Repository navigation
Conversation
Non-sample nodes that are in no edge can't be dated. They now keep their input time, have NaN posterior mean/variance in node metadata and in variational_gamma's fit.node_posteriors(), and no longer affect the dating of other nodes (mutational_area ignores edgeless nodes when building epochs). Isolated samples still raise an error. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
|
Sorry for delay here, I'll get to a review after the 13th |
|
Thanks. I don't think there's anything controversial here, but it would be good to have an extra set of eyes on it. |
|
We need some equivalence testing here, I think ("are estimated node times the same for the same ARG with and without isolated nodes"). Something like: import msprime
import numpy as np
import pytest
import tskit
import tsdate
def insert_isolated_nodes(ts, times, before):
"""
Return (tables, new_ids, old_to_new) where isolated non-sample nodes with the
given times are inserted into the node table immediately before the (original)
node ids in `before`.
"""
tables = ts.dump_tables()
old_nodes = tables.nodes.copy()
tables.nodes.clear()
insert = {}
for b, t in zip(before, times):
insert.setdefault(b, []).append(t)
old_to_new = np.full(ts.num_nodes, tskit.NULL, dtype=np.int32)
new_ids = []
for i in range(ts.num_nodes + 1):
for t in insert.get(i, []):
new_ids.append(tables.nodes.add_row(flags=0, time=t))
if i < ts.num_nodes:
old_to_new[i] = tables.nodes.append(old_nodes[i])
tables.edges.parent = old_to_new[tables.edges.parent]
tables.edges.child = old_to_new[tables.edges.child]
tables.mutations.node = old_to_new[tables.mutations.node]
return tables, np.array(new_ids), old_to_new
METHODS = {
"variational_gamma": lambda ts, **kw: tsdate.variational_gamma(
ts, mutation_rate=1e-8, **kw
),
"inside_outside": lambda ts, **kw: tsdate.inside_outside(
ts, mutation_rate=1e-8, population_size=1e4, **kw
),
"maximization": lambda ts, **kw: tsdate.maximization(
ts, mutation_rate=1e-8, population_size=1e4, **kw
),
}
# (times, position in node table as fraction of num_nodes; 1.0 == append)
PLACEMENTS = {
"append": ([5.5], [1.0]),
"prepend": ([5.5], [0.0]),
"middle_multiple": ([0.0, 7.0, 1e7], [0.5, 0.5, 0.25]),
"above_root": ([1e9], [0.3]),
"many": (list(np.linspace(0, 5e4, 50)), list(np.linspace(0, 1, 50))),
}
class TestIsolatedEquivalence:
@staticmethod
def simulate(num_samples, seed):
ts = msprime.sim_ancestry(
num_samples,
sequence_length=2e5,
recombination_rate=1e-8,
population_size=1e4,
random_seed=seed,
)
return msprime.sim_mutations(ts, rate=1e-8, random_seed=seed)
@pytest.mark.parametrize("method", METHODS)
@pytest.mark.parametrize("placement", PLACEMENTS)
@pytest.mark.parametrize("extra", [{}, {"rescaling_intervals": 5}])
def test_dating_is_equivalent(self, method, placement, extra):
if extra and method != "variational_gamma":
pytest.skip("rescaling only for variational_gamma")
ts = self.simulate(10, seed=2)
times, frac = PLACEMENTS[placement]
before = [int(f * ts.num_nodes) for f in frac]
tables, iso, old_to_new = insert_isolated_nodes(ts, times, before)
ts_iso = tables.tree_sequence()
func = METHODS[method]
kw = dict(extra)
if method == "variational_gamma":
dated, fit = func(ts, return_fit=True, **kw)
dated_iso, fit_iso = func(ts_iso, return_fit=True, **kw)
else:
dated, dated_iso = func(ts, **kw), func(ts_iso, **kw)
# isolated nodes are untouched
assert np.array_equal(dated_iso.nodes_time[iso], ts_iso.nodes_time[iso])
assert np.array_equal(np.sort(dated_iso.nodes_time[iso]), np.sort(times))
assert not np.any(dated_iso.nodes_flags[iso] & tskit.NODE_IS_SAMPLE)
# everything else is identical
assert np.allclose(dated.nodes_time, dated_iso.nodes_time[old_to_new])
assert np.allclose(dated.mutations_time, dated_iso.mutations_time)
assert dated_iso.num_edges == dated.num_edges
if method != "maximization": # maximization sets no posterior metadata
