Initial commit.

Yeah I know there are a lot of nonessential files but w/e.

.gitignore

@@ -0,0 +1,5 @@
+venv
+.env
+nohup.out
+__pycache__
+.ropeproject

README.md

@@ -0,0 +1,5 @@
+# ASCO Abstract Success Predictor
+
+Simple web interface to predict the chance a given abstract will be accepted for
+oral presentation, based only on its title. Code is currently in shambles, I
+might bring it up to my current standards some day.

app.py

@@ -0,0 +1,28 @@
+from flask import Flask, redirect, render_template, request, url_for
+import fit
+
+app = Flask(__name__)
+
+@app.route("/", methods=("GET", "POST"))
+def index():
+    if request.method == "POST":
+        title = request.form["title"]
+        embedding = fit.get(title);
+
+        #nclose = request.form["nclose"]
+
+        percent = fit.percent(embedding);
+        closest = fit.closest(embedding, 10);
+
+        tprob = closest[1]
+
+        return redirect(url_for("index", tprob=closest[1], result=percent, title=title, closest=closest[0]))
+
+    result = request.args.get("result")
+    title = request.args.get("title")
+    closest = request.args.get("closest")
+    #nclose = request.args.get("nclose")
+    tprob = request.args.get("tprob")
+    if title == None: title = ""
+    #if nclose == None: nclose = 10
+    return render_template("index.html", tprob=tprob, result=result, title=title, closest=closest)

fit.py

@@ -0,0 +1,58 @@
+import numpy as ν
+import csv
+import get_emb
+
+MAX_RES = 10
+
+φ = open('beta', 'r');                                      # coefficients
+Δ = [float(el) for el in (open('Delta', 'r').read().split('\n')[:-1])]
+#α = -0.8569279;                                             # some magic constant
+α = 0;
+M = open('titles2.text', 'r').read().strip().split('\n')     # TITLE\nAUTHORS\nACCEPTED?
+T = ν.array(M[0::4])
+A = ν.array(M[1::4])
+O = ν.array(M[2::4])
+X = ν.loadtxt(open('embeddings2.nsv', 'rb'), delimiter=',', skiprows=0)
+
+NN = open('NN', 'r').read().split('\n')
+
+β = φ.read().split('\n');
+β = β[:-1]
+β = [float(el) for el in β]
+
+def get(θ):
+    return(get_emb.get_embedding(θ))
+
+def percent(χ):
+    γ = α + ν.dot(χ, β)
+    π = ν.exp(γ) / (1 + ν.exp(γ))
+
+    return(str(π)[2:4] + '%')
+
+def closest(χ, n):
+    n = abs(n)
+    n = n % MAX_RES 
+    if n == 0: n = MAX_RES
+
+    ψ = ν.array(ν.dot(X, χ))
+    topn = T[ν.argsort(ψ)[-n:]]
+    aopn = A[ν.argsort(ψ)[-n:]]
+    oopn = O[ν.argsort(ψ)[-n:]]
+
+    print(ν.argsort(ψ)[-n:])
+
+    out = ""#"tail prob = " + str(percentile_far(ν.max(ψ))) + "\n"
+    for i in reversed(range(len(topn))):
+        if oopn[i] == "TRUE": 
+            p = "presented"
+        else:
+            p = "online-only"
+        out += topn[i] + " <i>(" + aopn[i] + ", " + p + ")</i>\n"
+
+    tailprob = int(percentile_far(ν.max(ψ))*100)
+
+    return [out, tailprob]
+
+def percentile_far(q_dist):
+    return sum(1*(ν.array(Δ)<=q_dist)) / len(Δ) # fraction of abstracts further from their nearest neighbor than χ
+

