Skip to content

TxGemma: Show example of converting regression output back to original label. #206

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

TxGemma: Show example of converting regression output back to original label. #206

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
165 changes: 155 additions & 10 deletions TxGemma/[TxGemma]Quickstart_with_Hugging_Face.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -662,6 +662,56 @@
"print(response)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"**Run a regression task and convert output back to the original label**"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"!wget https://github.com/google-gemini/gemma-cookbook/blob/main/TxGemma/regression_conversion_parameters.json"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def denormalize_str(s: str | float, min_value: str | float, max_value: str | float) -> float:\n",
" convert_func = lambda s: (float(s)/1000.0) * (float(max_value) - float(min_value)) + float(min_value)\n",
" try:\n",
" value = convert_func(s)\n",
" except:\n",
" value = convert_func(500.)\n",
" return value\n",
"\n",
"parameters_file = \"regression_conversion_parameters.json\"\n",
"\n",
"if os.path.isfile(parameters_file):\n",
" with open(parameters_file, \"r\") as f:\n",
" reg_params = json.load(f)\n",
"\n",
" task_name = \"Caco2_Wang\"\n",
" prompt = tdc_prompts_json[task_name].replace(input_type, drug_smiles)\n",
" input_ids = tokenizer(prompt, return_tensors=\"pt\").to(\"cuda\")\n",
" outputs = model.generate(**input_ids, max_new_tokens=8)\n",
" response = tokenizer.decode(outputs[0], skip_special_tokens=True)\n",
" integer = response.split(\"Answer:\")[1].strip()\n",
"\n",
" print(f\"TxGemma Output: {integer}\")\n",
" print(f\"Denormalized output: {denormalize_str(integer, *reg_params['Caco2_Wang'])}\")\n",
"\n",
"else:\n",
" print(f\"{parameters_file} not found for regression example.\")\n"
]
},
{
"cell_type": "markdown",
"id": "V4WlXpx2fGN4",
Expand Down Expand Up @@ -710,7 +760,12 @@
"outputs": [
{
"data": {
"text/markdown": "\n\n---\n\n",
"text/markdown": [
"\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand All @@ -720,7 +775,19 @@
},
{
"data": {
"text/markdown": "**User:**\n\nInstructions: Answer the following question about drug properties.\nContext: As a membrane separating circulating blood and brain extracellular fluid, the blood-brain barrier (BBB) is the protection layer that blocks most foreign drugs. Thus the ability of a drug to penetrate the barrier to deliver to the site of action forms a crucial challenge in development of drugs for central nervous system.\nQuestion: Given a drug SMILES string, predict whether it\n(A) does not cross the BBB (B) crosses the BBB\nDrug SMILES: CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21\nAnswer:\n\n---\n\n",
"text/markdown": [
"**User:**\n",
"\n",
"Instructions: Answer the following question about drug properties.\n",
"Context: As a membrane separating circulating blood and brain extracellular fluid, the blood-brain barrier (BBB) is the protection layer that blocks most foreign drugs. Thus the ability of a drug to penetrate the barrier to deliver to the site of action forms a crucial challenge in development of drugs for central nervous system.\n",
"Question: Given a drug SMILES string, predict whether it\n",
"(A) does not cross the BBB (B) crosses the BBB\n",
"Drug SMILES: CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21\n",
"Answer:\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand All @@ -730,7 +797,14 @@
},
{
"data": {
"text/markdown": "**TxGemma:**\n\n(B)\n\n---\n\n",
"text/markdown": [
"**TxGemma:**\n",
"\n",
"(B)\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand All @@ -740,7 +814,14 @@
},
{
"data": {
"text/markdown": "**User:**\n\nExplain your reasoning based on the molecule structure.\n\n---\n\n",
"text/markdown": [
"**User:**\n",
"\n",
"Explain your reasoning based on the molecule structure.\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand All @@ -750,7 +831,24 @@
},
{
"data": {
"text/markdown": "**TxGemma:**\n\nHere's the breakdown of why the drug with SMILES CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21 likely crosses the BBB:\n\n* **Small Size:** The molecule is relatively small, which is a general favorable characteristic for BBB penetration. Larger molecules have a harder time squeezing through the tight junctions between brain endothelial cells.\n* **Lipophilicity (Hydrophobicity):** The presence of multiple carbon and hydrogen atoms in the benzene rings and alkyl chains makes the molecule predominantly lipophilic (fat-loving). Lipophilicity is a key determinant of BBB permeability. The more lipophilic a drug, the easier it crosses the lipid-rich cell membranes of the BBB.\n* **Lack of Charged Groups:** The molecule lacks large, charged groups (like carboxyl or amine groups). Charged groups tend to be repelled by the lipid bilayer of the BBB, making it harder to cross. \n* **Absence of Specific BBB Targets:** While some drugs have specific transporters that help them across the BBB, this molecule doesn't appear to have any obvious targeting groups.\n\n**In summary, the drug's small size, lipophilic nature, lack of significant charge, and absence of specific BBB-targeting groups suggest that it likely possesses a good ability to cross the blood-brain barrier.**\n\n**Important Note:** This is a general prediction based on structural features. Actual BBB permeability is complex and can be influenced by many factors. Experimental validation is always necessary to confirm drug penetration ability. \n\n\n---\n\n",
"text/markdown": [
"**TxGemma:**\n",
"\n",
"Here's the breakdown of why the drug with SMILES CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21 likely crosses the BBB:\n",
"\n",
"* **Small Size:** The molecule is relatively small, which is a general favorable characteristic for BBB penetration. Larger molecules have a harder time squeezing through the tight junctions between brain endothelial cells.\n",
"* **Lipophilicity (Hydrophobicity):** The presence of multiple carbon and hydrogen atoms in the benzene rings and alkyl chains makes the molecule predominantly lipophilic (fat-loving). Lipophilicity is a key determinant of BBB permeability. The more lipophilic a drug, the easier it crosses the lipid-rich cell membranes of the BBB.\n",
