Display is still the canonical wet-lab closing step for AI-designed candidates — Praxis produces designs ready for display selection, not as a replacement for it.
Generation 3 — ML-augmented and de novo (2022–present). Sequence priors from protein language models (ProGen, IgLM, AntiBERTy, AbBFN), then full structure-based de novo design (RFdiffusion, RFantibody, BindCraft, Chroma, BoltzGen, PXDesign). This is the substrate Praxis composes.

The shift is not "AI replaces wet lab." It is "AI proposes a vastly narrower set of structurally rationalized hypotheses, which wet lab validates faster." Bennett et al. (2024 / 2025, Nature) demonstrated this end-to-end for RFantibody: structure-based design, ProteinMPNN sequencing, RoseTTAFold2 validation — confirmed by cryo-EM that the designs fold and bind exactly as predicted. The validation step is not optional, and the platform treats it as Step 6 of every workflow, not an afterthought.

2.2 The current state of generative antibody design

The field's headline number is the experimental hit rate — what fraction of designed sequences actually bind their intended target with the intended affinity.

RFantibody (Bennett et al., Nature 2024 / 2025): the first work to demonstrate atomically accurate de novo antibody design with experimental validation across four disease-relevant targets, including cryo-EM confirmation of the predicted binding pose.
BindCraft (Pacesa et al., Nature 2024 / 2025): an AlphaFold2-multimer hallucination pipeline; reported 10–100 % experimental success rates with nanomolar affinity on cell-surface receptors, common allergens, and complex multi-domain enzymes such as CRISPR-Cas9, without high-throughput screening.
Boltz-2 (MIT Jameel Clinic, 2025): the first open-weights structure-prediction model to materially outperform AlphaFold-Multimer on antibody–antigen complexes, while predicting binding affinity directly. Praxis uses Boltz-2 as the default complex scorer in Wave 1 of the scoring-audit roadmap.
Nabla Bio JAM-2 (2025): reported 100 % hit rate across 16 structurally diverse targets including GPCRs, in both VHH-Fc and full-length IgG formats. GPCRs have historically been the hardest class.
Chai-2 (2025): zero-shot binder design across scFv, VHH, and mini-binder formats; reported as the current overall benchmark leader.
Profluent ProGen3 + OpenAntibodies (April 2025): a 6.4-billion-parameter protein language model and 20 royalty-free designed antibodies against commercial drug targets.

These results are not interchangeable. Hit rate is measured against different target classes, different affinity thresholds, and different downstream developability bars. Praxis does not pick a single winner; it provides every tool listed above (and more) under one orchestrator, lets the workflow author pick the right tool per target class, and records reproducibility metadata for every run.

2.3 The commercial landscape

The AI-driven antibody-discovery services market is projected to grow from USD 1.9 B in 2025 to USD 3.5 B by 2030 (13.3 % CAGR). Funding has been intense:

Generate Biomedicines IPO'd for USD 400 M in 2024 on the strength of the Chroma platform.
Earendil raised USD 787 M in early 2026, backed by Sanofi, a Pfizer-tied biotech fund, and Hillhouse.
EvolutionaryScale closed a USD 142 M seed in 2024.
BigHat Biosciences signed a Lilly partnership in April 2025, with an equity investment.
Cradle signed a multi-year Bayer collaboration.
AbbVie's USD 2.1 B acquisition of Capstan in June 2025 set the precedent for in-vivo CAR-T as a strategic asset class.

Platform business models break into three rough buckets: (a) internal-pipeline biotechs that own their candidates (Generate, Nabla, Earendil), (b) services / partnership platforms that engineer antibodies under contract or in joint programs (BigHat, Cradle, Profluent), and (c) internal-platform-as-tooling for a parent company's own clinical pipeline.

Praxis is, today, the third — an internal platform that turns IASO Bio's BCMA CAR-T expertise and clinical-quality wet lab into a structurally-disciplined source of next-generation in-vivo candidates. The architecture is built so the platform can serve external partnerships without rewiring; the discipline of decoupling models from orchestration was a deliberate design choice.

3 · Design philosophy

Praxis is opinionated. Six choices govern almost every architectural decision.

3.1 The workflow is the product

Every page in the platform sits at a specific step of the canonical six-step workflow:

Input antigen structure — Import.
Define binding hotspot — Epitope Picker.
Generate backbones — Pipeline (default), with specialized entry points for VHH Design, Two-VHH CAR, Mini-Binder, and Programmable Design.
Design sequences — bundled into Pipeline (ProteinMPNN by default).
In-silico filtering — Interface, Candidates, Compare, with the Scoring Audit and Readout Guide as the operating manuals.
Experimental validation — off-platform; the CAR Design page packages the winning binder into a synthesis-ready cassette.

