RFT1 side-chain orientation — resolving a conflict between two structural analyses

Why

Two OpenScientist runs on the same AlphaFold model (AF-Q96AA3-F1) disagreed about E298:

RFT1-hypotheses/scramblase-vs-binding-cavity classified it "neither cavity nor portal;
exposed membrane-interface", with side-chain exposure +0.44 (opening outward) and radial
distance 13.8 Å.
RFT1-hypotheses/family-constraint-cavity-vs-portal listed it among the most constrained
central-cavity-lining positions (KL 3.31).

Both derived the membrane normal themselves, by different methods, so the disagreement is about
the axis estimate rather than the coordinates. Since one of these calls had already been written
into the gene notes, it needed settling before either was relied on further.

Method

axis_orientation.py — deliberately minimal. No cavity detection, no conservation, no docking.
The one question asked is whether a side chain points toward or away from the membrane-normal
axis, and how stable that answer is when the axis is perturbed.

Membrane normal = principal eigenvector of sum(v v^T) over local helix direction vectors
(Cα i→i+4 spans of 5–7 Å). The outer product is sign-invariant, so helices running in opposite
directions through the membrane reinforce rather than cancel.
Centre = centroid of all Cα. TM slab = |z| <= 15 Å along the normal (320 of 541 residues).
Per residue: dRad = r(side-chain centroid) − r(Cα), negative meaning the side chain turns
toward the axis.
Stability: the sign is recomputed under 24 perturbed axes (±10° and ±20° tilts × 12 azimuths).

uv run --no-project --with numpy python axis_orientation.py

Result

Median dRad across the TM slab is +0.31 Å — most side chains point outward, as expected for
a transmembrane bundle. Against that background, every residue tested points inward:

residue	pLDDT	z (Å)	r_Cα (Å)	dRad (Å)	call	axes agreeing
E64	94.4	+5.3	16.2	−2.57	inward	24/24
R67	95.2	+9.4	17.2	−1.54	inward	24/24
K152	91.6	+6.4	16.4	−3.37	inward	24/24
Q186	98.2	−11.0	16.8	−1.66	inward	24/24
N283	92.6	−7.9	2.7	−1.68	inward	18/24
R290	87.2	+1.4	6.0	−2.63	inward	24/24
E298	88.8	+12.2	7.0	−2.49	inward	24/24
Y378	97.4	−5.5	14.3	−2.45	inward	24/24
N435	93.0	−5.2	14.1	−2.34	inward	24/24

E298 points inward, not outward, at 7.0 Å from the axis — among the closest of the set, and
stable under every perturbation tried. The family-constraint run's classification is
corroborated; the scramblase-vs-binding-cavity run's "exposed, outward-facing" call for E298 is
not reproducible here and should not be relied on.

A secondary observation, not something either run stated: R67 is inward-facing but peripheral
(r_Cα 17.2 Å, against 7.0 Å for E298 and 2.7 Å for N283). That is consistent with R67 being absent
from the family-constraint run's top-constrained cavity list despite being an invariant arginine.
Inward-facing and cavity-lining are not the same thing, and the first run's r = 9.1 Å for R67 is
not reproduced either.

Limitations

One static AlphaFold model, apparently an occluded state. A different conformation could move
these side chains.
The membrane normal is estimated from helix geometry, not from an OPM/PPM membrane fit. The
perturbation test bounds sensitivity to that choice but does not remove it.
dRad measures orientation only. It does not establish that a residue lines a contiguous solvent
cavity, and says nothing about substrate contact.
Absolute radial distances depend on the centre and axis estimates, so they are comparable within
this table but not directly against either run's numbers.

Reproducing

The AlphaFold model is not committed; axis_orientation.py downloads
AF-Q96AA3-F1-model_v6.cif from the AlphaFold DB on first run.