File size: 16,523 Bytes
8476db7 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 | """In-process agent runner for contamination-safe BioDesignBench submissions.
Replaces the old endpoint-based dispatcher: the HF Space leaderboard
runs the agent loop here, using:
- the submitter's LLM provider + API key (passed at submission time,
never persisted -- forwarded straight to the provider SDK and
discarded after the run)
- either our reference protein-design-mcp endpoint or a custom
submitter-provided MCP URL (the latter is opt-in, off by default)
This keeps the 76 task descriptions inside Romero Lab infrastructure;
the only path data leaves the lab is via the submitter's chosen LLM
provider API call. The MCP server (custom or reference) sees only
operational tool arguments -- never the raw task prompt or evaluation
criteria.
"""
from __future__ import annotations
import json
import logging
import os
import re
import time
from pathlib import Path
from typing import Any
logger = logging.getLogger(__name__)
DEFAULT_MAX_STEPS = 50
TOOL_HTTP_TIMEOUT = 600 # per-tool-call timeout (seconds)
DEFAULT_SYSTEM_PROMPT = (
"You are an expert computational protein engineer participating in "
"the BioDesignBench evaluation. Your goal is to design protein "
"sequences that satisfy the user's task description by orchestrating "
"the available protein-design tools. Iterate: generate candidates, "
"evaluate them with multiple metrics (structure prediction, energy, "
"interface analysis), refine, and only output your final designs "
"after thorough validation. When you are done, end your final "
"message with a fasta-style block containing the designed sequences:\n"
"```fasta\n>design_1\nMKKL...\n>design_2\nMFQR...\n```"
)
# ---------------------------------------------------------------------------
# Tool-schema loading
# ---------------------------------------------------------------------------
def load_tool_schemas() -> list[dict]:
"""Load the 17 reference MCP tool schemas."""
p = Path(__file__).parent / "mcp_tool_schemas.json"
with open(p) as f:
return json.load(f)
def to_anthropic_tools(schemas: list[dict]) -> list[dict]:
"""Convert leaderboard tool schemas to Anthropic's `tools` format."""
out = []
for s in schemas:
out.append({
"name": s["name"],
"description": s.get("description", ""),
"input_schema": s.get("parameters") or s.get("input_schema") or {},
})
return out
def to_openai_tools(schemas: list[dict]) -> list[dict]:
"""Convert to OpenAI's `tools` format (also used by DeepSeek)."""
out = []
for s in schemas:
out.append({
"type": "function",
"function": {
"name": s["name"],
"description": s.get("description", ""),
"parameters": s.get("parameters") or s.get("input_schema") or {},
},
})
return out
# ---------------------------------------------------------------------------
# MCP HTTP client (one tool call per POST)
# ---------------------------------------------------------------------------
def call_mcp_tool(
mcp_url: str,
mcp_token: str,
tool_name: str,
arguments: dict[str, Any],
) -> dict[str, Any]:
"""POST a single tool call to the MCP server.
The server contract mirrors `protein-design-mcp/deploy/modal_app.py`:
POST / body: {"name": "<tool>", "arguments": {...}}
200 OK body: {...tool result...}
"""
if not mcp_url:
return {"error": "MCP endpoint not configured (PROTEIN_MCP_URL unset)"}
try:
import httpx
except ImportError:
return {"error": "httpx not installed in leaderboard image"}
headers = {"Content-Type": "application/json"}
if mcp_token:
headers["Authorization"] = f"Bearer {mcp_token}"
payload = {"name": tool_name, "arguments": arguments}
try:
resp = httpx.post(
mcp_url, json=payload, headers=headers, timeout=TOOL_HTTP_TIMEOUT,
)
except Exception as e:
return {"error": f"MCP POST failed: {e}"}
if resp.status_code != 200:
return {"error": f"MCP HTTP {resp.status_code}: {resp.text[:300]}"}
try:
return resp.json()
except Exception as e:
return {"error": f"MCP returned non-JSON: {e}"}
# ---------------------------------------------------------------------------
# Sequence extraction from agent's final answer
# ---------------------------------------------------------------------------
_FASTA_BLOCK = re.compile(r"```fasta\s*\n(.*?)\n```", re.DOTALL | re.IGNORECASE)
_FASTA_HEADER = re.compile(r"^>\S+", re.MULTILINE)
_AA_LINE = re.compile(r"^[ACDEFGHIKLMNPQRSTVWY*\-]+$", re.MULTILINE)
_AA_INLINE = re.compile(r"\b([ACDEFGHIKLMNPQRSTVWY]{20,})\b")
def extract_sequences(text: str, max_designs: int = 50) -> list[str]:
"""Pull amino acid sequences out of the agent's final assistant text.