Users never construct the workflow from primitives. They walk a predetermined narrative, the same one every credible structure-based antibody design program follows. Specialization happens by binder format (single-domain vs scFv vs mini-binder), not by re-inventing the pipeline. Power users can drop into Programmable Design for custom Chroma conditioners, but the default surface is the canonical path.

This is the opposite of the "expose every model as a tile" approach. It is the architectural bet that friction in onboarding is the binding constraint on platform adoption, not feature coverage. A scientist who can run their first end-to-end workflow in twenty minutes will continue to use the platform; a scientist who has to read three papers before they can compose one will not.

3.2 Honest scoring

Every metric the platform exposes is documented along six axes, surfaced in the Readout Guide and the Scoring Audit:

Range — where good lives.
Biological meaning — what the number is measuring.
Calculation — which model produced it, with what inputs.
Reliability boundary — where the number is known to fail.
Ranking usage — whether to optimize on it, gate on it, or use it only to break ties.
Worked example — a realistic number for a realistic candidate.

This is unusually candid for the field. The literature's standard "ipTM > 0.6 acceptable, > 0.8 excellent" cutoffs are inherited from generic protein–protein interactions and systematically under-predict on antibody–antigen interfaces, where contacts are flat, polar, and induced-fit. Praxis treats this as a public fact, not a marketing problem. The Scoring Audit page is bluntly titled "is the current scoring world-class" and answers honestly: Wave 1 closed the AlphaFold-Multimer gap with Boltz-2; four of five audited dimensions still need work.

Wet-lab data from the BCMA program is the closing loop. The platform's scoring layer will be re-calibrated against IASO-internal Kd and developability measurements in Wave 2, once the first BCMA maturation cohort returns.

3.3 Composable over monolithic

Praxis is built on the Model Context Protocol (MCP) — every external model and local computation is wrapped as an MCP server with a documented input / output schema. The Agent surface (Claude-backed) consumes these servers as tools. The deterministic UI surface (Pipeline, Interface, etc.) consumes the same servers as orchestrator stages.

This matters for three reasons:

No model lock-in. When Chai-3 ships in 2026, integration is a new MCP server, not a platform rewrite. When the field consolidates around three structure predictors, the platform's compute substrate consolidates with it without breaking the UI.
The Agent and the UI share a single tool surface. A natural-language request and a button click both end up at the same scoring function. Reproducibility metadata is identical across both entry points.
External partners can integrate at the MCP layer, not the UI. Pharma partners who already run RFantibody or BindCraft locally can plug into the orchestrator without sending sequences to a hosted API.

3.4 Hosted compute first; self-host when load-bearing

GPU work today dispatches to a hosted inference API. Build velocity wins out over infrastructure ownership at the current scale (cents to a few dollars per job; tens of jobs per day). Maintenance is upstream.

The trade-off is that the platform cannot, today, fine-tune ProteinMPNN, RFantibody, or any structural model on IASO's internal sequence library. When that becomes load-bearing — Wave 3 of the scoring roadmap — the Modal-based self-host stack already wired in pipeline/modal/ graduates from skeleton to production.

This is a deliberate sequencing. Self-hosted compute without internal training data is overhead; internal training data without self-hosted compute is unactionable. Praxis builds them in the order that makes each useful.

3.5 English-first, bilingual everywhere

Every user-facing surface — page copy, tutorials, the whitepaper, the scoring guide, sidebar tooltips, in-page InfoTips — ships English canonical plus Mandarin translation. This is a deliberate choice given the parent company's market (China) and the language of the literature the platform composes (English).

The bilingual convention is EN-first, not parallel. New copy is drafted in English; Mandarin is the translation. Drift between the two is treated as a translation defect, not a content fork.

3.6 Wet-lab grounded

The platform is built against IASO Bio's BCMA program, not a benchmark dataset. Every claim about scoring reliability is calibrated — or will be calibrated, in the next BCMA maturation closing loop — against measured Kd, on-cell expression, and tonic-signaling readouts on real binders the company has synthesized and characterized.

This is the disciplining force that keeps the platform from drifting into pure-benchmark territory. The published-paper hit rates are interesting; the wet-lab hit rate on IASO's targets is what matters.

4 · Platform capabilities

This section walks the canonical six steps with the concrete features that implement each step. Detailed per-feature walkthroughs live at /docs/tutorials.