mn = np.array([n.metadata["mn"] for n in dated_iso.nodes()])
vr = np.array([n.metadata["vr"] for n in dated_iso.nodes()])
mn0 = np.array([n.metadata["mn"] for n in dated.nodes()])
vr0 = np.array([n.metadata["vr"] for n in dated.nodes()])
assert np.all(np.isnan(mn[iso])) and np.all(np.isnan(vr[iso]))
assert np.allclose(mn[old_to_new], mn0)
assert np.allclose(vr[old_to_new], vr0)
if method == "variational_gamma":
mm = [m.metadata["mn"] for m in dated_iso.mutations()]
mm0 = [m.metadata["mn"] for m in dated.mutations()]
assert np.allclose(mm, mm0)
p, p0 = fit_iso.node_posteriors(), fit.node_posteriors()
for field in ("mean", "variance"):
assert np.all(np.isnan(p[field][iso]))
assert np.allclose(p[field][old_to_new], p0[field])
@pytest.mark.parametrize("method", METHODS)
def test_mutation_on_isolated_node(self, method):
ts = self.simulate(10, seed=3)
tables, iso, old_to_new = insert_isolated_nodes(ts, [5.0], [ts.num_nodes])
# new site with a single mutation above the isolated node
site = tables.sites.add_row(
position=ts.sequence_length - 0.5, ancestral_state="A"
)
tables.mutations.add_row(site=site, node=iso[0], derived_state="T")
tables.sort()
ts_iso = tables.tree_sequence()
func = METHODS[method]
dated, dated_iso = func(ts), func(ts_iso)
assert np.allclose(dated.nodes_time, dated_iso.nodes_time[old_to_new])
# mutation on the isolated node is dated at the node's (fixed) time
assert dated_iso.mutations_time[-1] == 5.0 |
nspope
left a comment
There was a problem hiding this comment.
Looks good materially, but I think we need a bit more testing and also want to make sure isolated samples can't creep through to the discrete-time methods.
| - Isolated non-sample nodes (those in no edges) are now allowed by all dating | ||
| methods. They cannot be dated, so keep their input times (with NaN | ||
| posterior mean and variance in their metadata) and do not affect the dating of | ||
| other nodes. Isolated samples still raise an error. |
There was a problem hiding this comment.
I don't think isolated samples raise an error for the discrete-grid methods? In fact, they will change the results because they alter the prior. I think we should disallow across the board.
Here's a sketch of a test:
import msprime
import numpy as np
import pytest
import tskit
import tsdate
METHODS = {
"variational_gamma": lambda ts: tsdate.variational_gamma(ts, mutation_rate=1e-8),
"inside_outside": lambda ts: tsdate.inside_outside(
ts, mutation_rate=1e-8, population_size=1e4
),
"maximization": lambda ts: tsdate.maximization(
ts, mutation_rate=1e-8, population_size=1e4
),
}
def simulate():
ts = msprime.sim_ancestry(
10,
sequence_length=2e5,
recombination_rate=1e-8,
population_size=1e4,
random_seed=1,
)
return msprime.sim_mutations(ts, rate=1e-8, random_seed=1)
def with_isolated_sample(ts, time=0.0):
tables = ts.dump_tables()
tables.nodes.add_row(flags=tskit.NODE_IS_SAMPLE, time=time)
return tables.tree_sequence()
class TestIsolatedSamplesProhibited:
@pytest.mark.parametrize("method", METHODS)
def test_isolated_sample_raises(self, method):
ts_iso = with_isolated_sample(simulate())
with pytest.raises(ValueError, match="disconnected"):
METHODS[method](ts_iso)
Non-sample nodes that are in no edge can't be dated, but previously caused an error (
variational_gamma: "disconnected nodes";inside_outside/maximization: a misleading "not simplified" error).Now, for all methods:
mn/vrnode metadata is NaN, as are their rows invariational_gamma'sfit.node_posteriors()(inside_outsidealready gives NaN for fixed nodes).rescaling.mutational_area, which built time epochs from all nodes, so an isolated node with a large time distorted the rescaling.Tests:
TestIsolatedNodesintests/test_inference.py(all three methods, several isolated times; checks other node times are unchanged vs. the tree sequence without the node). Docs and changelog updated. Full suite passes; timing on 500 samples x 10 Mb is unchanged.