Looks for fenced fasta blocks first; falls back to inline AA strings
of length >= 20.
"""
if not text:
return []
seqs: list[str] = []
for block in _FASTA_BLOCK.findall(text):
# Split on header lines and concatenate body lines per record
records: list[list[str]] = []
current: list[str] | None = None
for line in block.splitlines():
line = line.strip()
if not line:
continue
if line.startswith(">"):
if current is not None:
records.append(current)
current = []
continue
if current is None:
current = []
current.append(line)
if current:
records.append(current)
for rec in records:
joined = "".join(rec).replace(" ", "").replace("-", "").replace("*", "")
if len(joined) >= 20 and set(joined) <= set("ACDEFGHIKLMNPQRSTVWY"):
seqs.append(joined)
if not seqs:
for m in _AA_INLINE.finditer(text):
seqs.append(m.group(1))
# Dedup while preserving order
seen = set()
deduped = []
for s in seqs:
if s in seen:
continue
seen.add(s)
deduped.append(s)
if len(deduped) >= max_designs:
break
return deduped
# ---------------------------------------------------------------------------
# Anthropic agent loop
# ---------------------------------------------------------------------------
def _run_anthropic(
api_key: str,
model: str,
system: str,
user: str,
tool_schemas: list[dict],
mcp_url: str,
mcp_token: str,
max_steps: int,
) -> tuple[list[dict], str]:
"""Anthropic Claude tool-calling loop."""
from anthropic import Anthropic
client = Anthropic(api_key=api_key)
tools = to_anthropic_tools(tool_schemas)
messages: list[dict] = [{"role": "user", "content": user}]
run_log: list[dict] = []
final_text = ""
for step in range(max_steps):
resp = client.messages.create(
model=model,
max_tokens=8192,
system=system,
messages=messages,
tools=tools,
)
# Append assistant turn (Anthropic content blocks are passed back as-is)
messages.append(
{"role": "assistant",
"content": [block.model_dump() for block in resp.content]}
)
tool_uses = [b for b in resp.content if b.type == "tool_use"]
for b in resp.content:
if b.type == "text":
final_text = b.text
if not tool_uses or resp.stop_reason == "end_turn":
break
tool_results = []
for tu in tool_uses:
t0 = time.monotonic()
result = call_mcp_tool(mcp_url, mcp_token, tu.name, tu.input)
dt = round(time.monotonic() - t0, 2)
run_log.append({
"step": step,
"tool": tu.name,
"arguments": tu.input,
"success": "error" not in result,
"latency_sec": dt,
"result_summary": str(result)[:500],
})
tool_results.append({
"type": "tool_result",
"tool_use_id": tu.id,
"content": json.dumps(result)[:8000],
})
messages.append({"role": "user", "content": tool_results})
return run_log, final_text
# ---------------------------------------------------------------------------
# OpenAI / DeepSeek agent loop (DeepSeek uses the openai-compatible API)
# ---------------------------------------------------------------------------
def _run_openai_compat(
api_key: str,
model: str,
system: str,
user: str,
tool_schemas: list[dict],
mcp_url: str,
mcp_token: str,
max_steps: int,
base_url: str | None = None,
) -> tuple[list[dict], str]:
from openai import OpenAI
client = OpenAI(api_key=api_key, base_url=base_url) if base_url else OpenAI(api_key=api_key)
tools = to_openai_tools(tool_schemas)
messages: list[dict] = [
{"role": "system", "content": system},
{"role": "user", "content": user},
]
run_log: list[dict] = []
final_text = ""
token_param = (
"max_completion_tokens" if any(p in model for p in ("gpt-5", "o3", "o4")) else "max_tokens"
)
for step in range(max_steps):
resp = client.chat.completions.create(
model=model,
messages=messages,
tools=tools,
**{token_param: 8192},
)
msg = resp.choices[0].message
# Append assistant turn
assistant_dict: dict[str, Any] = {"role": "assistant"}
if msg.content:
assistant_dict["content"] = msg.content
final_text = msg.content
if msg.tool_calls:
assistant_dict["tool_calls"] = [
{"id": tc.id, "type": "function",
"function": {"name": tc.function.name, "arguments": tc.function.arguments}}
for tc in msg.tool_calls
]
messages.append(assistant_dict)
if not msg.tool_calls:
break
for tc in msg.tool_calls:
try:
args = json.loads(tc.function.arguments)
except Exception:
args = {}
t0 = time.monotonic()
result = call_mcp_tool(mcp_url, mcp_token, tc.function.name, args)
dt = round(time.monotonic() - t0, 2)
run_log.append({
"step": step,
"tool": tc.function.name,
"arguments": args,
"success": "error" not in result,
"latency_sec": dt,
