File size: 8,094 Bytes
3e1b7c7
6205b94
3e1b7c7
 
 
 
 
6205b94
3e1b7c7
 
 
6205b94
3e1b7c7
 
 
7fd8751
3e1b7c7
 
 
6205b94
 
 
 
 
3e1b7c7
6205b94
 
 
 
3e1b7c7
 
 
 
 
6205b94
 
3e1b7c7
 
 
 
7fd8751
 
 
3e1b7c7
 
6205b94
 
3e1b7c7
 
6205b94
3e1b7c7
 
 
 
 
 
6205b94
 
3e1b7c7
 
6205b94
3e1b7c7
 
6205b94
3e1b7c7
 
 
 
6205b94
3e1b7c7
7fd8751
 
3e1b7c7
 
7fd8751
6205b94
3e1b7c7
 
6205b94
3e1b7c7
 
 
 
6205b94
3e1b7c7
 
 
6205b94
3e1b7c7
 
6205b94
 
3e1b7c7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6205b94
 
 
 
 
 
3e1b7c7
6205b94
3e1b7c7
 
 
6205b94
3e1b7c7
6205b94
3e1b7c7
6205b94
3e1b7c7
 
6205b94
 
 
 
 
 
 
3e1b7c7
6205b94
 
3e1b7c7
 
 
6205b94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3e1b7c7
6205b94
3e1b7c7
6205b94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
"""Boltz-2 structure verification client (Phase B).

The HF Space leaderboard runs on cpu-basic, so it cannot host Boltz
directly. This module is a thin HTTP client that POSTs design sequences
to a Modal-deployed companion app (`modal_boltz_app.py`), which
provisions an A10G on demand, runs `boltz predict`, and returns
confidence metrics.

Two prediction modes (selected automatically by `run_boltz_posteval`):
  - Monomer (non-binding tasks)   -> pLDDT, pTM
  - Complex (binding tasks)       -> pLDDT, pTM, ipTM, i_pAE

Required HF Space secrets (set out-of-band via the leaderboard admin):
  MODAL_BOLTZ_URL    https://<workspace>--bdb-boltz-predict.modal.run
  MODAL_BOLTZ_TOKEN  shared bearer token matching the modal secret TOKEN

If `MODAL_BOLTZ_URL` is unset the predictors return a structured
failure dict with `success=False` and an actionable error message
rather than crashing the dispatcher.
"""

from __future__ import annotations

import logging
import os
from typing import Any

logger = logging.getLogger(__name__)

# Batch sizes large enough to amortize Modal cold-start, small enough
# to stay under the 1700s function timeout.
MONOMER_CHUNK_SIZE = 20
COMPLEX_CHUNK_SIZE = 10
HTTP_TIMEOUT_SEC = 1700


_NOT_CONFIGURED = (
    "Modal Boltz endpoint not configured. Set MODAL_BOLTZ_URL (and "
    "MODAL_BOLTZ_TOKEN) on the HF Space, or deploy the companion app "
    "with `modal deploy modal_boltz_app.py`."
)


def _modal_url() -> str | None:
    return os.environ.get("MODAL_BOLTZ_URL", "").strip() or None


def _modal_token() -> str:
    return os.environ.get("MODAL_BOLTZ_TOKEN", "").strip()


def _failure(error: str, complex_keys: bool = False) -> dict[str, Any]:
    out = {"pLDDT": 0.0, "pTM": 0.0, "success": False, "error": error}
    if complex_keys:
        out.update({"ipTM": 0.0, "i_pAE": 0.0})
    return out


def _post_predictions(items: list[dict[str, Any]]) -> dict[str, dict[str, Any]]:
    """POST a list of prediction items to the Modal endpoint.

    Returns a dict mapping each item's `name` to a metric dict, with
    structured failure entries on error.
    """
    url = _modal_url()
    if not url:
        return {item["name"]: _failure(_NOT_CONFIGURED) for item in items}

    try:
        import httpx
    except ImportError:
        return {
            item["name"]: _failure("httpx not installed in leaderboard image")
            for item in items
        }

    headers = {"Content-Type": "application/json"}
    payload = {"token": _modal_token(), "items": items}

    try:
        resp = httpx.post(
            url, json=payload, headers=headers, timeout=HTTP_TIMEOUT_SEC,
        )
    except Exception as e:
        return {item["name"]: _failure(f"Modal POST failed: {e}") for item in items}

    if resp.status_code != 200:
        return {
            item["name"]: _failure(f"Modal HTTP {resp.status_code}: {resp.text[:200]}")
            for item in items
        }

    try:
        body = resp.json()
    except Exception as e:
        return {item["name"]: _failure(f"Modal returned non-JSON: {e}") for item in items}

    if "error" in body:
        msg = body["error"]
        return {item["name"]: _failure(f"Modal: {msg}") for item in items}

