#!/usr/bin/env python3 """Score how much a genome contig raises Evo 2's likelihood of each candidate 16S (ssu_link_set.py output). For every (test genome, candidate 16S) pair: cond sum of log p(16S token | contig + preceding 16S tokens), over 16S tokens 2..L uncond sum of log p(16S token | preceding 16S tokens) for the 16S alone, same tokens gain cond - uncond: what the contig adds (a pointwise mutual information); the genome's own 16S should gain most Both orientations of the contig are scored (contig strand relative to the rRNA operon is unknown); gain takes the larger. Writes OUT.tsv: test, candidate, relation, len, uncond, cond_fwd, cond_rev, gain. usage (in the evo2 container): ssu_link_score.py SET OUT --model_name evo2_40b --local_path evo2_40b.pt SET = ssu_link_set.py prefix (SET.contigs.fa, SET.ssu.fa, SET.pairs.tsv) """ import argparse, time import torch from evo2 import Evo2 ap = argparse.ArgumentParser() ap.add_argument("set"); ap.add_argument("out") ap.add_argument("--model_name", default="evo2_40b"); ap.add_argument("--local_path", required=True) ap.add_argument("--start", type=int, default=0); ap.add_argument("--end", type=int, default=10 ** 9) a = ap.parse_args() def fasta(p): d, n = {}, None for l in open(p): if l[0] == ">": n = l[1:].strip(); d[n] = [] else: d[n].append(l.strip()) return {k: "".join(v) for k, v in d.items()} contig, ssu = fasta(a.set + ".contigs.fa"), fasta(a.set + ".ssu.fa") pairs = [l.rstrip("\n").split("\t") for l in open(a.set + ".pairs.tsv")][1:] tests = sorted({p[0] for p in pairs})[a.start:a.end] rc = lambda s: s[::-1].translate(str.maketrans("ACGTN", "TGCAN")) model = Evo2(a.model_name, local_path=a.local_path) dev = next(model.model.parameters()).device @torch.no_grad() def tail_logprob(prefix, target): """sum log p(target[1:] | prefix + target[:i]) for i = 1..len-1""" ids = torch.tensor(model.tokenizer.tokenize(prefix + target), dtype=torch.int).unsqueeze(0).to(dev) logits, _ = model.model.forward(ids) # as evo2's own test does; logits sit on the last GPU lp = torch.log_softmax(logits[0, :-1].float(), -1) nxt = ids[0, 1:].long().to(lp.device) tok = lp.gather(1, nxt.unsqueeze(1)).squeeze(1) return float(tok[len(prefix):].sum()) # tok[i] predicts token i+1: target[1:] is predicted from index len(prefix) uncond = {} out = open(a.out + ".tsv", "w") out.write("test\tcandidate\trelation\tlen\tuncond\tcond_fwd\tcond_rev\tgain\n") t0 = time.time() for k, g in enumerate(tests): c = contig[g] for p in (p for p in pairs if p[0] == g): s = ssu[p[1]] if p[1] not in uncond: uncond[p[1]] = tail_logprob("", s) f, r = tail_logprob(c, s), tail_logprob(rc(c), s) out.write(f"{g}\t{p[1]}\t{p[2]}\t{len(s)}\t{uncond[p[1]]:.3f}\t{f:.3f}\t{r:.3f}\t{max(f, r) - uncond[p[1]]:.3f}\n") out.flush() print(f"{k + 1}/{len(tests)} genomes, {time.time() - t0:.0f} s", flush=True)