"result_summary": str(result)[:500],
})
messages.append({
"role": "tool",
"tool_call_id": tc.id,
"content": json.dumps(result)[:8000],
})
return run_log, final_text
# ---------------------------------------------------------------------------
# Google Gemini agent loop
# ---------------------------------------------------------------------------
def _run_google(
api_key: str,
model: str,
system: str,
user: str,
tool_schemas: list[dict],
mcp_url: str,
mcp_token: str,
max_steps: int,
) -> tuple[list[dict], str]:
from google import genai
from google.genai import types as gtypes
client = genai.Client(api_key=api_key)
function_decls = []
for s in tool_schemas:
params = s.get("parameters") or {}
function_decls.append(
gtypes.FunctionDeclaration(
name=s["name"],
description=s.get("description", ""),
parameters=params,
)
)
tools = [gtypes.Tool(function_declarations=function_decls)]
history = [gtypes.Content(role="user", parts=[gtypes.Part(text=f"{system}\n\n{user}")])]
run_log: list[dict] = []
final_text = ""
for step in range(max_steps):
resp = client.models.generate_content(
model=model,
contents=history,
config=gtypes.GenerateContentConfig(tools=tools),
)
cand = resp.candidates[0]
history.append(cand.content)
function_calls = []
for part in cand.content.parts:
if getattr(part, "function_call", None):
function_calls.append(part.function_call)
elif getattr(part, "text", None):
final_text = part.text
if not function_calls:
break
function_responses = []
for fc in function_calls:
args = dict(fc.args) if fc.args else {}
t0 = time.monotonic()
result = call_mcp_tool(mcp_url, mcp_token, fc.name, args)
dt = round(time.monotonic() - t0, 2)
run_log.append({
"step": step, "tool": fc.name, "arguments": args,
"success": "error" not in result, "latency_sec": dt,
"result_summary": str(result)[:500],
})
function_responses.append(
gtypes.Part(function_response=gtypes.FunctionResponse(
name=fc.name, response=result,
))
)
history.append(gtypes.Content(role="user", parts=function_responses))
return run_log, final_text
# ---------------------------------------------------------------------------
# Provider dispatch
# ---------------------------------------------------------------------------
SUPPORTED_PROVIDERS = {"anthropic", "openai", "deepseek", "google"}
def run_agent_on_task(
provider: str,
api_key: str,
model: str,
task_prompt: str,
mcp_url: str,
mcp_token: str = "",
system_prompt: str | None = None,
max_steps: int = DEFAULT_MAX_STEPS,
) -> dict[str, Any]:
"""Run a single agent submission against one task.
The contract mirrors what the old HTTP dispatcher returned, so the
rest of the scoring pipeline (eval_dispatcher.score_cpu_components,
eval_boltz.run_boltz_posteval, eval_judge.run_judge_panel) works
unchanged.
Returns:
{
"success": bool,
"sequences": [str, ...],
"run_log": [{step, tool, arguments, success, latency_sec, ...}],
"total_steps": int,
"total_time_sec": float,
"metrics": {}, # reserved for future use
"error": str (only when success=False),
}
"""
if provider not in SUPPORTED_PROVIDERS:
return {
"success": False,
"error": f"Unknown provider '{provider}'. Use one of: {sorted(SUPPORTED_PROVIDERS)}",
}
system = system_prompt or DEFAULT_SYSTEM_PROMPT
schemas = load_tool_schemas()
start = time.monotonic()
try:
if provider == "anthropic":
run_log, final_text = _run_anthropic(
api_key, model, system, task_prompt, schemas,
mcp_url, mcp_token, max_steps,
)
elif provider == "openai":
run_log, final_text = _run_openai_compat(
api_key, model, system, task_prompt, schemas,
mcp_url, mcp_token, max_steps,
)
elif provider == "deepseek":
run_log, final_text = _run_openai_compat(
api_key, model, system, task_prompt, schemas,
mcp_url, mcp_token, max_steps,
base_url="https://api.deepseek.com",
)
elif provider == "google":
run_log, final_text = _run_google(
api_key, model, system, task_prompt, schemas,
mcp_url, mcp_token, max_steps,
)
else:
return {"success": False, "error": "unreachable"}
except Exception as e:
elapsed = round(time.monotonic() - start, 1)
logger.exception(f"Agent loop crashed for provider={provider}")
return {
"success": False,
"error": f"Agent loop crashed: {type(e).__name__}: {str(e)[:400]}",
"total_time_sec": elapsed,
}
sequences = extract_sequences(final_text)
elapsed = round(time.monotonic() - start, 1)
return {
"success": True,
"sequences": sequences,
"run_log": run_log,
"total_steps": len(run_log),
"total_time_sec": elapsed,
"metrics": {},
}
|