fit.r

@@ -0,0 +1,37 @@
+library('glmnet')
+library('pROC')
+library('survival')
+inv.logit <- function(x) {exp(x) / (1 + exp(x))}
+read.table('titles.text', header=FALSE, quote='', sep="\n") -> D
+matrix(as.vector(D[,1]), 3, nrow(D) / 3) -> M
+M <- t(M) 
+X <- as.matrix(read.csv('embeddings.nsv', header=F))  # matrix of embeddings
+
+# remove duplicates (if present)
+X <- X[!duplicated(M[,1]),]
+M <- M[!duplicated(M[,1]),]
+write.table(as.vector(t(cbind(M,""))), "titles2.text", sep="\n", row.names=FALSE, col.names=FALSE, quote=F)
+write.table(X, "embeddings2.nsv", sep=",", row.names=F, col.names=F, quote=F);
+
+Y <- as.numeric(as.logical(M[,3]))
+V <- cv.glmnet(x=X, y=Y, family="binomial", type.measure="auc")
+
+Z <- glmnet(x=X, y=Y, lambda=V$lambda.min, family="binomial")
+A <- predict(Z, newx=X, type="response")
+B <- glm(Y~A, family="binomial")
+R <- roc(Y, B$fitted.values)
+beta <- Z$beta;
+
+
+#P <- B$fitted.values
+#O1 <- X %*% as.vector(Z$beta)
+#O2 <- B$coefficients[1] + B$coefficients[2] * O1
+#O3 <- inv.logit(exp(1) + O2)
+#beta <- as.vector(Z$beta) * B$coefficients[2];
+#alpha <- B$coefficients[1] + exp(1);
+
+δ <- X %*% t(X)
+diag(δ) <- NA
+Δ <- apply(δ, 1, max, na.rm=T)
+write.table(as.vector(beta), "beta", sep="\n", row.names=F, col.names=F)
+write.table(Δ, "Delta", sep="\n", row.names=F, col.names=F)

get_emb.py

@@ -0,0 +1,5 @@
+import openai
+
+def get_embedding(text, model="text-embedding-ada-002"):
+    client = openai.OpenAI();
+    return client.embeddings.create(input = [text], model=model).data[0].embedding

gunicorn_config.py

@@ -0,0 +1,2 @@
+bind = "0.0.0.0:80"
+workers = 2

run.sh

@@ -0,0 +1,3 @@
+#!/bin/sh
+
+gunicorn --bind=0.0.0.0:8077 --worker-tmp-dir /dev/shm app:app

static/about.html

@@ -0,0 +1,27 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+   <meta charset="UTF-8">
+   <link rel="stylesheet" href="main.css">
+   <title>About</title>
+</head>
+<body>
+   <center><h1>About</h1></center>
+
+   <h2>The Data</h2>
+   <p>
+      Information was scraped from <a href="//ascopubs.org/jco/meeting">this page</a> on 2023-06-28.
+      Over 5000 abstract titles from 2023 were used as training data.
+      Embeddings for these were generated through the OpenAI API, with the <code>text-embedding-ada-002</code> model.
+   </p>
+
+   <h2>The Model</h2>
+   <p>
+      A penalized logistic regression model was fit using the <code>glmnet</code> R package.
+      The tuning parameter was selected using cross validation.
+      The area under the ROC curve was 0.83 in the training data.
+   </p>
+
+   <center><a href="..">back</a></center>
+</body>
+</html>