"* **Lack of Charged Groups:** The molecule lacks large, charged groups (like carboxyl or amine groups). Charged groups tend to be repelled by the lipid bilayer of the BBB, making it harder to cross. \n",
"* **Absence of Specific BBB Targets:** While some drugs have specific transporters that help them across the BBB, this molecule doesn't appear to have any obvious targeting groups.\n",
"\n",
"**In summary, the drug's small size, lipophilic nature, lack of significant charge, and absence of specific BBB-targeting groups suggest that it likely possesses a good ability to cross the blood-brain barrier.**\n",
"\n",
"**Important Note:** This is a general prediction based on structural features. Actual BBB permeability is complex and can be influenced by many factors. Experimental validation is always necessary to confirm drug penetration ability. \n",
"\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand Down Expand Up @@ -805,7 +903,12 @@
"outputs": [
{
"data": {
"text/markdown": "\n\n---\n\n",
"text/markdown": [
"\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand All @@ -815,7 +918,19 @@
},
{
"data": {
"text/markdown": "**User:**\n\nInstructions: Answer the following question about drug properties.\nContext: As a membrane separating circulating blood and brain extracellular fluid, the blood-brain barrier (BBB) is the protection layer that blocks most foreign drugs. Thus the ability of a drug to penetrate the barrier to deliver to the site of action forms a crucial challenge in development of drugs for central nervous system.\nQuestion: Given a drug SMILES string, predict whether it\n(A) does not cross the BBB (B) crosses the BBB\nDrug SMILES: CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21\nAnswer:\n\n---\n\n",
"text/markdown": [
"**User:**\n",
"\n",
"Instructions: Answer the following question about drug properties.\n",
"Context: As a membrane separating circulating blood and brain extracellular fluid, the blood-brain barrier (BBB) is the protection layer that blocks most foreign drugs. Thus the ability of a drug to penetrate the barrier to deliver to the site of action forms a crucial challenge in development of drugs for central nervous system.\n",
"Question: Given a drug SMILES string, predict whether it\n",
"(A) does not cross the BBB (B) crosses the BBB\n",
"Drug SMILES: CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21\n",
"Answer:\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand All @@ -825,7 +940,14 @@
},
{
"data": {
"text/markdown": "**TxGemma:**\n\n(B)\n\n---\n\n",
"text/markdown": [
"**TxGemma:**\n",
"\n",
"(B)\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand All @@ -835,7 +957,14 @@
},
{
"data": {
"text/markdown": "**User:**\n\nExplain your reasoning based on the molecule structure.\n\n---\n\n",
"text/markdown": [
"**User:**\n",
"\n",
"Explain your reasoning based on the molecule structure.\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand All @@ -845,7 +974,23 @@
},
{
"data": {
"text/markdown": "**TxGemma:**\n\nHere's the breakdown of why the drug with SMILES CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21 likely crosses the BBB:\n\n* **Small Size:** The molecule is relatively small, which is a general favorable characteristic for BBB penetration. Larger molecules have a harder time squeezing through the tight junctions between brain endothelial cells.\n* **Lipophilicity (Hydrophobicity):** The presence of multiple carbon and hydrogen atoms in the benzene rings and alkyl chains makes the molecule predominantly lipophilic (fat-loving). Lipophilicity is a key determinant of BBB permeability. The more lipophilic a drug, the easier it crosses the lipid-rich cell membranes of the BBB.\n* **Lack of Charged Groups:** The molecule lacks large, charged groups (like carboxyl or amine groups). Charged groups tend to be repelled by the lipid bilayer of the BBB, making it harder to cross. \n* **Absence of Specific BBB Targets:** While some drugs have specific transporters that help them across the BBB, this molecule doesn't appear to have any obvious targeting groups.\n\n**In summary, the drug's small size, lipophilic nature, lack of significant charge, and absence of specific BBB-targeting groups suggest that it likely possesses a good ability to cross the blood-brain barrier.**\n\n**Important Note:** This is a general prediction based on structural features. Actual BBB permeability is complex and can be influenced by many factors. Experimental validation is always necessary to confirm drug penetration ability.\n\n---\n\n",
"text/markdown": [
"**TxGemma:**\n",
"\n",
"Here's the breakdown of why the drug with SMILES CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21 likely crosses the BBB:\n",
"\n",
"* **Small Size:** The molecule is relatively small, which is a general favorable characteristic for BBB penetration. Larger molecules have a harder time squeezing through the tight junctions between brain endothelial cells.\n",
"* **Lipophilicity (Hydrophobicity):** The presence of multiple carbon and hydrogen atoms in the benzene rings and alkyl chains makes the molecule predominantly lipophilic (fat-loving). Lipophilicity is a key determinant of BBB permeability. The more lipophilic a drug, the easier it crosses the lipid-rich cell membranes of the BBB.\n",
"* **Lack of Charged Groups:** The molecule lacks large, charged groups (like carboxyl or amine groups). Charged groups tend to be repelled by the lipid bilayer of the BBB, making it harder to cross. \n",
"* **Absence of Specific BBB Targets:** While some drugs have specific transporters that help them across the BBB, this molecule doesn't appear to have any obvious targeting groups.\n",
"\n",
"**In summary, the drug's small size, lipophilic nature, lack of significant charge, and absence of specific BBB-targeting groups suggest that it likely possesses a good ability to cross the blood-brain barrier.**\n",
"\n",
"**Important Note:** This is a general prediction based on structural features. Actual BBB permeability is complex and can be influenced by many factors. Experimental validation is always necessary to confirm drug penetration ability.\n",
"\n",
"---\n",
"\n"
],
"text/plain": [
"<IPython.core.display.Markdown object>"
]
Expand Down
1 change: 1 addition & 0 deletions TxGemma/regression_conversion_parameters.json
View file
Open in desktop
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

