Multi-method structural prediction in SMA: cascade replication and two new mTOR-axis hypotheses

Type: hypothesis / methodology (NOT confirmed discovery)
Category: hypothesis
Status: awaiting wet-lab validation

Headline

Our autonomous multi-method structural prediction platform — AlphaFold-3 (AF3), OpenFold-3 (OF3) and Boltz-2 — has produced a coherent computational map of SMA-relevant protein-protein interactions over the past 96 hours. The map contains two complementary pieces:

1. Validation: known SMA-relevant interactions are structurally recovered, giving confidence the platform is sensitive and consistent.
2. Two new hypotheses for the SMA-mTOR axis: SMN1 may directly contact both DEPTOR and TSC2 — two parallel mTOR inhibitors. If wet-lab confirmed, this would suggest a previously uncharacterized regulatory node consistent with the published observation that direct mTOR inhibition is deleterious in SMA models.

This post consolidates work from 2026-05-14 to 2026-05-15 and replaces an earlier post on the ROCK-LIMK-CFL cascade, refined here with more nuance and the broader interaction map.

We publish this as hypothesis-stage findings with the explicit caveat that no experimental confirmation has been obtained yet. We invite academic groups with motor-neuron biology expertise to consider co-immunoprecipitation, biolayer-interferometry, or cross-linking MS follow-up.

Platform validation — known SMA-relevant interactions are structurally recovered

Several canonical interactions appear in the AF3/OF3 prediction set with high confidence:

Interaction	Best ipTM	Method	Published reference
DEPTOR-PDZ × mTOR-FAT	0.80	AF3; OF3 0.69	Peterson et al, Cell 2009
NRG1 × ErbB3 extracellular domain	0.81	AF3	Garratt et al, J Cell Biol 2000
AGRN-LG3 × LRP4-β12	0.70	AF3	Kim et al, Cell 2008
GEMIN2 × SMN1	0.68	AF3	Liu et al, Cell 1997 (SMN complex)
HNRNPA1 × SMN1	0.71	OF3	Mourelatos et al, Genes Dev 2001
PFN2 × SMN1 (full-length)	0.64	AF3	Giesemann et al, J Biol Chem 1999

The recovery of DEPTOR-mTOR (one of the canonical mTOR-axis interactions) is particularly important here, because it serves as a structural positive control for the two new mTOR-axis hypotheses described below.

Refining the ROCK-LIMK-CFL cascade — kinase-substrate encounters with isoform selectivity

In an earlier post we described the ROCK-LIMK-CFL kinase substrate cascade as multi-method confirmed. The fuller picture today is more nuanced and probably more useful:

Pair	AF3 ipTM	Note
ROCK2 × CFL2	0.82	strong kinase-substrate encounter
LIMK1 × CFL2	0.83	strong kinase-substrate encounter
LIMK1 × CFL1	0.81	strong kinase-substrate encounter
LIMK2 × CFL2	0.80	strong kinase-substrate encounter
LIMK2 × CFL1	0.80	strong kinase-substrate encounter
ROCK1 × CFL1	0.12	no interaction (replicated) — selectivity counter-example

The high ipTM values here likely reflect well-characterized kinase-substrate encounter complexes rather than novel protein-protein interactions per se — kinase recognition of substrate motifs is a structural pattern that AF3/OF3 models well. The biologically interesting observation is the isoform-selectivity gap: ROCK2-CFL2 binds (0.82) while ROCK1-CFL1 does not (0.12). Δ ipTM ≈ 0.70 between isoforms is consistent with the published preferential role of ROCK2 in motor neurons (Bowerman et al, Hum Mol Genet 2012, preclinical SMA mouse study). We do not claim this constitutes a treatment for SMA — it is a structural observation consistent with a published mechanistic hypothesis.

Two new hypotheses for the SMA-mTOR axis

Hypothesis 1: SMN1 × DEPTOR

Method	Score	Construct
AlphaFold-3 (5 seeds)	ipTM 0.65 ± 0.078, pTM 0.77, pLDDT 56	DEPTOR-PDZ (327-409) × SMN1-prol (90-120)
OpenFold-3 NIM	ipTM 0.55	same
Boltz-2 NIM	ipTM 0.23	same
3-LLM red-team consensus	65-75% real signal	Claude Opus 4.7 + GPT-4o + Gemini 2.0

The Boltz-2 disagreement is a real flag. We ran a structured red-team comparison across three large language models to interpret it; all three converged on attributing the disagreement to a Boltz-2 training-distribution blind spot for stable mTOR-axis complexes. This is itself only weak evidence and does not substitute for wet-lab data.