    results = body.get("results", {})
    out: dict[str, dict[str, Any]] = {}
    for item in items:
        name = item["name"]
        out[name] = results.get(name) or _failure(
            "Modal returned no result for this item"
        )
    return out


def predict_monomer_batch(sequences: list[str]) -> list[dict[str, float]]:
    """Predict structures for a batch of monomer sequences."""
    items = [
        {"name": f"mono_{i}", "kind": "monomer", "sequences": [seq]}
        for i, seq in enumerate(sequences[:MONOMER_CHUNK_SIZE])
    ]
    by_name = _post_predictions(items)
    return [by_name[item["name"]] for item in items]


def predict_complex_batch(
    pairs: list[tuple[str, str]],
) -> list[dict[str, float]]:
    """Predict structures for a batch of (binder, target) pairs."""
    items = [
        {"name": f"cmplx_{i}", "kind": "complex", "sequences": [b, t]}
        for i, (b, t) in enumerate(pairs[:COMPLEX_CHUNK_SIZE])
    ]
    by_name = _post_predictions(items)
    return [by_name[item["name"]] for item in items]


def run_boltz_posteval(
    per_task_results: dict[str, dict[str, Any]],
    progress_callback=None,
) -> dict[str, dict[str, Any]]:
    """Run Boltz post-assessment on every task that needs it.

    For each successful task:
      - Non-binding: pick the first design -> monomer prediction
      - Binding: pick the first design + target sequence -> complex prediction
      - Merge Boltz metrics into existing results
      - Re-score the quality component
    """
    from eval_scorer import _is_binding_task

    monomer_tasks: list[tuple[str, str]] = []
    complex_tasks: list[tuple[str, str, str]] = []

    for task_id, result in per_task_results.items():
        if not result.get("success") or not result.get("quality_pending"):
            continue
        sequences = result.get("sequences", [])
        if not sequences:
            continue
        best_seq = sequences[0]

        if _is_binding_task(task_id):
            target_seq = (
                result.get("ground_truth_thresholds", {}).get("target_sequence")
            )
            if target_seq:
                complex_tasks.append((task_id, best_seq, target_seq))
            else:
                monomer_tasks.append((task_id, best_seq))
        else:
            monomer_tasks.append((task_id, best_seq))

    total = len(monomer_tasks) + len(complex_tasks)
    done = 0

    for chunk_start in range(0, len(monomer_tasks), MONOMER_CHUNK_SIZE):
        chunk = monomer_tasks[chunk_start:chunk_start + MONOMER_CHUNK_SIZE]
        seqs = [seq for _, seq in chunk]
        boltz_results = predict_monomer_batch(seqs)
        for (task_id, _), metrics in zip(chunk, boltz_results):
            if metrics.get("success"):
                _merge_boltz_metrics(per_task_results[task_id], metrics)
            done += 1
            if progress_callback:
                progress_callback(task_id, done, total, metrics)

    for chunk_start in range(0, len(complex_tasks), COMPLEX_CHUNK_SIZE):
        chunk = complex_tasks[chunk_start:chunk_start + COMPLEX_CHUNK_SIZE]
        pairs = [(binder, target) for _, binder, target in chunk]
        boltz_results = predict_complex_batch(pairs)
        for (task_id, _, _), metrics in zip(chunk, boltz_results):
            if metrics.get("success"):
                _merge_boltz_metrics(per_task_results[task_id], metrics)
            done += 1
            if progress_callback:
                progress_callback(task_id, done, total, metrics)

    return per_task_results


def _merge_boltz_metrics(
    task_result: dict[str, Any],
    boltz_metrics: dict[str, Any],
) -> None:
    """Merge Boltz prediction metrics into a task result and re-score quality."""
    from eval_scorer import apply_design_gate, score_quality

    merged_metrics = task_result.get("agent_metrics", {}).copy()
    for key in ("pLDDT", "pTM", "ipTM", "i_pAE"):
        if key in boltz_metrics and boltz_metrics[key] > 0:
            merged_metrics[key] = boltz_metrics[key]

    quality_result = score_quality(
        agent_metrics=merged_metrics,
        thresholds=task_result.get("ground_truth_thresholds", {}),
        task_id=task_result.get("task_id", ""),
        designs=task_result.get("sequences"),
        oracle_sequences=task_result.get("oracle_sequences"),
    )

    task_result["boltz_metrics"] = boltz_metrics
    task_result["quality_pending"] = False

    if "cpu_scores" in task_result:
        task_result["cpu_scores"]["quality"] = quality_result["score"]
        component_scores = dict(task_result["cpu_scores"])
        gated = apply_design_gate(component_scores, task_result.get("num_designs", 0))
        task_result["final_scores"] = gated
        task_result["total_score"] = sum(gated.values())

    if "cpu_details" in task_result:
        task_result["cpu_details"]["quality"] = quality_result