4.1 Step 1 — Input antigen structure

/import brings external structures and sequences into the platform database. Single PDB or batched. Source-tagged (BoltzGen / mBER / RFantibody / BindCraft / Chroma / External) so downstream provenance is preserved. Sequence is extracted automatically from the CA backbone; the new candidate appears in /candidates immediately.

For targets without a deposited structure, /import accepts UniProt IDs and resolves to the corresponding AlphaFold prediction. For competitor or external-collaborator binders (e.g. CT103a, Cilta-cel), /import is the entry point — once ingested, they participate in the same scoring and comparison pipeline as platform-designed candidates.

4.2 Step 2 — Define the binding hotspot

/epitope-picker is the structural interface for epitope selection. Curated presets for BCMA, CD20, HER2, and GPRC5D produce the literature-derived hotspot in one click; the residue list is canonicalized to chain:residue tokens (e.g. A:40,A:42,A:45) and exported to every downstream design surface.

The platform treats epitope selection as the single most consequential choice in the workflow. A misplaced hotspot turns the rest of the run into wasted compute. The 3D Mol* viewer is there to make epitope selection physically grounded, not abstract — patches scattered across opposite faces of the antigen visibly do not form a coherent binding patch, and the platform deliberately exposes that.

4.3 Step 3 — Generate backbones

/pipeline is the default orchestrator. It composes a backbone-generation model (RFantibody by default; BoltzGen, BindCraft, Chroma, mBER, or Germinal as alternatives) with the downstream stages.

The platform's specialized step-3 entry points exist for cases where the default surface is too generic:

/vhh-design — VHH-specific UI with four design modes (de novo, germline-guided, template-based, humanization), camelid IGHV family selection, and the VHH hallmark stringency gate.
/two-vhh — dual-VHH design for tandem CAR constructs (Carvykti-style), with three architectures: in-vivo LNP delivery, bispecific tandem, and hybrid (dual tumor antigen plus T-cell engager arm). The platform actively models the tonic-clustering risk this architecture amplifies.
/mini-binder — small de novo protein binders (~80–120 aa). The mRNA-LNP-compatible alternative when ~120-aa VHH framework is still too large.
/programmable-design — full Chroma conditioner system (substructure, symmetry, natural language, subsequence, shape, domain class). The power-user surface; not the place to start.

4.4 Step 4 — Sequence design

Bundled into /pipeline. ProteinMPNN designs CDR sequences compatible with the generated backbones; framework residues stay grafted from the chosen germline. CDR-H3 receives special attention because it dominates antigen contact and is the highest-leverage redesign target.

For warm-start affinity maturation — taking a wet-lab-validated parent VHH and re-designing only CDR-H3 — the platform uses partial diffusion mode, which preserves the parent backbone and only re-samples HCDR3. This is the path the BCMA program walks for its first wet-lab-driven design cohort.

4.5 Step 5 — In-silico filtering

The bulk of the platform's value compounds here. Praxis runs every designed sequence through a scoring stack that includes:

Boltz-2 ipTM, iPAE, pLDDT. The Wave-1 default complex predictor; outperforms AlphaFold-Multimer on antibody–antigen interfaces.
AlphaFold-Multimer. Retained as a secondary scorer for ensemble agreement.
ImmuneBuilder. Antibody-specific confidence score with better calibration than generic AF pLDDT on the Ig fold.
TAP (Therapeutic Antibody Profiler). Five-axis developability gate (CDR length, hydrophobic patches, charge patches, hydrophobicity, isoelectric point).
ThermoMPNN. Per-mutation ΔΔG prediction for stability screening.
AggreScan3D. Structure-based aggregation propensity.
CDR geometry validator. ANARCI / IMGT-based check that CDR lengths fall in canonical ranges.
BSA, contact pair table, paratope footprint. Per-complex geometry on every accepted design.
Tonic clustering risk proxy. A first-principles heuristic; explicitly marked v0 in the Scoring Audit, awaiting wet-lab calibration in Wave 2.
CAR-fit composite. Format-specific scoring for CAR construct viability, currently hand-tuned; the natural landing point for the first internal-data-trained predictor.

Three pages consume this scoring stack:

/candidates — the triage surface. Filter by min ipTM and max Pareto rank, sort by any column, multi-select, export FASTA / PDB / CSV.
/compare — side-by-side head-to-head with dual Mol* viewers, metric table with deltas, and sequence alignment. The page where CT103a-vs-Cilta-cel-vs-internal-designs analyses happen.
/interface — fresh-score any sequence against any target with a focused Boltz-2 or AlphaFold-Multimer run. The platform's primary on-demand scoring entry point.