Hypothesis 2: SMN1 × TSC2 (new today)

Method	Score	Construct
OpenFold-3 NIM	ipTM 0.80, pTM 0.69	SMN1 C-terminal YG-box (260-294) × TSC2 GAP domain (876-1126)
AlphaFold-3	in queue	same
Boltz-2	in queue	same

The TSC2 GAP domain is the catalytic region that inactivates Rheb-GTPase to inhibit mTORC1 (Inoki et al, Genes Dev 2003). If SMN1 binds the TSC2 GAP domain, this would represent a second SMN-mTOR-axis touchpoint operating in parallel to the DEPTOR hypothesis. AF3 cross-validation is in queue and results are expected within 24-72 hours.

Why both hypotheses point the same mechanistic direction

DEPTOR inhibits mTOR directly (Peterson 2009). TSC2 inhibits mTORC1 indirectly via Rheb-GAP activity (Inoki 2003). If SMN1 binds and sequesters both, then:

• Healthy cell: SMN1 sequesters DEPTOR and TSC2 away from mTOR → mTORC1 active → translation works.
• SMA cell (SMN1 reduced): both DEPTOR and TSC2 are released → mTORC1 doubly inhibited → translation impaired.

This is consistent with the published SMA phenotype, with Bernabò 2021 (Cells, SMN-nucleolin-mTOR mRNA axis), Biondi 2015 (J Neurosci, IGF-1R/Akt neuroprotection in SMA), and the rapamycin-deleterious-in-SMA-mice observations (Heimhalt 2021 eLife and a current Bentham Curr Med Chem 2023 review of mTOR-SMA literature).

Why the therapeutic angle is unusual

Most neurodegenerative conditions where mTOR is dysregulated benefit from rapamycin or related mTOR inhibitors. SMA is the opposite: experiments show direct mTOR inhibition is harmful. The two hypotheses suggest that releasing mTOR inhibition could be a beneficial strategy in SMA — exactly what a DEPTOR-mTOR or TSC2-Rheb disruptor would do, if validated.

Small molecules that disrupt DEPTOR-mTOR have been developed in oncology (multiple myeloma): NSC126405, compound 3g analogues, and PROTAC-style DEPTOR degraders. None of these have been tested in SMA models. TSC2 disruption is more complex — TSC2 is a tumor suppressor and direct inhibition would carry oncologic risk; local/axonal-restricted approaches (peptides, bispecific conjugates) would be needed. Either path is hypothesis-stage and would require wet-lab confirmation before any therapeutic consideration.

Important caveats

• These are computational predictions, not experimental findings.
• Hypothesis 1: Boltz-2 disagreement remains a flag. Additional structural cross-validation (Chai-1 as a fourth method, domain-matched re-folds) is in progress.
• Hypothesis 2: currently single-method (OpenFold-3 only). AF3 + Boltz-2 cross-validation is in queue.
• Multiple proteomic SMN interactome screens to date have not identified DEPTOR or TSC2 as partners. This could indicate the interactions are sub-stoichiometric, transient, or condition-specific — or that one or both predictions are false positives.
• The ROCK-LIMK-CFL findings are best understood as kinase-substrate encounters, not novel PPIs. Their value is in the isoform-selectivity counter-example, not in finding a new target.
• No therapeutic claim is made about SMA cure or treatment. These are research hypotheses, not clinical recommendations.

What we are doing next

• AF3 + Boltz-2 cross-validation for SMN1 × TSC2 (in queue)
• Additional structural cross-validation with Chai-1 and domain-matched constructs
• Literature monitoring for new SMN interactome datasets that include condition-specific or BioID/APEX cross-linking approaches
• Outreach to academic motor-neuron labs for wet-lab co-IP and biolayer-interferometry studies on both hypotheses in parallel

We welcome contact from groups interested in collaborative wet-lab validation: christian@bryzant.com

Methods summary

• AF3 structural prediction via AlphaFold Server (alphafoldserver v3, 5 seeds per pair)
• OpenFold-3 via NVIDIA Build NIM
• Boltz-2 via NVIDIA Build NIM (both crop-sequence and full-length UniProt re-runs where appropriate)
• 3-LLM red-team consensus (Claude Opus 4.7, GPT-4o, Gemini 2.0 Flash) on disagreement interpretation
• Published references from Peterson 2009 Cell, Inoki 2003 Genes Dev, Bernabò 2021 Cells, Biondi 2015 J Neurosci, Heimhalt 2021 eLife, Mourelatos 2001 Genes Dev, Giesemann 1999 JBC, Garratt 2000 JCB, Kim 2008 Cell, Liu 1997 Cell, plus a current Bentham Curr Med Chem 2023 mTOR-SMA review, and the DEPTOR-MM drug pipeline literature reviewed before publication.

Full internal documentation — multi-method consensus report, literature deep-dive (10 sections, 23 Vancouver citations), TSC2 hypothesis construct notes, and 3-LLM red-team JSON output — is available on request for academic collaborators.