how where these values set? should the min and max values match the min and max values found in the CSS column in the DrugComb data set from tdcommons? In this example the min and max values are [-54.0485, 99.8405]. however it is set to ["-9.325965", "80.077045"] here. -9.325965 or 80.077045 does not exist in the CSS column from the DrugComb data set as far as I can tell.

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

+1, for BindingDB_ic50 we see 2 and 10 currently while thle column Y in bindingdb_ic50.csv shows value 0 as the lowest (row 149124) and 10000000 (row 5161) as the highest

Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
{"Caco2_Wang": ["-7.76", "-3.72948"], "VDss_Lombardo": ["0.01", "700.0"], "Half_Life_Obach": ["0.13", "1200.0"], "Clearance_Hepatocyte_AZ": ["3.0", "150.0"], "Clearance_Microsome_AZ": ["3.0", "150.0"], "LD50_Zhu": ["-0.343", "7.1"], "PPBR_AZ": ["12.63", "99.95"], "BindingDB_kd": ["2.0", "10.0"], "BindingDB_ic50": ["2.0", "10.0"], "BindingDB_ki": ["2.0", "10.0"], "DAVIS": ["5.0", "9.93554"], "USPTO_Yields": ["0.0", "0.9262238423"], "Buchwald_Hartwig": ["0.06", "1.0"], "OncoPolyPharmacology": ["-49.43829775", "71.29880244"], "DrugComb_CSS": ["-9.325965", "80.077045"], "DrugComb_HSA": ["-17.25489767", "15.41516102"], "DrugComb_Loewe": ["-65.01337758", "11.96498908"], "DrugComb_Bliss": ["-16.47171678", "16.74003994"], "DrugComb_ZIP": ["-12.48189979", "15.68920757"], "Protein_SAbDab": ["4.811506144", "9.989323594"], "Lipophilicity_AstraZeneca": ["-1.48", "4.5"], "Solubility_AqSolDB": ["-11.99893786", "1.9675127"], "TAP_CDR_Length": ["39.0", "60.0"], "TAP_PSH": ["83.8389", "173.8525"], "TAP_PPC": ["0.0", "3.1617"], "TAP_PNC": ["0.0", "3.4997"], "TAP_SFvCSP": ["-20.4", "36.0"], "Leenay_Fraction_Insertions": ["0.0011676397", "0.8024803216"], "Leenay_Avg_Insertion_Length": ["0.0", "20.7738817"], "Leenay_Avg_Deletion_Length": ["2.110951518", "46.02702703"], "Leenay_Indel_Diversity": ["0.8654522802", "5.548347734"], "Leenay_Fraction_Frameshifts": ["0.001676409904", "0.940090951"], "KIBA": ["7.1", "17.2001795"], "BindingDB_Patent": ["-11.51291547", "16.11809565"]}
Loading