4.6 Step 6 — Experimental validation prep

/car-design is the bridge to wet lab. It takes a winning binder and assembles the full CAR cassette (signal peptide + binder + hinge + transmembrane + costimulatory + CD3ζ), then codon-optimizes for mRNA-LNP delivery (CAI, GC%, LNP-fit scoring). The output is a synthesis-ready cassette sequence, plus a downloadable construct file ready for the wet-lab team's preferred synthesis vendor.

4.7 The Agent surface

/agent is the natural-language entry point. Claude-backed; eleven platform tools available (run_design, query_candidates, score_candidate, chroma_generate, export_candidates, propose_dag, and others). Used most often for exploratory queries ("show me BCMA candidates with ipTM > 0.7 and aggregation < 0.3"), composition tasks ("plan a maturation run as a DAG and submit it"), and onboarding (the Agent can explain platform concepts and surface the relevant tutorial).

The Agent and the deterministic UI share an identical tool surface; the difference is the planner. The Agent's plans render as in-chat DAGs the user can review before launching.

5 · Differentiation

Five characteristics, taken together, distinguish Praxis from both the open-source generative-biology stack and the closed-platform AI biotech competitors.

5.1 The workflow is end-to-end and opinionated

Most academic codebases stop at "here is a model and a notebook." Most commercial platforms expose a model marketplace — pick one, configure it. Praxis ships a single canonical workflow with specialized variants. The path from a target structure to a wet-lab-ready candidate is a guided eight-step click-through, not an integration project. The platform handles file format conversions, units, score normalization, run provenance, and downstream linkage automatically.

5.2 The wet-lab loop is real, not future

IASO Bio operates an active CAR-T clinical-stage wet lab. The platform's BCMA cohort, scheduled for synthesis and characterization in mid-2026, will return measured affinity, expression, and tonic-signaling readouts on designed binders. Those numbers will calibrate the scoring layer in Wave 2. No other AI antibody platform we are aware of has this closing loop wired into a single organization at this clinical maturity.

5.3 Scoring is auditable, not a black box

The Scoring Audit page exists. The Readout Guide exists. Every metric is documented in plain language with its known failure modes. When a candidate scores high, the user can trace each contribution back to the model that produced it. This is unusual in a market where most platforms surface a single composite "design score" with no decomposition.

Agent-driven exploration and deterministic UI runs hit the same MCP servers. The same scoring function is called. The same provenance is recorded. A scientist who prefers buttons and a research lead who prefers natural language are working on the same substrate. This is not the case for platforms that build Agent and UI as separate stacks.

5.5 Bilingual is native, not bolted on

Every user-facing string ships in English and Mandarin. The platform's research lead, wet-lab team, and IASO collaborators operate in different language preferences without context switching. This is not a translation layer pasted on at the end; it is the convention every page and component is built against.

6 · Future development

6.1 Near-term — the next two quarters

Wave 2 scoring calibration. Once the BCMA wet-lab cohort returns (~mid-2026), use measured Kd and on-cell expression to recalibrate Boltz-2 score thresholds, Pareto weights, and the tonic-risk composite. Replace heuristic v0 labels with calibrated v1.
Mutation-resistance screen. The /mutation-resistance surface (in development) substitutes antigen residues — post-treatment BCMA escape mutations, for example — and re-scores per-mutation binding retention. The competitive-narrative angle: design binders that retain affinity where competitors lose it.
Cross-reactivity counter-screen. Already shipped as a pipeline stage; the next iteration uses ortholog panels to flag candidates with elevated off-target risk before they reach wet lab.
CAR-T format-specific scoring. A first internal-data-trained predictor for CAR-fit, leveraging Fucaso-era construct data plus the new wet-lab cohort. The first surface where Praxis stops borrowing public-model scores and starts producing IASO-proprietary ones.

6.2 Medium-term — the year ahead

Wave 3 self-host migration. ProteinMPNN, RFantibody, and the structure predictors migrate to the Modal self-host stack. This unlocks fine-tuning on IASO's internal sequence library and removes the per-job cloud-API cost ceiling.
Second and third therapeutic targets. CD20 and GPRC5D are the obvious follow-ons (hematologic oncology, BCMA-program-adjacent expertise). A rare-disease antigen pipeline is tracked in the Research surface and will be triaged by feasibility (structural data quality, public-patent landscape, clinical pull) as the platform matures.
External partnership surface. The MCP architecture decouples external partner integrations from the UI. A pharma partner running RFantibody locally can plug into the platform's orchestrator without sending sequences to a hosted API; the same applies to a partner running their own scoring stack.