static/main.css

@@ -0,0 +1,129 @@
+/* global */
+:root {
+   font-family:         monospace;
+
+   --fgm:               #bdae93;
+   --bgm:               #1d2021;
+   --bgs:               #282828;
+   --bgr:               #32302f;
+}
+
+* {
+   border-radius:       0px !important;
+}
+
+/* background */
+html {
+   background-color:    var(--bgm);
+   padding:             0;
+   margin:              0;
+   display:             flex;
+   justify-content:     center;
+   font-size:           16px;
+}
+
+/* page */
+body {
+   background-color:    var(--bgs);
+   color:               var(--fgm);
+   margin:              0;
+   padding-inline:      2em;
+   padding-block:       1em;
+   height:              fit-content;
+   min-height:          100vh;
+   max-width:           50em;
+   flex-grow:           1;
+   font-size:           18px;
+}
+
+/* headers */
+h1 {
+   color:               #fb4934;
+} h2 {
+   color:               #fabd2f;
+} h3 {
+   color:               #b8bb26;
+} h4 {
+   color:               #8ec07c
+} h5 {
+   color:               #83a598;
+} h1, h2, 
+  h3, h4, 
+  h5 {
+   font-weight:         bold;
+}
+
+/* text styles */
+b, i, u {
+   color:               #d79921;
+}
+
+/* links */
+a {
+   color:               #458588;
+} a:hover {
+   color:               var(--bgs);
+   background-color:    #458588;
+   text-decoration:     none;
+}
+
+/* preformatted */
+pre,code {
+   color:               #8ec07c;
+   background-color:    var(--bgr);
+   width:               min-content;
+}
+
+/* misc */
+.c {
+   color:               #458588;
+   transition:          filter 0.2s;
+   cursor:              help;
+} .c:hover {
+   filter:              blur(2px)
+}
+
+form {
+   margin:              1em;
+   padding:             1em;
+}
+
+input[type="text"],
+textarea,
+input[type="number"],
+input[type="submit"] {
+   background-color:    var(--bgr);
+   border:              none;
+   outline:             none;
+   color:               #8ec07c !important;
+   font-size:           1em;
+   font-family:         monospace;
+   width:               min-content !important;
+}
+
+textarea:active,
+textarea:focus,
+input[type="submit"]:active,
+input[type="submit"]:focus,
+input[type="text"]:active,
+input[type="number"]:active,
+input[type="text"]:focus,
+input[type="number"]:focus {
+   background-color:    #83c07c;
+   color:               var(--bgs) !important;
+   border:              none;
+   outline:             none;
+   font-size:           1em;
+   font-family:         monospace;
+}
+
+input {
+   margin:              0.2em;
+   border-radius:       0px !important;
+}
+
+input[type="submit"],
+input[type="submit"]:active,
+input[type="submit"]:focus {
+   border:              3px solid #83c07c
+}

templates/index.html

@@ -0,0 +1,67 @@
+<!DOCTYPE html>
+<head>
+  <title>ASCO Abstract Title Evaluator</title>
+  <link rel="stylesheet" href="static/main.css">
+</head>
+
+<body>
+   <center><h1>ASCO Abstract Title Evaluator</h1></center>
+   <center><form action="/" method="post">
+         <textarea   rows="3"
+                     cols="64"
+                     name="title" 
+                     placeholder="Enter abstract title..." 
+                     required>{{ title }}</textarea><br>
+         <input type="submit" value="Evaluate"/>
+   </form></center>
+
+   {% if tprob %}
+      <center style="margin-inline:10em; text-align: justify;"><p>
+         <h3>{{ tprob }}% Similarity</h3>
+         The similarity of your abstract title to accepted 
+         ASCO 2023 titles is in this percentile.
+      </p></center>
+      <br>
+   {% endif %}
+
+   {% if result %}
+      <center style="margin-inline:10em; text-align: justify;"><p>
+         <h3>{{ result }} Chance of on-site presentation</h3>
+         Probability of on-site presentation 
+         versus online only publication of your 
+         abstract, based on the title, if
+         accepted.
+         {% if tprob|int < 5 %}
+         <p><i>Warning: the similarity is too low to be confident in this prediction.</i></p><br>
+         {% endif %}
+      </p></center>
+      <br>
+   {% endif %}
+
+   {% if closest %}
+   {% set closed = closest.split('\n')[:-1] %}
+   <h3>Closest {{ closed|length }} titles from ASCO 2023:</h3>
+   <ol>
+      {% for close in closed %}
+      <li>{{ close|safe }}</li>
+      {% endfor %}
+   </ol>
+   {% endif %}
+
+   <center><a href="{{ url_for('static', filename='about.html') }}">about</a></center>
+
+	<script>
+      function sub(event) {
+          if (event.key == 'Enter' && !event.shiftKey) {
+              if (!event.repeat) {
+                  const newEvent = new Event("submit", {cancelable: true});
+                  event.target.form.dispatchEvent(newEvent);
+              }
+
+              event.preventDefault(); 
+          }
+      }
+
+      document.querySelector("textarea").addEventListener("keydown", sub);
+	</script>
+</body>