Skip to content

SonoForge — Technical Whitepaper

A closed-loop DBTL optimization backbone for de novo design of ultrasound-coupling neural biomolecules

Dr. Sanjay Anbu · Working draft

1. Motivation

Merge Labs and adjacent efforts are building non-invasive, molecular brain–computer interfaces: instead of implanting electrodes, they express genetically-encoded proteins that let neurons exchange information with focused ultrasound through intact skull and tissue. The two enabling protein classes are:

  1. Acoustic biosensors (READ). Gas vesicles (GVs) are hollow, gas-filled protein nanostructures, ~50–250 nm, whose stiff shells scatter ultrasound and produce nonlinear contrast. Expressed from acoustic reporter genes (ARGs), they turn gene expression and, increasingly, neural activity into an ultrasound-visible signal (Shapiro et al., Nat. Chem. Biol. 2014; Farhadi et al., Science 2019).
  2. Sonogenetic actuators (WRITE). Mechanosensitive membrane channels (e.g. MscL, TRP variants) gated by ultrasound pressure enable cell-type-specific neuromodulation.

The key engineered handles on a gas vesicle are set at the level of its constituent proteins — principally the primary shell protein GvpA and the scaffolding/reinforcing protein GvpC. Varying GvpC controls the vesicle's collapse pressure (the acoustic pressure at which it irreversibly buckles), which in turn sets ultrasound contrast, harmonic behaviour, and — critically — the ability to do collapse-based acoustic multiplexing (imaging different populations by selectively collapsing them at distinct pressures) (Lakshmanan et al., ACS Nano 2016). Directed evolution of ARGs is an established DBTL campaign (ACS Synth. Biol. 2024).

The ML problem. Optimize protein sequences/structures to jointly (i) maximize ultrasound contrast, (ii) hit a target collapse pressure, (iii) remain expressible and soluble in mammalian cells, and (iv) stay below an immunogenicity ceiling — from few, noisy, expensive measurements. This is exactly the "closed-loop optimization backbone" the three Merge ML Research Scientist roles are chartered to build "from a blank slate." SonoForge is a concrete, reproducible reference for that backbone.

2. Design principles

  • Closed-loop first. The unit of value is a cycle (design → build → test → learn), not a single model. Everything is built so that a real wet-lab data stream can replace a simulated oracle with a one-line config change.
  • Physics where it's cheap and informative. First-principles molecular dynamics is used for the one property (collapse pressure) where mechanics dominate and a learned proxy would be data-starved — encoding a genuine first-principles prior rather than a black box.
  • Uncertainty is a first-class citizen. Expensive oracles + tiny budgets mean the acquisition function, not the point prediction, is what matters. We track calibration explicitly.
  • Safety is a constraint, not an afterthought. Immunogenicity enters as a hard feasibility constraint, mirroring the two dedicated immunology functions at Merge.
  • Graceful degradation. Each heavy module has a torch-free / CPU fallback so the whole loop and the test suite run on a laptop with no weights.

3. System architecture

             ┌──────────────────────── DBTL cycle k ─────────────────────────┐
 seed lib ──▶│ DESIGN                     TEST                                │──▶ archive
             │  ├ SE(3) frame flow-        ├ ESMFold self-consistency (scRMSD) │    (Pareto front,
             │  │  matching backbones      ├ contrast proxy                    │     de-immunized,
             │  └ Mamba PLM sequences ─────├ collapse-pressure  (OpenMM MD)    │     UQ-annotated)
             │                             ├ expressibility / solubility       │
             │                             └ immunogenicity (MHC-II)  ⟂ CONSTR │
             │ LEARN                                                           │
             │  └ constrained qNEHVI  ⊕  GFlowNet/RL proposal  ⊕  DPO  ⊕  UQ   │
             └───────────────────────────────────────────────────────────────┘

3.1 Representation — Mamba/S4 protein language model (plm/)

A state-space sequence model (Mamba/S4) over amino-acid tokens, pretrained on a broad protein corpus and transfer-learned onto the sparse gas-vesicle / ARG family. It serves three jobs: (a) inverse-folding-style sequence design conditioned on a backbone, (b) property-prediction heads (expressibility, solubility, contrast surrogate), and (c) learned embeddings that become the search space for Bayesian optimization. SSMs give linear-time long-context modeling and are explicitly listed as desirable in the De Novo and Biophysics roles. Fallback: a profile-HMM / ESM-2 pseudo-log-likelihood scorer.

3.2 Geometry — SE(3) frame flow-matching generator (generative/)

Conditional flow matching over residue frames (translations + SO(3) rotations), with an equivariant network as the velocity field, generalizing the Cα-only prior work to full backbone frames (à la FrameFlow / FoldFlow). Equivariance is unit-tested. A compact JAX/Flax reference of the equivariant layer ships alongside the PyTorch trainer to satisfy the "Python / PyTorch / Jax" requirement and to serve as a clean pedagogical artifact.

3.3 Physics — molecular dynamics oracle (physics/)

An OpenMM pipeline estimates mechanical observables of the GV shell — a collapse-pressure proxy from shell rigidity / buckling response and per-variant stability deltas. This is the "first-principles prior" the Computational Biophysics role centers on. MD is expensive, so it is (a) cached, (b) invoked only on the optimizer's most promising candidates, and (c) distilled into a fast surrogate over cycles. Fallback: an elastic-network / sequence-based rigidity proxy.