6.3 Long-term — the platform thesis

The long arc is to make in-vivo CAR-T discovery a first-class computational discipline. Today every approved CAR-T derives from a binder discovered in the 2010s by either an animal-immunization program (Carvykti VHHs, Abecma scFv) or a small-scale phage-display campaign (Fucaso 026). The next decade's approved CAR-Ts should derive from binders generated, scored, and triaged on platforms like Praxis — with a six-month design-to-IND timeline rather than a six-year one.

The constraint that gates this is not generative capability. The capability is already approximately good enough; the field's hit rates against well-characterized targets are in double-digit percentages. The constraint is trustworthy scoring under in-vivo CAR-T-specific stresses — tonic signaling, expression, persistence, off-tumor margin. Praxis's investment in honest scoring and the wet-lab feedback loop is a long bet that this is where defensibility accumulates.

7 · Open questions

In keeping with the platform's honest-scoring doctrine, here is the list of things Praxis does not, today, do well. These are the items that will move from this list into the roadmap as the wet-lab loop closes.

Tonic-clustering proxy is heuristic v0. No published in-silico predictor for tonic signaling has been validated against clinical outcomes. The platform's composite (surface hydrophobicity + pH 7.4 net charge + Fv-Fv self-association docking) is a first-principles starting point only.
No metric has been calibrated against IASO-internal wet-lab data yet. The Wave-1 Boltz-2 swap closed the structural-prediction gap; calibration of cutoffs and Pareto weights against measured Kd is the Wave-2 task.
In-vivo expression prediction is not in the stack. A binder that scores well by ipTM, BSA, and TAP may still misfold or fail to surface-express in the CAR context. The first internal-data-trained predictor (Wave 3) targets this gap.
Mutation-resistance scoring is on the page but not on the calibration curve. Predicting which competitor binders lose binding on the BCMA escape-mutation panel is feasible mechanistically; whether the platform's predictions agree with what clinicians see in real escape is unknown without wet-lab data.
The current orchestrator handles several specialized workflow types as stubs. The Mini-Binder, Two-VHH, and VHH workflows POST to the workflows endpoint and receive workflow IDs, but the underlying executor for those types currently simulates rather than dispatches to the real orchestrator. The fix is well-scoped and is in the queue.

These gaps are the work. The platform's design philosophy is that surfacing them keeps the work pointed at the right thing.

8 · Appendix

8.1 Integrated models (selected)

Backbone generation: RFantibody, BoltzGen, BindCraft, Chroma, mBER, Germinal.
Sequence design: ProteinMPNN, AntiFold, IgLM, AbBFN.
Structure prediction: AlphaFold-Multimer, AlphaFold 3, Boltz-1, Boltz-2, Chai-1, ImmuneBuilder, ESMFold (fallback).
Scoring / developability: TAP, ThermoMPNN, AggreScan3D, DeepImmuno, BioPhi (humanness), ANARCI / IMGT (CDR geometry), FoldX (ΔΔG).
Optimization: partial-diffusion HCDR3 redesign, Antibody Evolution, planned contextual-bandit and GP-surrogate maturation policies.

8.2 Selected references

Bennett, N. R. et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature, 2025. (RFantibody.)
Pacesa, M. et al. One-shot design of functional protein binders with BindCraft. Nature, 2024 / 2025.
Passaro, S. et al. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv / MIT Jameel Clinic, 2025.
Ingraham, J. B. et al. Chroma: a generative model for programmable protein design. Nature, 2023.
Watson, J. L. et al. RFdiffusion: De novo design of protein structure and function. Nature, 2023.
Dauparas, J. et al. Robust deep learning–based protein sequence design using ProteinMPNN. Science, 2022.
Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 2024.
Almagro, J. C. et al. Evolution of phage display libraries for therapeutic antibody discovery. 2023 review.
Carvykti FDA prescribing information. Boxed warning on CRS, HLH/MAS, ICANS, Parkinsonism, Guillain–Barré syndrome.

8.3 Glossary

The platform ships an in-product glossary surfaced as hover-tips throughout the UI. Definitions for epitope, paratope, CDR-H3, VHH, scFv, Ig fold, ipTM, pLDDT, PAE, BSA, TAP, tonic signaling, AggreScan3D, Pareto rank, RFantibody, BindCraft, BoltzGen, Chroma, ProteinMPNN, AlphaFold-Multimer, CAR, in-vivo CAR-T, hinge, costimulatory domain, BCMA, LNP, CAI, partial diffusion, warm-start regenerate, and MCP server live in the Readout Guide and Knowledge Map.

Praxis — an antibody discovery platform by IASO Bio.

This whitepaper is a living document and will be revised as the platform's wet-lab calibration cohorts close. The version of record is whatever this page renders at the time you load it; previous revisions are recoverable from the git history in the platform repository.