3.4 Oracle stack (oracle/)

A combined OracleStack returns a vector objective plus a feasibility flag per candidate: - Contrast — acoustic scattering proxy from shell geometry/stiffness (maximize). - Collapse pressure — MD-derived (target a set-point; enables multiplexing). - Expressibility & solubility — sequence/structure proxies (maximize). - Immunogenicity — MHC-II binding / T-cell epitope load (constraint: stay ≤ ceiling; drives de-immunization). All oracles are disk-cached and clearly flagged as in-silico proxies.

3.5 Learn — constrained multi-objective active learning (optimize/)

  • qNEHVI (BoTorch) for batched, noisy, constrained multi-objective Bayesian optimization over PLM embeddings, with outcome constraints for immunogenicity/tolerability.
  • GFlowNet / RL proposal to generate diverse high-reward sequences where the acquisition function is the reward — hitting the RL requirement and combating mode collapse.
  • DPO (direct preference optimization) to learn from pairwise acoustic screens ("variant A scattered more than B"), which are cheaper and lower-variance than absolute measurements — a direct fit for "preference optimization" + "sparse, noisy, high-cost data."
  • Uncertainty quantification via deep ensembles / GP posteriors, with a calibration report (coverage vs. nominal) each cycle.
  • Baselines: random search + single-objective BO + NSGA-II, for honest benchmarking.

3.6 Serve — democratization (serve/)

A typed Python API, a Gradio "design-an-ARG" app, and a Streamlit dashboard (hypervolume trajectory, interactive Pareto front, calibration curves, py3Dmol structure gallery). All three JDs require "serving models to non-domain experts for democratization."

4. Data

  • Targets & seeds: GvpA/GvpC sequences and available GV/ARG structures (RCSB PDB, UniProt), plus published ARG variant panels where available.
  • Pretraining corpus: a broad protein set for the PLM; fine-tune on the GV family (transfer learning under extreme sparsity).
  • Schema: every Candidate carries sequence, backbone, and a property record — contrast, collapse_pressure, expressibility, solubility, immunogenicity, plus provenance (cycle, parents, oracle versions). Dataset and model cards live in docs/cards/.

5. Evaluation

  1. Optimization quality — hypervolume trajectory vs. random / NSGA-II / single-objective BO across cycles; feasibility rate under the immunogenicity constraint.
  2. Sample efficiency — hypervolume vs. number of (expensive) oracle calls; the metric that matters for real wet-lab budgets.
  3. Calibration — predicted-vs-observed coverage of the surrogate's uncertainty.
  4. Structural validity — ESMFold self-consistency (scRMSD, pLDDT) of designed sequences against intended backbones.
  5. Ablations — physics-prior on/off; PLM embeddings vs. one-hot for BO; DPO vs. regression targets; GFlowNet diversity vs. BO-only.

6. Limitations

  • Oracles are in-silico proxies, not validated assays; absolute numbers are not experimental claims. The contribution is the closed-loop method and engineering, into which real data plugs directly.
  • The MD collapse-pressure estimate is a coarse-grained proxy, not a first-principles buckling simulation of a full vesicle.
  • Immunogenicity prediction is a screening filter, not a clinical immunogenicity assessment.
  • Gas-vesicle assembly is multi-protein and cooperative; single-chain structure prediction captures only part of the biology. These are stated plainly, and every simulated signal is labelled in code.

7. Roadmap

See PLAN.md for the phased build (0–6), hardware budget, and per-phase acceptance tests.

Appendix A — Role coverage matrix

Merge requirement Role(s) SonoForge component
SSMs / LLMs De Novo Mamba/S4 protein LM (plm/)
SE(3)-equivariance, flow-matching De Novo frame flow-matching generator (generative/)
PyTorch + Jax De Novo, Biophysics PyTorch trainer + JAX/Flax equivariant reference
Molecular dynamics, first-principles priors Biophysics OpenMM collapse-pressure oracle (physics/)
Protein-structure modeling Biophysics ESMFold self-consistency (oracle/)
Probabilistic modeling, UQ BayesOpt ensembles/GP + calibration report (optimize/)
Representation learning BayesOpt PLM embeddings as BO search space
Bayesian optimization, active learning BayesOpt constrained qNEHVI loop (optimize/, loop/)
Preference optimization BayesOpt DPO on pairwise acoustic screens
Reinforcement learning BayesOpt GFlowNet/RL proposal
Multi-objective / constrained all three qNEHVI + outcome constraints
Transfer learning all three pretrain → fine-tune PLM on sparse ARG data
Sparse / noisy / high-cost data all three UQ-driven active learning + DPO + oracle caching
Democratization to non-experts all three typed API + Gradio + dashboard (serve/)
Production-grade, reproducible code all three typed, tested, CI, Docker, Hydra, cards
Domain priors / wet-lab constraints all three immunogenicity constraint; expressibility/delivery fields

Appendix B — Selected references

  • Shapiro et al. Biogenic gas nanostructures as ultrasound molecular reporters. Nat. Chem. Biol. 2014.
  • Lakshmanan et al. Molecular engineering of acoustic protein nanostructures. ACS Nano 2016.
  • Farhadi et al. Ultrasound imaging of gene expression in mammalian cells. Science 2019.
  • Directed evolution of acoustic reporter genes. ACS Synth. Biol. 2024.
  • Yim et al. Fast protein backbone generation with SE(3) flow matching (FrameFlow), 2023.
  • Gu & Dao. Mamba: linear-time sequence modeling with selective state spaces, 2023.
  • Daulton et al. Parallel noisy-constrained EHVI (qNEHVI), NeurIPS 2021.
  • Bengio et al. GFlowNet foundations; Jain et al. Biological sequence design with GFlowNets, 2022.
  • Rafailov et al. Direct Preference Optimization, 2023.